]>
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 LT |
16 | #include <linux/skbuff.h> |
17 | #include <linux/ip.h> | |
18 | #include <linux/tcp.h> | |
19 | #include <linux/udp.h> | |
20 | #include <linux/icmp.h> | |
21 | #include <linux/if_arp.h> | |
22 | #include <linux/proc_fs.h> | |
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 PM |
28 | #include <net/netfilter/nf_conntrack.h> |
29 | #include <net/checksum.h> | |
1da177e4 | 30 | |
136e92bb | 31 | #define CLUSTERIP_VERSION "0.8" |
1da177e4 LT |
32 | |
33 | #define DEBUG_CLUSTERIP | |
34 | ||
35 | #ifdef DEBUG_CLUSTERIP | |
36 | #define DEBUGP printk | |
37 | #else | |
38 | #define DEBUGP | |
39 | #endif | |
40 | ||
41 | MODULE_LICENSE("GPL"); | |
42 | MODULE_AUTHOR("Harald Welte <laforge@netfilter.org>"); | |
43 | MODULE_DESCRIPTION("iptables target for CLUSTERIP"); | |
44 | ||
45 | struct clusterip_config { | |
46 | struct list_head list; /* list of all configs */ | |
47 | atomic_t refcount; /* reference count */ | |
44513624 KK |
48 | atomic_t entries; /* number of entries/rules |
49 | * referencing us */ | |
1da177e4 | 50 | |
6a19d614 | 51 | __be32 clusterip; /* the IP address */ |
1da177e4 LT |
52 | u_int8_t clustermac[ETH_ALEN]; /* the MAC address */ |
53 | struct net_device *dev; /* device */ | |
54 | u_int16_t num_total_nodes; /* total number of nodes */ | |
136e92bb | 55 | unsigned long local_nodes; /* node number array */ |
1da177e4 LT |
56 | |
57 | #ifdef CONFIG_PROC_FS | |
58 | struct proc_dir_entry *pde; /* proc dir entry */ | |
59 | #endif | |
60 | enum clusterip_hashmode hash_mode; /* which hashing mode */ | |
61 | u_int32_t hash_initval; /* hash initialization */ | |
62 | }; | |
63 | ||
64 | static LIST_HEAD(clusterip_configs); | |
65 | ||
136e92bb | 66 | /* clusterip_lock protects the clusterip_configs list */ |
e45b1be8 | 67 | static DEFINE_RWLOCK(clusterip_lock); |
1da177e4 LT |
68 | |
69 | #ifdef CONFIG_PROC_FS | |
9a32144e | 70 | static const struct file_operations clusterip_proc_fops; |
1da177e4 LT |
71 | static struct proc_dir_entry *clusterip_procdir; |
72 | #endif | |
73 | ||
74 | static inline void | |
44513624 KK |
75 | clusterip_config_get(struct clusterip_config *c) |
76 | { | |
1da177e4 LT |
77 | atomic_inc(&c->refcount); |
78 | } | |
79 | ||
80 | static inline void | |
44513624 KK |
81 | clusterip_config_put(struct clusterip_config *c) |
82 | { | |
83 | if (atomic_dec_and_test(&c->refcount)) | |
84 | kfree(c); | |
85 | } | |
86 | ||
87 | /* increase the count of entries(rules) using/referencing this config */ | |
88 | static inline void | |
89 | clusterip_config_entry_get(struct clusterip_config *c) | |
90 | { | |
91 | atomic_inc(&c->entries); | |
92 | } | |
93 | ||
94 | /* decrease the count of entries using/referencing this config. If last | |
95 | * entry(rule) is removed, remove the config from lists, but don't free it | |
96 | * yet, since proc-files could still be holding references */ | |
97 | static inline void | |
98 | clusterip_config_entry_put(struct clusterip_config *c) | |
99 | { | |
100 | if (atomic_dec_and_test(&c->entries)) { | |
e45b1be8 | 101 | write_lock_bh(&clusterip_lock); |
1da177e4 | 102 | list_del(&c->list); |
e45b1be8 | 103 | write_unlock_bh(&clusterip_lock); |
44513624 | 104 | |
1da177e4 LT |
105 | dev_mc_delete(c->dev, c->clustermac, ETH_ALEN, 0); |
106 | dev_put(c->dev); | |
44513624 KK |
107 | |
108 | /* In case anyone still accesses the file, the open/close | |
109 | * functions are also incrementing the refcount on their own, | |
110 | * so it's safe to remove the entry even if it's in use. */ | |
111 | #ifdef CONFIG_PROC_FS | |
112 | remove_proc_entry(c->pde->name, c->pde->parent); | |
113 | #endif | |
1da177e4 LT |
114 | } |
115 | } | |
116 | ||
1da177e4 | 117 | static struct clusterip_config * |
6a19d614 | 118 | __clusterip_config_find(__be32 clusterip) |
1da177e4 LT |
119 | { |
120 | struct list_head *pos; | |
121 | ||
1da177e4 | 122 | list_for_each(pos, &clusterip_configs) { |
e905a9ed | 123 | struct clusterip_config *c = list_entry(pos, |
1da177e4 LT |
124 | struct clusterip_config, list); |
125 | if (c->clusterip == clusterip) { | |
126 | return c; | |
127 | } | |
128 | } | |
129 | ||
130 | return NULL; | |
131 | } | |
132 | ||
133 | static inline struct clusterip_config * | |
6a19d614 | 134 | clusterip_config_find_get(__be32 clusterip, int entry) |
1da177e4 LT |
135 | { |
136 | struct clusterip_config *c; | |
137 | ||
e45b1be8 | 138 | read_lock_bh(&clusterip_lock); |
1da177e4 LT |
139 | c = __clusterip_config_find(clusterip); |
140 | if (!c) { | |
e45b1be8 | 141 | read_unlock_bh(&clusterip_lock); |
1da177e4 LT |
142 | return NULL; |
143 | } | |
144 | atomic_inc(&c->refcount); | |
44513624 KK |
145 | if (entry) |
146 | atomic_inc(&c->entries); | |
e45b1be8 | 147 | read_unlock_bh(&clusterip_lock); |
1da177e4 LT |
148 | |
149 | return c; | |
150 | } | |
151 | ||
136e92bb KK |
152 | static void |
153 | clusterip_config_init_nodelist(struct clusterip_config *c, | |
154 | const struct ipt_clusterip_tgt_info *i) | |
155 | { | |
156 | int n; | |
157 | ||
158 | for (n = 0; n < i->num_local_nodes; n++) { | |
159 | set_bit(i->local_nodes[n] - 1, &c->local_nodes); | |
160 | } | |
161 | } | |
162 | ||
1da177e4 | 163 | static struct clusterip_config * |
6a19d614 | 164 | clusterip_config_init(struct ipt_clusterip_tgt_info *i, __be32 ip, |
1da177e4 LT |
165 | struct net_device *dev) |
166 | { | |
167 | struct clusterip_config *c; | |
1da177e4 | 168 | |
0da974f4 | 169 | c = kzalloc(sizeof(*c), GFP_ATOMIC); |
1da177e4 LT |
170 | if (!c) |
171 | return NULL; | |
172 | ||
1da177e4 LT |
173 | c->dev = dev; |
174 | c->clusterip = ip; | |
175 | memcpy(&c->clustermac, &i->clustermac, ETH_ALEN); | |
176 | c->num_total_nodes = i->num_total_nodes; | |
136e92bb | 177 | clusterip_config_init_nodelist(c, i); |
1da177e4 LT |
178 | c->hash_mode = i->hash_mode; |
179 | c->hash_initval = i->hash_initval; | |
180 | atomic_set(&c->refcount, 1); | |
44513624 | 181 | atomic_set(&c->entries, 1); |
1da177e4 LT |
182 | |
183 | #ifdef CONFIG_PROC_FS | |
76592584 PM |
184 | { |
185 | char buffer[16]; | |
186 | ||
187 | /* create proc dir entry */ | |
188 | sprintf(buffer, "%u.%u.%u.%u", NIPQUAD(ip)); | |
189 | c->pde = create_proc_entry(buffer, S_IWUSR|S_IRUSR, | |
190 | clusterip_procdir); | |
191 | if (!c->pde) { | |
192 | kfree(c); | |
193 | return NULL; | |
194 | } | |
1da177e4 LT |
195 | } |
196 | c->pde->proc_fops = &clusterip_proc_fops; | |
197 | c->pde->data = c; | |
198 | #endif | |
199 | ||
e45b1be8 | 200 | write_lock_bh(&clusterip_lock); |
1da177e4 | 201 | list_add(&c->list, &clusterip_configs); |
e45b1be8 | 202 | write_unlock_bh(&clusterip_lock); |
1da177e4 LT |
203 | |
204 | return c; | |
205 | } | |
206 | ||
76592584 | 207 | #ifdef CONFIG_PROC_FS |
1da177e4 LT |
208 | static int |
209 | clusterip_add_node(struct clusterip_config *c, u_int16_t nodenum) | |
210 | { | |
1da177e4 | 211 | |
136e92bb KK |
212 | if (nodenum == 0 || |
213 | nodenum > c->num_total_nodes) | |
1da177e4 | 214 | return 1; |
1da177e4 | 215 | |
136e92bb KK |
216 | /* check if we already have this number in our bitfield */ |
217 | if (test_and_set_bit(nodenum - 1, &c->local_nodes)) | |
218 | return 1; | |
1da177e4 | 219 | |
1da177e4 LT |
220 | return 0; |
221 | } | |
222 | ||
223 | static int | |
224 | clusterip_del_node(struct clusterip_config *c, u_int16_t nodenum) | |
225 | { | |
136e92bb KK |
226 | if (nodenum == 0 || |
227 | nodenum > c->num_total_nodes) | |
1da177e4 | 228 | return 1; |
e905a9ed | 229 | |
136e92bb KK |
230 | if (test_and_clear_bit(nodenum - 1, &c->local_nodes)) |
231 | return 0; | |
1da177e4 | 232 | |
1da177e4 LT |
233 | return 1; |
234 | } | |
76592584 | 235 | #endif |
1da177e4 LT |
236 | |
237 | static inline u_int32_t | |
238 | clusterip_hashfn(struct sk_buff *skb, struct clusterip_config *config) | |
239 | { | |
eddc9ec5 | 240 | struct iphdr *iph = ip_hdr(skb); |
1da177e4 LT |
241 | unsigned long hashval; |
242 | u_int16_t sport, dport; | |
957dc80a | 243 | u_int16_t *ports; |
1da177e4 LT |
244 | |
245 | switch (iph->protocol) { | |
246 | case IPPROTO_TCP: | |
1da177e4 | 247 | case IPPROTO_UDP: |
a8d0f952 | 248 | case IPPROTO_UDPLITE: |
957dc80a PM |
249 | case IPPROTO_SCTP: |
250 | case IPPROTO_DCCP: | |
1da177e4 | 251 | case IPPROTO_ICMP: |
957dc80a PM |
252 | ports = (void *)iph+iph->ihl*4; |
253 | sport = ports[0]; | |
254 | dport = ports[1]; | |
1da177e4 LT |
255 | break; |
256 | default: | |
257 | if (net_ratelimit()) { | |
258 | printk(KERN_NOTICE "CLUSTERIP: unknown protocol `%u'\n", | |
259 | iph->protocol); | |
260 | } | |
261 | sport = dport = 0; | |
262 | } | |
263 | ||
264 | switch (config->hash_mode) { | |
265 | case CLUSTERIP_HASHMODE_SIP: | |
266 | hashval = jhash_1word(ntohl(iph->saddr), | |
267 | config->hash_initval); | |
268 | break; | |
269 | case CLUSTERIP_HASHMODE_SIP_SPT: | |
e905a9ed | 270 | hashval = jhash_2words(ntohl(iph->saddr), sport, |
1da177e4 LT |
271 | config->hash_initval); |
272 | break; | |
273 | case CLUSTERIP_HASHMODE_SIP_SPT_DPT: | |
274 | hashval = jhash_3words(ntohl(iph->saddr), sport, dport, | |
275 | config->hash_initval); | |
276 | break; | |
277 | default: | |
278 | /* to make gcc happy */ | |
279 | hashval = 0; | |
280 | /* This cannot happen, unless the check function wasn't called | |
281 | * at rule load time */ | |
282 | printk("CLUSTERIP: unknown mode `%u'\n", config->hash_mode); | |
283 | BUG(); | |
284 | break; | |
285 | } | |
286 | ||
287 | /* node numbers are 1..n, not 0..n */ | |
288 | return ((hashval % config->num_total_nodes)+1); | |
289 | } | |
290 | ||
291 | static inline int | |
292 | clusterip_responsible(struct clusterip_config *config, u_int32_t hash) | |
293 | { | |
136e92bb | 294 | return test_bit(hash - 1, &config->local_nodes); |
1da177e4 LT |
295 | } |
296 | ||
e905a9ed YH |
297 | /*********************************************************************** |
298 | * IPTABLES TARGET | |
1da177e4 LT |
299 | ***********************************************************************/ |
300 | ||
301 | static unsigned int | |
302 | target(struct sk_buff **pskb, | |
303 | const struct net_device *in, | |
304 | const struct net_device *out, | |
305 | unsigned int hooknum, | |
c4986734 | 306 | const struct xt_target *target, |
fe1cb108 | 307 | const void *targinfo) |
1da177e4 LT |
308 | { |
309 | const struct ipt_clusterip_tgt_info *cipinfo = targinfo; | |
587aa641 | 310 | struct nf_conn *ct; |
1da177e4 | 311 | enum ip_conntrack_info ctinfo; |
587aa641 | 312 | u_int32_t hash; |
1da177e4 LT |
313 | |
314 | /* don't need to clusterip_config_get() here, since refcount | |
315 | * is only decremented by destroy() - and ip_tables guarantees | |
316 | * that the ->target() function isn't called after ->destroy() */ | |
317 | ||
587aa641 PM |
318 | ct = nf_ct_get(*pskb, &ctinfo); |
319 | if (ct == NULL) { | |
1da177e4 LT |
320 | printk(KERN_ERR "CLUSTERIP: no conntrack!\n"); |
321 | /* FIXME: need to drop invalid ones, since replies | |
e905a9ed | 322 | * to outgoing connections of other nodes will be |
1da177e4 LT |
323 | * marked as INVALID */ |
324 | return NF_DROP; | |
325 | } | |
326 | ||
327 | /* special case: ICMP error handling. conntrack distinguishes between | |
328 | * error messages (RELATED) and information requests (see below) */ | |
eddc9ec5 | 329 | if (ip_hdr(*pskb)->protocol == IPPROTO_ICMP |
e905a9ed | 330 | && (ctinfo == IP_CT_RELATED |
5d927eb0 | 331 | || ctinfo == IP_CT_RELATED+IP_CT_IS_REPLY)) |
6709dbbb | 332 | return XT_CONTINUE; |
1da177e4 | 333 | |
e905a9ed | 334 | /* ip_conntrack_icmp guarantees us that we only have ICMP_ECHO, |
1da177e4 LT |
335 | * TIMESTAMP, INFO_REQUEST or ADDRESS type icmp packets from here |
336 | * on, which all have an ID field [relevant for hashing]. */ | |
337 | ||
338 | hash = clusterip_hashfn(*pskb, cipinfo->config); | |
339 | ||
340 | switch (ctinfo) { | |
341 | case IP_CT_NEW: | |
587aa641 | 342 | ct->mark = hash; |
1da177e4 LT |
343 | break; |
344 | case IP_CT_RELATED: | |
345 | case IP_CT_RELATED+IP_CT_IS_REPLY: | |
346 | /* FIXME: we don't handle expectations at the | |
347 | * moment. they can arrive on a different node than | |
348 | * the master connection (e.g. FTP passive mode) */ | |
349 | case IP_CT_ESTABLISHED: | |
350 | case IP_CT_ESTABLISHED+IP_CT_IS_REPLY: | |
351 | break; | |
352 | default: | |
353 | break; | |
354 | } | |
355 | ||
356 | #ifdef DEBUG_CLUSTERP | |
357 | DUMP_TUPLE(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple); | |
358 | #endif | |
587aa641 | 359 | DEBUGP("hash=%u ct_hash=%u ", hash, ct->mark); |
1da177e4 LT |
360 | if (!clusterip_responsible(cipinfo->config, hash)) { |
361 | DEBUGP("not responsible\n"); | |
362 | return NF_DROP; | |
363 | } | |
364 | DEBUGP("responsible\n"); | |
365 | ||
366 | /* despite being received via linklayer multicast, this is | |
367 | * actually a unicast IP packet. TCP doesn't like PACKET_MULTICAST */ | |
368 | (*pskb)->pkt_type = PACKET_HOST; | |
369 | ||
6709dbbb | 370 | return XT_CONTINUE; |
1da177e4 LT |
371 | } |
372 | ||
373 | static int | |
374 | checkentry(const char *tablename, | |
2e4e6a17 | 375 | const void *e_void, |
c4986734 | 376 | const struct xt_target *target, |
e905a9ed YH |
377 | void *targinfo, |
378 | unsigned int hook_mask) | |
1da177e4 LT |
379 | { |
380 | struct ipt_clusterip_tgt_info *cipinfo = targinfo; | |
2e4e6a17 | 381 | const struct ipt_entry *e = e_void; |
1da177e4 LT |
382 | |
383 | struct clusterip_config *config; | |
384 | ||
1da177e4 LT |
385 | if (cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP && |
386 | cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP_SPT && | |
387 | cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP_SPT_DPT) { | |
388 | printk(KERN_WARNING "CLUSTERIP: unknown mode `%u'\n", | |
389 | cipinfo->hash_mode); | |
390 | return 0; | |
391 | ||
392 | } | |
6a19d614 | 393 | if (e->ip.dmsk.s_addr != htonl(0xffffffff) |
1da177e4 LT |
394 | || e->ip.dst.s_addr == 0) { |
395 | printk(KERN_ERR "CLUSTERIP: Please specify destination IP\n"); | |
396 | return 0; | |
397 | } | |
398 | ||
399 | /* FIXME: further sanity checks */ | |
400 | ||
44513624 KK |
401 | config = clusterip_config_find_get(e->ip.dst.s_addr, 1); |
402 | if (config) { | |
403 | if (cipinfo->config != NULL) { | |
404 | /* Case A: This is an entry that gets reloaded, since | |
405 | * it still has a cipinfo->config pointer. Simply | |
406 | * increase the entry refcount and return */ | |
407 | if (cipinfo->config != config) { | |
408 | printk(KERN_ERR "CLUSTERIP: Reloaded entry " | |
409 | "has invalid config pointer!\n"); | |
410 | return 0; | |
411 | } | |
44513624 KK |
412 | } else { |
413 | /* Case B: This is a new rule referring to an existing | |
414 | * clusterip config. */ | |
415 | cipinfo->config = config; | |
44513624 KK |
416 | } |
417 | } else { | |
418 | /* Case C: This is a completely new clusterip config */ | |
1da177e4 LT |
419 | if (!(cipinfo->flags & CLUSTERIP_FLAG_NEW)) { |
420 | printk(KERN_WARNING "CLUSTERIP: no config found for %u.%u.%u.%u, need 'new'\n", NIPQUAD(e->ip.dst.s_addr)); | |
421 | return 0; | |
422 | } else { | |
423 | struct net_device *dev; | |
424 | ||
425 | if (e->ip.iniface[0] == '\0') { | |
426 | printk(KERN_WARNING "CLUSTERIP: Please specify an interface name\n"); | |
427 | return 0; | |
428 | } | |
429 | ||
430 | dev = dev_get_by_name(e->ip.iniface); | |
431 | if (!dev) { | |
432 | printk(KERN_WARNING "CLUSTERIP: no such interface %s\n", e->ip.iniface); | |
433 | return 0; | |
434 | } | |
435 | ||
e905a9ed | 436 | config = clusterip_config_init(cipinfo, |
1da177e4 LT |
437 | e->ip.dst.s_addr, dev); |
438 | if (!config) { | |
439 | printk(KERN_WARNING "CLUSTERIP: cannot allocate config\n"); | |
440 | dev_put(dev); | |
441 | return 0; | |
442 | } | |
443 | dev_mc_add(config->dev,config->clustermac, ETH_ALEN, 0); | |
444 | } | |
44513624 | 445 | cipinfo->config = config; |
1da177e4 LT |
446 | } |
447 | ||
11078c37 YK |
448 | if (nf_ct_l3proto_try_module_get(target->family) < 0) { |
449 | printk(KERN_WARNING "can't load conntrack support for " | |
450 | "proto=%d\n", target->family); | |
451 | return 0; | |
452 | } | |
453 | ||
1da177e4 LT |
454 | return 1; |
455 | } | |
456 | ||
457 | /* drop reference count of cluster config when rule is deleted */ | |
efa74165 | 458 | static void destroy(const struct xt_target *target, void *targinfo) |
1da177e4 | 459 | { |
c4986734 | 460 | struct ipt_clusterip_tgt_info *cipinfo = targinfo; |
1da177e4 | 461 | |
44513624 KK |
462 | /* if no more entries are referencing the config, remove it |
463 | * from the list and destroy the proc entry */ | |
464 | clusterip_config_entry_put(cipinfo->config); | |
465 | ||
1da177e4 | 466 | clusterip_config_put(cipinfo->config); |
11078c37 YK |
467 | |
468 | nf_ct_l3proto_module_put(target->family); | |
1da177e4 LT |
469 | } |
470 | ||
6709dbbb | 471 | static struct xt_target clusterip_tgt = { |
1d5cd909 | 472 | .name = "CLUSTERIP", |
6709dbbb | 473 | .family = AF_INET, |
1d5cd909 PM |
474 | .target = target, |
475 | .targetsize = sizeof(struct ipt_clusterip_tgt_info), | |
476 | .checkentry = checkentry, | |
477 | .destroy = destroy, | |
478 | .me = THIS_MODULE | |
1da177e4 LT |
479 | }; |
480 | ||
481 | ||
e905a9ed YH |
482 | /*********************************************************************** |
483 | * ARP MANGLING CODE | |
1da177e4 LT |
484 | ***********************************************************************/ |
485 | ||
486 | /* hardcoded for 48bit ethernet and 32bit ipv4 addresses */ | |
487 | struct arp_payload { | |
488 | u_int8_t src_hw[ETH_ALEN]; | |
6a19d614 | 489 | __be32 src_ip; |
1da177e4 | 490 | u_int8_t dst_hw[ETH_ALEN]; |
6a19d614 | 491 | __be32 dst_ip; |
1da177e4 LT |
492 | } __attribute__ ((packed)); |
493 | ||
494 | #ifdef CLUSTERIP_DEBUG | |
e905a9ed | 495 | static void arp_print(struct arp_payload *payload) |
1da177e4 LT |
496 | { |
497 | #define HBUFFERLEN 30 | |
498 | char hbuffer[HBUFFERLEN]; | |
499 | int j,k; | |
500 | const char hexbuf[]= "0123456789abcdef"; | |
501 | ||
502 | for (k=0, j=0; k < HBUFFERLEN-3 && j < ETH_ALEN; j++) { | |
503 | hbuffer[k++]=hexbuf[(payload->src_hw[j]>>4)&15]; | |
504 | hbuffer[k++]=hexbuf[payload->src_hw[j]&15]; | |
505 | hbuffer[k++]=':'; | |
506 | } | |
507 | hbuffer[--k]='\0'; | |
508 | ||
e905a9ed | 509 | printk("src %u.%u.%u.%u@%s, dst %u.%u.%u.%u\n", |
1da177e4 LT |
510 | NIPQUAD(payload->src_ip), hbuffer, |
511 | NIPQUAD(payload->dst_ip)); | |
512 | } | |
513 | #endif | |
514 | ||
515 | static unsigned int | |
516 | arp_mangle(unsigned int hook, | |
517 | struct sk_buff **pskb, | |
518 | const struct net_device *in, | |
519 | const struct net_device *out, | |
520 | int (*okfn)(struct sk_buff *)) | |
521 | { | |
d0a92be0 | 522 | struct arphdr *arp = arp_hdr(*pskb); |
1da177e4 LT |
523 | struct arp_payload *payload; |
524 | struct clusterip_config *c; | |
525 | ||
526 | /* we don't care about non-ethernet and non-ipv4 ARP */ | |
527 | if (arp->ar_hrd != htons(ARPHRD_ETHER) | |
528 | || arp->ar_pro != htons(ETH_P_IP) | |
529 | || arp->ar_pln != 4 || arp->ar_hln != ETH_ALEN) | |
530 | return NF_ACCEPT; | |
531 | ||
4095ebf1 HW |
532 | /* we only want to mangle arp requests and replies */ |
533 | if (arp->ar_op != htons(ARPOP_REPLY) | |
534 | && arp->ar_op != htons(ARPOP_REQUEST)) | |
1da177e4 LT |
535 | return NF_ACCEPT; |
536 | ||
537 | payload = (void *)(arp+1); | |
538 | ||
e905a9ed | 539 | /* if there is no clusterip configuration for the arp reply's |
1da177e4 | 540 | * source ip, we don't want to mangle it */ |
44513624 | 541 | c = clusterip_config_find_get(payload->src_ip, 0); |
1da177e4 LT |
542 | if (!c) |
543 | return NF_ACCEPT; | |
544 | ||
e905a9ed | 545 | /* normally the linux kernel always replies to arp queries of |
1da177e4 LT |
546 | * addresses on different interfacs. However, in the CLUSTERIP case |
547 | * this wouldn't work, since we didn't subscribe the mcast group on | |
548 | * other interfaces */ | |
549 | if (c->dev != out) { | |
550 | DEBUGP("CLUSTERIP: not mangling arp reply on different " | |
551 | "interface: cip'%s'-skb'%s'\n", c->dev->name, out->name); | |
552 | clusterip_config_put(c); | |
553 | return NF_ACCEPT; | |
554 | } | |
555 | ||
556 | /* mangle reply hardware address */ | |
557 | memcpy(payload->src_hw, c->clustermac, arp->ar_hln); | |
558 | ||
559 | #ifdef CLUSTERIP_DEBUG | |
560 | DEBUGP(KERN_DEBUG "CLUSTERIP mangled arp reply: "); | |
561 | arp_print(payload); | |
562 | #endif | |
563 | ||
564 | clusterip_config_put(c); | |
565 | ||
566 | return NF_ACCEPT; | |
567 | } | |
568 | ||
569 | static struct nf_hook_ops cip_arp_ops = { | |
570 | .hook = arp_mangle, | |
571 | .pf = NF_ARP, | |
572 | .hooknum = NF_ARP_OUT, | |
573 | .priority = -1 | |
574 | }; | |
575 | ||
e905a9ed YH |
576 | /*********************************************************************** |
577 | * PROC DIR HANDLING | |
1da177e4 LT |
578 | ***********************************************************************/ |
579 | ||
580 | #ifdef CONFIG_PROC_FS | |
581 | ||
136e92bb KK |
582 | struct clusterip_seq_position { |
583 | unsigned int pos; /* position */ | |
584 | unsigned int weight; /* number of bits set == size */ | |
585 | unsigned int bit; /* current bit */ | |
586 | unsigned long val; /* current value */ | |
587 | }; | |
588 | ||
1da177e4 LT |
589 | static void *clusterip_seq_start(struct seq_file *s, loff_t *pos) |
590 | { | |
591 | struct proc_dir_entry *pde = s->private; | |
592 | struct clusterip_config *c = pde->data; | |
136e92bb KK |
593 | unsigned int weight; |
594 | u_int32_t local_nodes; | |
595 | struct clusterip_seq_position *idx; | |
596 | ||
597 | /* FIXME: possible race */ | |
598 | local_nodes = c->local_nodes; | |
599 | weight = hweight32(local_nodes); | |
600 | if (*pos >= weight) | |
1da177e4 LT |
601 | return NULL; |
602 | ||
136e92bb KK |
603 | idx = kmalloc(sizeof(struct clusterip_seq_position), GFP_KERNEL); |
604 | if (!idx) | |
1da177e4 LT |
605 | return ERR_PTR(-ENOMEM); |
606 | ||
136e92bb KK |
607 | idx->pos = *pos; |
608 | idx->weight = weight; | |
609 | idx->bit = ffs(local_nodes); | |
610 | idx->val = local_nodes; | |
611 | clear_bit(idx->bit - 1, &idx->val); | |
612 | ||
613 | return idx; | |
1da177e4 LT |
614 | } |
615 | ||
616 | static void *clusterip_seq_next(struct seq_file *s, void *v, loff_t *pos) | |
617 | { | |
136e92bb | 618 | struct clusterip_seq_position *idx = (struct clusterip_seq_position *)v; |
1da177e4 | 619 | |
136e92bb KK |
620 | *pos = ++idx->pos; |
621 | if (*pos >= idx->weight) { | |
1da177e4 LT |
622 | kfree(v); |
623 | return NULL; | |
624 | } | |
136e92bb KK |
625 | idx->bit = ffs(idx->val); |
626 | clear_bit(idx->bit - 1, &idx->val); | |
627 | return idx; | |
1da177e4 LT |
628 | } |
629 | ||
630 | static void clusterip_seq_stop(struct seq_file *s, void *v) | |
631 | { | |
632 | kfree(v); | |
1da177e4 LT |
633 | } |
634 | ||
635 | static int clusterip_seq_show(struct seq_file *s, void *v) | |
636 | { | |
136e92bb | 637 | struct clusterip_seq_position *idx = (struct clusterip_seq_position *)v; |
1da177e4 | 638 | |
e905a9ed | 639 | if (idx->pos != 0) |
1da177e4 | 640 | seq_putc(s, ','); |
1da177e4 | 641 | |
136e92bb KK |
642 | seq_printf(s, "%u", idx->bit); |
643 | ||
644 | if (idx->pos == idx->weight - 1) | |
1da177e4 LT |
645 | seq_putc(s, '\n'); |
646 | ||
647 | return 0; | |
648 | } | |
649 | ||
650 | static struct seq_operations clusterip_seq_ops = { | |
651 | .start = clusterip_seq_start, | |
652 | .next = clusterip_seq_next, | |
653 | .stop = clusterip_seq_stop, | |
654 | .show = clusterip_seq_show, | |
655 | }; | |
656 | ||
657 | static int clusterip_proc_open(struct inode *inode, struct file *file) | |
658 | { | |
659 | int ret = seq_open(file, &clusterip_seq_ops); | |
660 | ||
661 | if (!ret) { | |
662 | struct seq_file *sf = file->private_data; | |
663 | struct proc_dir_entry *pde = PDE(inode); | |
664 | struct clusterip_config *c = pde->data; | |
665 | ||
666 | sf->private = pde; | |
667 | ||
668 | clusterip_config_get(c); | |
669 | } | |
670 | ||
671 | return ret; | |
672 | } | |
673 | ||
674 | static int clusterip_proc_release(struct inode *inode, struct file *file) | |
675 | { | |
676 | struct proc_dir_entry *pde = PDE(inode); | |
677 | struct clusterip_config *c = pde->data; | |
678 | int ret; | |
679 | ||
680 | ret = seq_release(inode, file); | |
681 | ||
682 | if (!ret) | |
683 | clusterip_config_put(c); | |
684 | ||
685 | return ret; | |
686 | } | |
687 | ||
688 | static ssize_t clusterip_proc_write(struct file *file, const char __user *input, | |
689 | size_t size, loff_t *ofs) | |
690 | { | |
691 | #define PROC_WRITELEN 10 | |
692 | char buffer[PROC_WRITELEN+1]; | |
6df81ab2 | 693 | struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode); |
1da177e4 LT |
694 | struct clusterip_config *c = pde->data; |
695 | unsigned long nodenum; | |
696 | ||
697 | if (copy_from_user(buffer, input, PROC_WRITELEN)) | |
698 | return -EFAULT; | |
699 | ||
700 | if (*buffer == '+') { | |
701 | nodenum = simple_strtoul(buffer+1, NULL, 10); | |
702 | if (clusterip_add_node(c, nodenum)) | |
703 | return -ENOMEM; | |
704 | } else if (*buffer == '-') { | |
705 | nodenum = simple_strtoul(buffer+1, NULL,10); | |
706 | if (clusterip_del_node(c, nodenum)) | |
707 | return -ENOENT; | |
708 | } else | |
709 | return -EIO; | |
710 | ||
711 | return size; | |
712 | } | |
713 | ||
9a32144e | 714 | static const struct file_operations clusterip_proc_fops = { |
1da177e4 LT |
715 | .owner = THIS_MODULE, |
716 | .open = clusterip_proc_open, | |
717 | .read = seq_read, | |
718 | .write = clusterip_proc_write, | |
719 | .llseek = seq_lseek, | |
720 | .release = clusterip_proc_release, | |
721 | }; | |
722 | ||
723 | #endif /* CONFIG_PROC_FS */ | |
724 | ||
32292a7f | 725 | static int __init ipt_clusterip_init(void) |
1da177e4 LT |
726 | { |
727 | int ret; | |
728 | ||
6709dbbb | 729 | ret = xt_register_target(&clusterip_tgt); |
32292a7f PM |
730 | if (ret < 0) |
731 | return ret; | |
1da177e4 | 732 | |
32292a7f PM |
733 | ret = nf_register_hook(&cip_arp_ops); |
734 | if (ret < 0) | |
1da177e4 | 735 | goto cleanup_target; |
1da177e4 LT |
736 | |
737 | #ifdef CONFIG_PROC_FS | |
738 | clusterip_procdir = proc_mkdir("ipt_CLUSTERIP", proc_net); | |
739 | if (!clusterip_procdir) { | |
740 | printk(KERN_ERR "CLUSTERIP: Unable to proc dir entry\n"); | |
741 | ret = -ENOMEM; | |
742 | goto cleanup_hook; | |
743 | } | |
744 | #endif /* CONFIG_PROC_FS */ | |
745 | ||
746 | printk(KERN_NOTICE "ClusterIP Version %s loaded successfully\n", | |
747 | CLUSTERIP_VERSION); | |
1da177e4 LT |
748 | return 0; |
749 | ||
76592584 | 750 | #ifdef CONFIG_PROC_FS |
1da177e4 LT |
751 | cleanup_hook: |
752 | nf_unregister_hook(&cip_arp_ops); | |
76592584 | 753 | #endif /* CONFIG_PROC_FS */ |
1da177e4 | 754 | cleanup_target: |
6709dbbb | 755 | xt_unregister_target(&clusterip_tgt); |
32292a7f | 756 | return ret; |
1da177e4 LT |
757 | } |
758 | ||
65b4b4e8 | 759 | static void __exit ipt_clusterip_fini(void) |
1da177e4 | 760 | { |
32292a7f PM |
761 | printk(KERN_NOTICE "ClusterIP Version %s unloading\n", |
762 | CLUSTERIP_VERSION); | |
763 | #ifdef CONFIG_PROC_FS | |
764 | remove_proc_entry(clusterip_procdir->name, clusterip_procdir->parent); | |
765 | #endif | |
766 | nf_unregister_hook(&cip_arp_ops); | |
6709dbbb | 767 | xt_unregister_target(&clusterip_tgt); |
1da177e4 LT |
768 | } |
769 | ||
65b4b4e8 AM |
770 | module_init(ipt_clusterip_init); |
771 | module_exit(ipt_clusterip_fini); |