]>
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> | |
24 | ||
25 | #include <net/checksum.h> | |
26 | ||
27 | #include <linux/netfilter_arp.h> | |
28 | ||
6709dbbb | 29 | #include <linux/netfilter/x_tables.h> |
1da177e4 LT |
30 | #include <linux/netfilter_ipv4/ip_tables.h> |
31 | #include <linux/netfilter_ipv4/ipt_CLUSTERIP.h> | |
9fb9cbb1 | 32 | #include <net/netfilter/nf_conntrack_compat.h> |
1da177e4 | 33 | |
136e92bb | 34 | #define CLUSTERIP_VERSION "0.8" |
1da177e4 LT |
35 | |
36 | #define DEBUG_CLUSTERIP | |
37 | ||
38 | #ifdef DEBUG_CLUSTERIP | |
39 | #define DEBUGP printk | |
40 | #else | |
41 | #define DEBUGP | |
42 | #endif | |
43 | ||
44 | MODULE_LICENSE("GPL"); | |
45 | MODULE_AUTHOR("Harald Welte <laforge@netfilter.org>"); | |
46 | MODULE_DESCRIPTION("iptables target for CLUSTERIP"); | |
47 | ||
48 | struct clusterip_config { | |
49 | struct list_head list; /* list of all configs */ | |
50 | atomic_t refcount; /* reference count */ | |
44513624 KK |
51 | atomic_t entries; /* number of entries/rules |
52 | * referencing us */ | |
1da177e4 | 53 | |
6a19d614 | 54 | __be32 clusterip; /* the IP address */ |
1da177e4 LT |
55 | u_int8_t clustermac[ETH_ALEN]; /* the MAC address */ |
56 | struct net_device *dev; /* device */ | |
57 | u_int16_t num_total_nodes; /* total number of nodes */ | |
136e92bb | 58 | unsigned long local_nodes; /* node number array */ |
1da177e4 LT |
59 | |
60 | #ifdef CONFIG_PROC_FS | |
61 | struct proc_dir_entry *pde; /* proc dir entry */ | |
62 | #endif | |
63 | enum clusterip_hashmode hash_mode; /* which hashing mode */ | |
64 | u_int32_t hash_initval; /* hash initialization */ | |
65 | }; | |
66 | ||
67 | static LIST_HEAD(clusterip_configs); | |
68 | ||
136e92bb | 69 | /* clusterip_lock protects the clusterip_configs list */ |
e45b1be8 | 70 | static DEFINE_RWLOCK(clusterip_lock); |
1da177e4 LT |
71 | |
72 | #ifdef CONFIG_PROC_FS | |
9a32144e | 73 | static const struct file_operations clusterip_proc_fops; |
1da177e4 LT |
74 | static struct proc_dir_entry *clusterip_procdir; |
75 | #endif | |
76 | ||
77 | static inline void | |
44513624 KK |
78 | clusterip_config_get(struct clusterip_config *c) |
79 | { | |
1da177e4 LT |
80 | atomic_inc(&c->refcount); |
81 | } | |
82 | ||
83 | static inline void | |
44513624 KK |
84 | clusterip_config_put(struct clusterip_config *c) |
85 | { | |
86 | if (atomic_dec_and_test(&c->refcount)) | |
87 | kfree(c); | |
88 | } | |
89 | ||
90 | /* increase the count of entries(rules) using/referencing this config */ | |
91 | static inline void | |
92 | clusterip_config_entry_get(struct clusterip_config *c) | |
93 | { | |
94 | atomic_inc(&c->entries); | |
95 | } | |
96 | ||
97 | /* decrease the count of entries using/referencing this config. If last | |
98 | * entry(rule) is removed, remove the config from lists, but don't free it | |
99 | * yet, since proc-files could still be holding references */ | |
100 | static inline void | |
101 | clusterip_config_entry_put(struct clusterip_config *c) | |
102 | { | |
103 | if (atomic_dec_and_test(&c->entries)) { | |
e45b1be8 | 104 | write_lock_bh(&clusterip_lock); |
1da177e4 | 105 | list_del(&c->list); |
e45b1be8 | 106 | write_unlock_bh(&clusterip_lock); |
44513624 | 107 | |
1da177e4 LT |
108 | dev_mc_delete(c->dev, c->clustermac, ETH_ALEN, 0); |
109 | dev_put(c->dev); | |
44513624 KK |
110 | |
111 | /* In case anyone still accesses the file, the open/close | |
112 | * functions are also incrementing the refcount on their own, | |
113 | * so it's safe to remove the entry even if it's in use. */ | |
114 | #ifdef CONFIG_PROC_FS | |
115 | remove_proc_entry(c->pde->name, c->pde->parent); | |
116 | #endif | |
1da177e4 LT |
117 | } |
118 | } | |
119 | ||
1da177e4 | 120 | static struct clusterip_config * |
6a19d614 | 121 | __clusterip_config_find(__be32 clusterip) |
1da177e4 LT |
122 | { |
123 | struct list_head *pos; | |
124 | ||
1da177e4 | 125 | list_for_each(pos, &clusterip_configs) { |
e905a9ed | 126 | struct clusterip_config *c = list_entry(pos, |
1da177e4 LT |
127 | struct clusterip_config, list); |
128 | if (c->clusterip == clusterip) { | |
129 | return c; | |
130 | } | |
131 | } | |
132 | ||
133 | return NULL; | |
134 | } | |
135 | ||
136 | static inline struct clusterip_config * | |
6a19d614 | 137 | clusterip_config_find_get(__be32 clusterip, int entry) |
1da177e4 LT |
138 | { |
139 | struct clusterip_config *c; | |
140 | ||
e45b1be8 | 141 | read_lock_bh(&clusterip_lock); |
1da177e4 LT |
142 | c = __clusterip_config_find(clusterip); |
143 | if (!c) { | |
e45b1be8 | 144 | read_unlock_bh(&clusterip_lock); |
1da177e4 LT |
145 | return NULL; |
146 | } | |
147 | atomic_inc(&c->refcount); | |
44513624 KK |
148 | if (entry) |
149 | atomic_inc(&c->entries); | |
e45b1be8 | 150 | read_unlock_bh(&clusterip_lock); |
1da177e4 LT |
151 | |
152 | return c; | |
153 | } | |
154 | ||
136e92bb KK |
155 | static void |
156 | clusterip_config_init_nodelist(struct clusterip_config *c, | |
157 | const struct ipt_clusterip_tgt_info *i) | |
158 | { | |
159 | int n; | |
160 | ||
161 | for (n = 0; n < i->num_local_nodes; n++) { | |
162 | set_bit(i->local_nodes[n] - 1, &c->local_nodes); | |
163 | } | |
164 | } | |
165 | ||
1da177e4 | 166 | static struct clusterip_config * |
6a19d614 | 167 | clusterip_config_init(struct ipt_clusterip_tgt_info *i, __be32 ip, |
1da177e4 LT |
168 | struct net_device *dev) |
169 | { | |
170 | struct clusterip_config *c; | |
1da177e4 | 171 | |
0da974f4 | 172 | c = kzalloc(sizeof(*c), GFP_ATOMIC); |
1da177e4 LT |
173 | if (!c) |
174 | return NULL; | |
175 | ||
1da177e4 LT |
176 | c->dev = dev; |
177 | c->clusterip = ip; | |
178 | memcpy(&c->clustermac, &i->clustermac, ETH_ALEN); | |
179 | c->num_total_nodes = i->num_total_nodes; | |
136e92bb | 180 | clusterip_config_init_nodelist(c, i); |
1da177e4 LT |
181 | c->hash_mode = i->hash_mode; |
182 | c->hash_initval = i->hash_initval; | |
183 | atomic_set(&c->refcount, 1); | |
44513624 | 184 | atomic_set(&c->entries, 1); |
1da177e4 LT |
185 | |
186 | #ifdef CONFIG_PROC_FS | |
76592584 PM |
187 | { |
188 | char buffer[16]; | |
189 | ||
190 | /* create proc dir entry */ | |
191 | sprintf(buffer, "%u.%u.%u.%u", NIPQUAD(ip)); | |
192 | c->pde = create_proc_entry(buffer, S_IWUSR|S_IRUSR, | |
193 | clusterip_procdir); | |
194 | if (!c->pde) { | |
195 | kfree(c); | |
196 | return NULL; | |
197 | } | |
1da177e4 LT |
198 | } |
199 | c->pde->proc_fops = &clusterip_proc_fops; | |
200 | c->pde->data = c; | |
201 | #endif | |
202 | ||
e45b1be8 | 203 | write_lock_bh(&clusterip_lock); |
1da177e4 | 204 | list_add(&c->list, &clusterip_configs); |
e45b1be8 | 205 | write_unlock_bh(&clusterip_lock); |
1da177e4 LT |
206 | |
207 | return c; | |
208 | } | |
209 | ||
76592584 | 210 | #ifdef CONFIG_PROC_FS |
1da177e4 LT |
211 | static int |
212 | clusterip_add_node(struct clusterip_config *c, u_int16_t nodenum) | |
213 | { | |
1da177e4 | 214 | |
136e92bb KK |
215 | if (nodenum == 0 || |
216 | nodenum > c->num_total_nodes) | |
1da177e4 | 217 | return 1; |
1da177e4 | 218 | |
136e92bb KK |
219 | /* check if we already have this number in our bitfield */ |
220 | if (test_and_set_bit(nodenum - 1, &c->local_nodes)) | |
221 | return 1; | |
1da177e4 | 222 | |
1da177e4 LT |
223 | return 0; |
224 | } | |
225 | ||
226 | static int | |
227 | clusterip_del_node(struct clusterip_config *c, u_int16_t nodenum) | |
228 | { | |
136e92bb KK |
229 | if (nodenum == 0 || |
230 | nodenum > c->num_total_nodes) | |
1da177e4 | 231 | return 1; |
e905a9ed | 232 | |
136e92bb KK |
233 | if (test_and_clear_bit(nodenum - 1, &c->local_nodes)) |
234 | return 0; | |
1da177e4 | 235 | |
1da177e4 LT |
236 | return 1; |
237 | } | |
76592584 | 238 | #endif |
1da177e4 LT |
239 | |
240 | static inline u_int32_t | |
241 | clusterip_hashfn(struct sk_buff *skb, struct clusterip_config *config) | |
242 | { | |
243 | struct iphdr *iph = skb->nh.iph; | |
244 | unsigned long hashval; | |
245 | u_int16_t sport, dport; | |
957dc80a | 246 | u_int16_t *ports; |
1da177e4 LT |
247 | |
248 | switch (iph->protocol) { | |
249 | case IPPROTO_TCP: | |
1da177e4 | 250 | case IPPROTO_UDP: |
a8d0f952 | 251 | case IPPROTO_UDPLITE: |
957dc80a PM |
252 | case IPPROTO_SCTP: |
253 | case IPPROTO_DCCP: | |
1da177e4 | 254 | case IPPROTO_ICMP: |
957dc80a PM |
255 | ports = (void *)iph+iph->ihl*4; |
256 | sport = ports[0]; | |
257 | dport = ports[1]; | |
1da177e4 LT |
258 | break; |
259 | default: | |
260 | if (net_ratelimit()) { | |
261 | printk(KERN_NOTICE "CLUSTERIP: unknown protocol `%u'\n", | |
262 | iph->protocol); | |
263 | } | |
264 | sport = dport = 0; | |
265 | } | |
266 | ||
267 | switch (config->hash_mode) { | |
268 | case CLUSTERIP_HASHMODE_SIP: | |
269 | hashval = jhash_1word(ntohl(iph->saddr), | |
270 | config->hash_initval); | |
271 | break; | |
272 | case CLUSTERIP_HASHMODE_SIP_SPT: | |
e905a9ed | 273 | hashval = jhash_2words(ntohl(iph->saddr), sport, |
1da177e4 LT |
274 | config->hash_initval); |
275 | break; | |
276 | case CLUSTERIP_HASHMODE_SIP_SPT_DPT: | |
277 | hashval = jhash_3words(ntohl(iph->saddr), sport, dport, | |
278 | config->hash_initval); | |
279 | break; | |
280 | default: | |
281 | /* to make gcc happy */ | |
282 | hashval = 0; | |
283 | /* This cannot happen, unless the check function wasn't called | |
284 | * at rule load time */ | |
285 | printk("CLUSTERIP: unknown mode `%u'\n", config->hash_mode); | |
286 | BUG(); | |
287 | break; | |
288 | } | |
289 | ||
290 | /* node numbers are 1..n, not 0..n */ | |
291 | return ((hashval % config->num_total_nodes)+1); | |
292 | } | |
293 | ||
294 | static inline int | |
295 | clusterip_responsible(struct clusterip_config *config, u_int32_t hash) | |
296 | { | |
136e92bb | 297 | return test_bit(hash - 1, &config->local_nodes); |
1da177e4 LT |
298 | } |
299 | ||
e905a9ed YH |
300 | /*********************************************************************** |
301 | * IPTABLES TARGET | |
1da177e4 LT |
302 | ***********************************************************************/ |
303 | ||
304 | static unsigned int | |
305 | target(struct sk_buff **pskb, | |
306 | const struct net_device *in, | |
307 | const struct net_device *out, | |
308 | unsigned int hooknum, | |
c4986734 | 309 | const struct xt_target *target, |
fe1cb108 | 310 | const void *targinfo) |
1da177e4 LT |
311 | { |
312 | const struct ipt_clusterip_tgt_info *cipinfo = targinfo; | |
313 | enum ip_conntrack_info ctinfo; | |
9fb9cbb1 | 314 | u_int32_t *mark, hash; |
1da177e4 LT |
315 | |
316 | /* don't need to clusterip_config_get() here, since refcount | |
317 | * is only decremented by destroy() - and ip_tables guarantees | |
318 | * that the ->target() function isn't called after ->destroy() */ | |
319 | ||
9fb9cbb1 YK |
320 | mark = nf_ct_get_mark((*pskb), &ctinfo); |
321 | if (mark == NULL) { | |
1da177e4 LT |
322 | printk(KERN_ERR "CLUSTERIP: no conntrack!\n"); |
323 | /* FIXME: need to drop invalid ones, since replies | |
e905a9ed | 324 | * to outgoing connections of other nodes will be |
1da177e4 LT |
325 | * marked as INVALID */ |
326 | return NF_DROP; | |
327 | } | |
328 | ||
329 | /* special case: ICMP error handling. conntrack distinguishes between | |
330 | * error messages (RELATED) and information requests (see below) */ | |
331 | if ((*pskb)->nh.iph->protocol == IPPROTO_ICMP | |
e905a9ed | 332 | && (ctinfo == IP_CT_RELATED |
5d927eb0 | 333 | || ctinfo == IP_CT_RELATED+IP_CT_IS_REPLY)) |
6709dbbb | 334 | return XT_CONTINUE; |
1da177e4 | 335 | |
e905a9ed | 336 | /* ip_conntrack_icmp guarantees us that we only have ICMP_ECHO, |
1da177e4 LT |
337 | * TIMESTAMP, INFO_REQUEST or ADDRESS type icmp packets from here |
338 | * on, which all have an ID field [relevant for hashing]. */ | |
339 | ||
340 | hash = clusterip_hashfn(*pskb, cipinfo->config); | |
341 | ||
342 | switch (ctinfo) { | |
343 | case IP_CT_NEW: | |
9fb9cbb1 | 344 | *mark = hash; |
1da177e4 LT |
345 | break; |
346 | case IP_CT_RELATED: | |
347 | case IP_CT_RELATED+IP_CT_IS_REPLY: | |
348 | /* FIXME: we don't handle expectations at the | |
349 | * moment. they can arrive on a different node than | |
350 | * the master connection (e.g. FTP passive mode) */ | |
351 | case IP_CT_ESTABLISHED: | |
352 | case IP_CT_ESTABLISHED+IP_CT_IS_REPLY: | |
353 | break; | |
354 | default: | |
355 | break; | |
356 | } | |
357 | ||
358 | #ifdef DEBUG_CLUSTERP | |
359 | DUMP_TUPLE(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple); | |
360 | #endif | |
9fb9cbb1 | 361 | DEBUGP("hash=%u ct_hash=%u ", hash, *mark); |
1da177e4 LT |
362 | if (!clusterip_responsible(cipinfo->config, hash)) { |
363 | DEBUGP("not responsible\n"); | |
364 | return NF_DROP; | |
365 | } | |
366 | DEBUGP("responsible\n"); | |
367 | ||
368 | /* despite being received via linklayer multicast, this is | |
369 | * actually a unicast IP packet. TCP doesn't like PACKET_MULTICAST */ | |
370 | (*pskb)->pkt_type = PACKET_HOST; | |
371 | ||
6709dbbb | 372 | return XT_CONTINUE; |
1da177e4 LT |
373 | } |
374 | ||
375 | static int | |
376 | checkentry(const char *tablename, | |
2e4e6a17 | 377 | const void *e_void, |
c4986734 | 378 | const struct xt_target *target, |
e905a9ed YH |
379 | void *targinfo, |
380 | unsigned int hook_mask) | |
1da177e4 LT |
381 | { |
382 | struct ipt_clusterip_tgt_info *cipinfo = targinfo; | |
2e4e6a17 | 383 | const struct ipt_entry *e = e_void; |
1da177e4 LT |
384 | |
385 | struct clusterip_config *config; | |
386 | ||
1da177e4 LT |
387 | if (cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP && |
388 | cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP_SPT && | |
389 | cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP_SPT_DPT) { | |
390 | printk(KERN_WARNING "CLUSTERIP: unknown mode `%u'\n", | |
391 | cipinfo->hash_mode); | |
392 | return 0; | |
393 | ||
394 | } | |
6a19d614 | 395 | if (e->ip.dmsk.s_addr != htonl(0xffffffff) |
1da177e4 LT |
396 | || e->ip.dst.s_addr == 0) { |
397 | printk(KERN_ERR "CLUSTERIP: Please specify destination IP\n"); | |
398 | return 0; | |
399 | } | |
400 | ||
401 | /* FIXME: further sanity checks */ | |
402 | ||
44513624 KK |
403 | config = clusterip_config_find_get(e->ip.dst.s_addr, 1); |
404 | if (config) { | |
405 | if (cipinfo->config != NULL) { | |
406 | /* Case A: This is an entry that gets reloaded, since | |
407 | * it still has a cipinfo->config pointer. Simply | |
408 | * increase the entry refcount and return */ | |
409 | if (cipinfo->config != config) { | |
410 | printk(KERN_ERR "CLUSTERIP: Reloaded entry " | |
411 | "has invalid config pointer!\n"); | |
412 | return 0; | |
413 | } | |
414 | clusterip_config_entry_get(cipinfo->config); | |
415 | } else { | |
416 | /* Case B: This is a new rule referring to an existing | |
417 | * clusterip config. */ | |
418 | cipinfo->config = config; | |
419 | clusterip_config_entry_get(cipinfo->config); | |
420 | } | |
421 | } else { | |
422 | /* Case C: This is a completely new clusterip config */ | |
1da177e4 LT |
423 | if (!(cipinfo->flags & CLUSTERIP_FLAG_NEW)) { |
424 | printk(KERN_WARNING "CLUSTERIP: no config found for %u.%u.%u.%u, need 'new'\n", NIPQUAD(e->ip.dst.s_addr)); | |
425 | return 0; | |
426 | } else { | |
427 | struct net_device *dev; | |
428 | ||
429 | if (e->ip.iniface[0] == '\0') { | |
430 | printk(KERN_WARNING "CLUSTERIP: Please specify an interface name\n"); | |
431 | return 0; | |
432 | } | |
433 | ||
434 | dev = dev_get_by_name(e->ip.iniface); | |
435 | if (!dev) { | |
436 | printk(KERN_WARNING "CLUSTERIP: no such interface %s\n", e->ip.iniface); | |
437 | return 0; | |
438 | } | |
439 | ||
e905a9ed | 440 | config = clusterip_config_init(cipinfo, |
1da177e4 LT |
441 | e->ip.dst.s_addr, dev); |
442 | if (!config) { | |
443 | printk(KERN_WARNING "CLUSTERIP: cannot allocate config\n"); | |
444 | dev_put(dev); | |
445 | return 0; | |
446 | } | |
447 | dev_mc_add(config->dev,config->clustermac, ETH_ALEN, 0); | |
448 | } | |
44513624 | 449 | cipinfo->config = config; |
1da177e4 LT |
450 | } |
451 | ||
11078c37 YK |
452 | if (nf_ct_l3proto_try_module_get(target->family) < 0) { |
453 | printk(KERN_WARNING "can't load conntrack support for " | |
454 | "proto=%d\n", target->family); | |
455 | return 0; | |
456 | } | |
457 | ||
1da177e4 LT |
458 | return 1; |
459 | } | |
460 | ||
461 | /* drop reference count of cluster config when rule is deleted */ | |
efa74165 | 462 | static void destroy(const struct xt_target *target, void *targinfo) |
1da177e4 | 463 | { |
c4986734 | 464 | struct ipt_clusterip_tgt_info *cipinfo = targinfo; |
1da177e4 | 465 | |
44513624 KK |
466 | /* if no more entries are referencing the config, remove it |
467 | * from the list and destroy the proc entry */ | |
468 | clusterip_config_entry_put(cipinfo->config); | |
469 | ||
1da177e4 | 470 | clusterip_config_put(cipinfo->config); |
11078c37 YK |
471 | |
472 | nf_ct_l3proto_module_put(target->family); | |
1da177e4 LT |
473 | } |
474 | ||
6709dbbb | 475 | static struct xt_target clusterip_tgt = { |
1d5cd909 | 476 | .name = "CLUSTERIP", |
6709dbbb | 477 | .family = AF_INET, |
1d5cd909 PM |
478 | .target = target, |
479 | .targetsize = sizeof(struct ipt_clusterip_tgt_info), | |
480 | .checkentry = checkentry, | |
481 | .destroy = destroy, | |
482 | .me = THIS_MODULE | |
1da177e4 LT |
483 | }; |
484 | ||
485 | ||
e905a9ed YH |
486 | /*********************************************************************** |
487 | * ARP MANGLING CODE | |
1da177e4 LT |
488 | ***********************************************************************/ |
489 | ||
490 | /* hardcoded for 48bit ethernet and 32bit ipv4 addresses */ | |
491 | struct arp_payload { | |
492 | u_int8_t src_hw[ETH_ALEN]; | |
6a19d614 | 493 | __be32 src_ip; |
1da177e4 | 494 | u_int8_t dst_hw[ETH_ALEN]; |
6a19d614 | 495 | __be32 dst_ip; |
1da177e4 LT |
496 | } __attribute__ ((packed)); |
497 | ||
498 | #ifdef CLUSTERIP_DEBUG | |
e905a9ed | 499 | static void arp_print(struct arp_payload *payload) |
1da177e4 LT |
500 | { |
501 | #define HBUFFERLEN 30 | |
502 | char hbuffer[HBUFFERLEN]; | |
503 | int j,k; | |
504 | const char hexbuf[]= "0123456789abcdef"; | |
505 | ||
506 | for (k=0, j=0; k < HBUFFERLEN-3 && j < ETH_ALEN; j++) { | |
507 | hbuffer[k++]=hexbuf[(payload->src_hw[j]>>4)&15]; | |
508 | hbuffer[k++]=hexbuf[payload->src_hw[j]&15]; | |
509 | hbuffer[k++]=':'; | |
510 | } | |
511 | hbuffer[--k]='\0'; | |
512 | ||
e905a9ed | 513 | printk("src %u.%u.%u.%u@%s, dst %u.%u.%u.%u\n", |
1da177e4 LT |
514 | NIPQUAD(payload->src_ip), hbuffer, |
515 | NIPQUAD(payload->dst_ip)); | |
516 | } | |
517 | #endif | |
518 | ||
519 | static unsigned int | |
520 | arp_mangle(unsigned int hook, | |
521 | struct sk_buff **pskb, | |
522 | const struct net_device *in, | |
523 | const struct net_device *out, | |
524 | int (*okfn)(struct sk_buff *)) | |
525 | { | |
526 | struct arphdr *arp = (*pskb)->nh.arph; | |
527 | struct arp_payload *payload; | |
528 | struct clusterip_config *c; | |
529 | ||
530 | /* we don't care about non-ethernet and non-ipv4 ARP */ | |
531 | if (arp->ar_hrd != htons(ARPHRD_ETHER) | |
532 | || arp->ar_pro != htons(ETH_P_IP) | |
533 | || arp->ar_pln != 4 || arp->ar_hln != ETH_ALEN) | |
534 | return NF_ACCEPT; | |
535 | ||
4095ebf1 HW |
536 | /* we only want to mangle arp requests and replies */ |
537 | if (arp->ar_op != htons(ARPOP_REPLY) | |
538 | && arp->ar_op != htons(ARPOP_REQUEST)) | |
1da177e4 LT |
539 | return NF_ACCEPT; |
540 | ||
541 | payload = (void *)(arp+1); | |
542 | ||
e905a9ed | 543 | /* if there is no clusterip configuration for the arp reply's |
1da177e4 | 544 | * source ip, we don't want to mangle it */ |
44513624 | 545 | c = clusterip_config_find_get(payload->src_ip, 0); |
1da177e4 LT |
546 | if (!c) |
547 | return NF_ACCEPT; | |
548 | ||
e905a9ed | 549 | /* normally the linux kernel always replies to arp queries of |
1da177e4 LT |
550 | * addresses on different interfacs. However, in the CLUSTERIP case |
551 | * this wouldn't work, since we didn't subscribe the mcast group on | |
552 | * other interfaces */ | |
553 | if (c->dev != out) { | |
554 | DEBUGP("CLUSTERIP: not mangling arp reply on different " | |
555 | "interface: cip'%s'-skb'%s'\n", c->dev->name, out->name); | |
556 | clusterip_config_put(c); | |
557 | return NF_ACCEPT; | |
558 | } | |
559 | ||
560 | /* mangle reply hardware address */ | |
561 | memcpy(payload->src_hw, c->clustermac, arp->ar_hln); | |
562 | ||
563 | #ifdef CLUSTERIP_DEBUG | |
564 | DEBUGP(KERN_DEBUG "CLUSTERIP mangled arp reply: "); | |
565 | arp_print(payload); | |
566 | #endif | |
567 | ||
568 | clusterip_config_put(c); | |
569 | ||
570 | return NF_ACCEPT; | |
571 | } | |
572 | ||
573 | static struct nf_hook_ops cip_arp_ops = { | |
574 | .hook = arp_mangle, | |
575 | .pf = NF_ARP, | |
576 | .hooknum = NF_ARP_OUT, | |
577 | .priority = -1 | |
578 | }; | |
579 | ||
e905a9ed YH |
580 | /*********************************************************************** |
581 | * PROC DIR HANDLING | |
1da177e4 LT |
582 | ***********************************************************************/ |
583 | ||
584 | #ifdef CONFIG_PROC_FS | |
585 | ||
136e92bb KK |
586 | struct clusterip_seq_position { |
587 | unsigned int pos; /* position */ | |
588 | unsigned int weight; /* number of bits set == size */ | |
589 | unsigned int bit; /* current bit */ | |
590 | unsigned long val; /* current value */ | |
591 | }; | |
592 | ||
1da177e4 LT |
593 | static void *clusterip_seq_start(struct seq_file *s, loff_t *pos) |
594 | { | |
595 | struct proc_dir_entry *pde = s->private; | |
596 | struct clusterip_config *c = pde->data; | |
136e92bb KK |
597 | unsigned int weight; |
598 | u_int32_t local_nodes; | |
599 | struct clusterip_seq_position *idx; | |
600 | ||
601 | /* FIXME: possible race */ | |
602 | local_nodes = c->local_nodes; | |
603 | weight = hweight32(local_nodes); | |
604 | if (*pos >= weight) | |
1da177e4 LT |
605 | return NULL; |
606 | ||
136e92bb KK |
607 | idx = kmalloc(sizeof(struct clusterip_seq_position), GFP_KERNEL); |
608 | if (!idx) | |
1da177e4 LT |
609 | return ERR_PTR(-ENOMEM); |
610 | ||
136e92bb KK |
611 | idx->pos = *pos; |
612 | idx->weight = weight; | |
613 | idx->bit = ffs(local_nodes); | |
614 | idx->val = local_nodes; | |
615 | clear_bit(idx->bit - 1, &idx->val); | |
616 | ||
617 | return idx; | |
1da177e4 LT |
618 | } |
619 | ||
620 | static void *clusterip_seq_next(struct seq_file *s, void *v, loff_t *pos) | |
621 | { | |
136e92bb | 622 | struct clusterip_seq_position *idx = (struct clusterip_seq_position *)v; |
1da177e4 | 623 | |
136e92bb KK |
624 | *pos = ++idx->pos; |
625 | if (*pos >= idx->weight) { | |
1da177e4 LT |
626 | kfree(v); |
627 | return NULL; | |
628 | } | |
136e92bb KK |
629 | idx->bit = ffs(idx->val); |
630 | clear_bit(idx->bit - 1, &idx->val); | |
631 | return idx; | |
1da177e4 LT |
632 | } |
633 | ||
634 | static void clusterip_seq_stop(struct seq_file *s, void *v) | |
635 | { | |
636 | kfree(v); | |
1da177e4 LT |
637 | } |
638 | ||
639 | static int clusterip_seq_show(struct seq_file *s, void *v) | |
640 | { | |
136e92bb | 641 | struct clusterip_seq_position *idx = (struct clusterip_seq_position *)v; |
1da177e4 | 642 | |
e905a9ed | 643 | if (idx->pos != 0) |
1da177e4 | 644 | seq_putc(s, ','); |
1da177e4 | 645 | |
136e92bb KK |
646 | seq_printf(s, "%u", idx->bit); |
647 | ||
648 | if (idx->pos == idx->weight - 1) | |
1da177e4 LT |
649 | seq_putc(s, '\n'); |
650 | ||
651 | return 0; | |
652 | } | |
653 | ||
654 | static struct seq_operations clusterip_seq_ops = { | |
655 | .start = clusterip_seq_start, | |
656 | .next = clusterip_seq_next, | |
657 | .stop = clusterip_seq_stop, | |
658 | .show = clusterip_seq_show, | |
659 | }; | |
660 | ||
661 | static int clusterip_proc_open(struct inode *inode, struct file *file) | |
662 | { | |
663 | int ret = seq_open(file, &clusterip_seq_ops); | |
664 | ||
665 | if (!ret) { | |
666 | struct seq_file *sf = file->private_data; | |
667 | struct proc_dir_entry *pde = PDE(inode); | |
668 | struct clusterip_config *c = pde->data; | |
669 | ||
670 | sf->private = pde; | |
671 | ||
672 | clusterip_config_get(c); | |
673 | } | |
674 | ||
675 | return ret; | |
676 | } | |
677 | ||
678 | static int clusterip_proc_release(struct inode *inode, struct file *file) | |
679 | { | |
680 | struct proc_dir_entry *pde = PDE(inode); | |
681 | struct clusterip_config *c = pde->data; | |
682 | int ret; | |
683 | ||
684 | ret = seq_release(inode, file); | |
685 | ||
686 | if (!ret) | |
687 | clusterip_config_put(c); | |
688 | ||
689 | return ret; | |
690 | } | |
691 | ||
692 | static ssize_t clusterip_proc_write(struct file *file, const char __user *input, | |
693 | size_t size, loff_t *ofs) | |
694 | { | |
695 | #define PROC_WRITELEN 10 | |
696 | char buffer[PROC_WRITELEN+1]; | |
6df81ab2 | 697 | struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode); |
1da177e4 LT |
698 | struct clusterip_config *c = pde->data; |
699 | unsigned long nodenum; | |
700 | ||
701 | if (copy_from_user(buffer, input, PROC_WRITELEN)) | |
702 | return -EFAULT; | |
703 | ||
704 | if (*buffer == '+') { | |
705 | nodenum = simple_strtoul(buffer+1, NULL, 10); | |
706 | if (clusterip_add_node(c, nodenum)) | |
707 | return -ENOMEM; | |
708 | } else if (*buffer == '-') { | |
709 | nodenum = simple_strtoul(buffer+1, NULL,10); | |
710 | if (clusterip_del_node(c, nodenum)) | |
711 | return -ENOENT; | |
712 | } else | |
713 | return -EIO; | |
714 | ||
715 | return size; | |
716 | } | |
717 | ||
9a32144e | 718 | static const struct file_operations clusterip_proc_fops = { |
1da177e4 LT |
719 | .owner = THIS_MODULE, |
720 | .open = clusterip_proc_open, | |
721 | .read = seq_read, | |
722 | .write = clusterip_proc_write, | |
723 | .llseek = seq_lseek, | |
724 | .release = clusterip_proc_release, | |
725 | }; | |
726 | ||
727 | #endif /* CONFIG_PROC_FS */ | |
728 | ||
32292a7f | 729 | static int __init ipt_clusterip_init(void) |
1da177e4 LT |
730 | { |
731 | int ret; | |
732 | ||
6709dbbb | 733 | ret = xt_register_target(&clusterip_tgt); |
32292a7f PM |
734 | if (ret < 0) |
735 | return ret; | |
1da177e4 | 736 | |
32292a7f PM |
737 | ret = nf_register_hook(&cip_arp_ops); |
738 | if (ret < 0) | |
1da177e4 | 739 | goto cleanup_target; |
1da177e4 LT |
740 | |
741 | #ifdef CONFIG_PROC_FS | |
742 | clusterip_procdir = proc_mkdir("ipt_CLUSTERIP", proc_net); | |
743 | if (!clusterip_procdir) { | |
744 | printk(KERN_ERR "CLUSTERIP: Unable to proc dir entry\n"); | |
745 | ret = -ENOMEM; | |
746 | goto cleanup_hook; | |
747 | } | |
748 | #endif /* CONFIG_PROC_FS */ | |
749 | ||
750 | printk(KERN_NOTICE "ClusterIP Version %s loaded successfully\n", | |
751 | CLUSTERIP_VERSION); | |
1da177e4 LT |
752 | return 0; |
753 | ||
76592584 | 754 | #ifdef CONFIG_PROC_FS |
1da177e4 LT |
755 | cleanup_hook: |
756 | nf_unregister_hook(&cip_arp_ops); | |
76592584 | 757 | #endif /* CONFIG_PROC_FS */ |
1da177e4 | 758 | cleanup_target: |
6709dbbb | 759 | xt_unregister_target(&clusterip_tgt); |
32292a7f | 760 | return ret; |
1da177e4 LT |
761 | } |
762 | ||
65b4b4e8 | 763 | static void __exit ipt_clusterip_fini(void) |
1da177e4 | 764 | { |
32292a7f PM |
765 | printk(KERN_NOTICE "ClusterIP Version %s unloading\n", |
766 | CLUSTERIP_VERSION); | |
767 | #ifdef CONFIG_PROC_FS | |
768 | remove_proc_entry(clusterip_procdir->name, clusterip_procdir->parent); | |
769 | #endif | |
770 | nf_unregister_hook(&cip_arp_ops); | |
6709dbbb | 771 | xt_unregister_target(&clusterip_tgt); |
1da177e4 LT |
772 | } |
773 | ||
65b4b4e8 AM |
774 | module_init(ipt_clusterip_init); |
775 | module_exit(ipt_clusterip_fini); |