]>
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 | 32 | |
1da177e4 LT |
33 | MODULE_LICENSE("GPL"); |
34 | MODULE_AUTHOR("Harald Welte <laforge@netfilter.org>"); | |
35 | MODULE_DESCRIPTION("iptables target for CLUSTERIP"); | |
36 | ||
37 | struct clusterip_config { | |
38 | struct list_head list; /* list of all configs */ | |
39 | atomic_t refcount; /* reference count */ | |
44513624 KK |
40 | atomic_t entries; /* number of entries/rules |
41 | * referencing us */ | |
1da177e4 | 42 | |
6a19d614 | 43 | __be32 clusterip; /* the IP address */ |
1da177e4 LT |
44 | u_int8_t clustermac[ETH_ALEN]; /* the MAC address */ |
45 | struct net_device *dev; /* device */ | |
46 | u_int16_t num_total_nodes; /* total number of nodes */ | |
136e92bb | 47 | unsigned long local_nodes; /* node number array */ |
1da177e4 LT |
48 | |
49 | #ifdef CONFIG_PROC_FS | |
50 | struct proc_dir_entry *pde; /* proc dir entry */ | |
51 | #endif | |
52 | enum clusterip_hashmode hash_mode; /* which hashing mode */ | |
53 | u_int32_t hash_initval; /* hash initialization */ | |
54 | }; | |
55 | ||
56 | static LIST_HEAD(clusterip_configs); | |
57 | ||
136e92bb | 58 | /* clusterip_lock protects the clusterip_configs list */ |
e45b1be8 | 59 | static DEFINE_RWLOCK(clusterip_lock); |
1da177e4 LT |
60 | |
61 | #ifdef CONFIG_PROC_FS | |
9a32144e | 62 | static const struct file_operations clusterip_proc_fops; |
1da177e4 LT |
63 | static struct proc_dir_entry *clusterip_procdir; |
64 | #endif | |
65 | ||
66 | static inline void | |
44513624 KK |
67 | clusterip_config_get(struct clusterip_config *c) |
68 | { | |
1da177e4 LT |
69 | atomic_inc(&c->refcount); |
70 | } | |
71 | ||
72 | static inline void | |
44513624 KK |
73 | clusterip_config_put(struct clusterip_config *c) |
74 | { | |
75 | if (atomic_dec_and_test(&c->refcount)) | |
76 | kfree(c); | |
77 | } | |
78 | ||
79 | /* increase the count of entries(rules) using/referencing this config */ | |
80 | static inline void | |
81 | clusterip_config_entry_get(struct clusterip_config *c) | |
82 | { | |
83 | atomic_inc(&c->entries); | |
84 | } | |
85 | ||
86 | /* decrease the count of entries using/referencing this config. If last | |
87 | * entry(rule) is removed, remove the config from lists, but don't free it | |
88 | * yet, since proc-files could still be holding references */ | |
89 | static inline void | |
90 | clusterip_config_entry_put(struct clusterip_config *c) | |
91 | { | |
92 | if (atomic_dec_and_test(&c->entries)) { | |
e45b1be8 | 93 | write_lock_bh(&clusterip_lock); |
1da177e4 | 94 | list_del(&c->list); |
e45b1be8 | 95 | write_unlock_bh(&clusterip_lock); |
44513624 | 96 | |
1da177e4 LT |
97 | dev_mc_delete(c->dev, c->clustermac, ETH_ALEN, 0); |
98 | dev_put(c->dev); | |
44513624 KK |
99 | |
100 | /* In case anyone still accesses the file, the open/close | |
101 | * functions are also incrementing the refcount on their own, | |
102 | * so it's safe to remove the entry even if it's in use. */ | |
103 | #ifdef CONFIG_PROC_FS | |
104 | remove_proc_entry(c->pde->name, c->pde->parent); | |
105 | #endif | |
1da177e4 LT |
106 | } |
107 | } | |
108 | ||
1da177e4 | 109 | static struct clusterip_config * |
6a19d614 | 110 | __clusterip_config_find(__be32 clusterip) |
1da177e4 LT |
111 | { |
112 | struct list_head *pos; | |
113 | ||
1da177e4 | 114 | list_for_each(pos, &clusterip_configs) { |
e905a9ed | 115 | struct clusterip_config *c = list_entry(pos, |
1da177e4 | 116 | struct clusterip_config, list); |
7c4e36bc | 117 | if (c->clusterip == clusterip) |
1da177e4 | 118 | return c; |
1da177e4 LT |
119 | } |
120 | ||
121 | return NULL; | |
122 | } | |
123 | ||
124 | static inline struct clusterip_config * | |
6a19d614 | 125 | clusterip_config_find_get(__be32 clusterip, int entry) |
1da177e4 LT |
126 | { |
127 | struct clusterip_config *c; | |
128 | ||
e45b1be8 | 129 | read_lock_bh(&clusterip_lock); |
1da177e4 LT |
130 | c = __clusterip_config_find(clusterip); |
131 | if (!c) { | |
e45b1be8 | 132 | read_unlock_bh(&clusterip_lock); |
1da177e4 LT |
133 | return NULL; |
134 | } | |
135 | atomic_inc(&c->refcount); | |
44513624 KK |
136 | if (entry) |
137 | atomic_inc(&c->entries); | |
e45b1be8 | 138 | read_unlock_bh(&clusterip_lock); |
1da177e4 LT |
139 | |
140 | return c; | |
141 | } | |
142 | ||
136e92bb KK |
143 | static void |
144 | clusterip_config_init_nodelist(struct clusterip_config *c, | |
145 | const struct ipt_clusterip_tgt_info *i) | |
146 | { | |
147 | int n; | |
148 | ||
7c4e36bc | 149 | for (n = 0; n < i->num_local_nodes; n++) |
136e92bb | 150 | set_bit(i->local_nodes[n] - 1, &c->local_nodes); |
136e92bb KK |
151 | } |
152 | ||
1da177e4 | 153 | static struct clusterip_config * |
6a19d614 | 154 | clusterip_config_init(struct ipt_clusterip_tgt_info *i, __be32 ip, |
1da177e4 LT |
155 | struct net_device *dev) |
156 | { | |
157 | struct clusterip_config *c; | |
1da177e4 | 158 | |
0da974f4 | 159 | c = kzalloc(sizeof(*c), GFP_ATOMIC); |
1da177e4 LT |
160 | if (!c) |
161 | return NULL; | |
162 | ||
1da177e4 LT |
163 | c->dev = dev; |
164 | c->clusterip = ip; | |
165 | memcpy(&c->clustermac, &i->clustermac, ETH_ALEN); | |
166 | c->num_total_nodes = i->num_total_nodes; | |
136e92bb | 167 | clusterip_config_init_nodelist(c, i); |
1da177e4 LT |
168 | c->hash_mode = i->hash_mode; |
169 | c->hash_initval = i->hash_initval; | |
170 | atomic_set(&c->refcount, 1); | |
44513624 | 171 | atomic_set(&c->entries, 1); |
1da177e4 LT |
172 | |
173 | #ifdef CONFIG_PROC_FS | |
76592584 PM |
174 | { |
175 | char buffer[16]; | |
176 | ||
177 | /* create proc dir entry */ | |
178 | sprintf(buffer, "%u.%u.%u.%u", NIPQUAD(ip)); | |
179 | c->pde = create_proc_entry(buffer, S_IWUSR|S_IRUSR, | |
180 | clusterip_procdir); | |
181 | if (!c->pde) { | |
182 | kfree(c); | |
183 | return NULL; | |
184 | } | |
1da177e4 LT |
185 | } |
186 | c->pde->proc_fops = &clusterip_proc_fops; | |
187 | c->pde->data = c; | |
188 | #endif | |
189 | ||
e45b1be8 | 190 | write_lock_bh(&clusterip_lock); |
1da177e4 | 191 | list_add(&c->list, &clusterip_configs); |
e45b1be8 | 192 | write_unlock_bh(&clusterip_lock); |
1da177e4 LT |
193 | |
194 | return c; | |
195 | } | |
196 | ||
76592584 | 197 | #ifdef CONFIG_PROC_FS |
1da177e4 LT |
198 | static int |
199 | clusterip_add_node(struct clusterip_config *c, u_int16_t nodenum) | |
200 | { | |
1da177e4 | 201 | |
136e92bb KK |
202 | if (nodenum == 0 || |
203 | nodenum > c->num_total_nodes) | |
1da177e4 | 204 | return 1; |
1da177e4 | 205 | |
136e92bb KK |
206 | /* check if we already have this number in our bitfield */ |
207 | if (test_and_set_bit(nodenum - 1, &c->local_nodes)) | |
208 | return 1; | |
1da177e4 | 209 | |
1da177e4 LT |
210 | return 0; |
211 | } | |
212 | ||
e1931b78 | 213 | static bool |
1da177e4 LT |
214 | clusterip_del_node(struct clusterip_config *c, u_int16_t nodenum) |
215 | { | |
136e92bb KK |
216 | if (nodenum == 0 || |
217 | nodenum > c->num_total_nodes) | |
e1931b78 | 218 | return true; |
e905a9ed | 219 | |
136e92bb | 220 | if (test_and_clear_bit(nodenum - 1, &c->local_nodes)) |
e1931b78 | 221 | return false; |
1da177e4 | 222 | |
e1931b78 | 223 | return true; |
1da177e4 | 224 | } |
76592584 | 225 | #endif |
1da177e4 LT |
226 | |
227 | static inline u_int32_t | |
a47362a2 JE |
228 | clusterip_hashfn(const struct sk_buff *skb, |
229 | const struct clusterip_config *config) | |
1da177e4 | 230 | { |
a47362a2 | 231 | const struct iphdr *iph = ip_hdr(skb); |
1da177e4 LT |
232 | unsigned long hashval; |
233 | u_int16_t sport, dport; | |
a47362a2 | 234 | const u_int16_t *ports; |
1da177e4 LT |
235 | |
236 | switch (iph->protocol) { | |
237 | case IPPROTO_TCP: | |
1da177e4 | 238 | case IPPROTO_UDP: |
a8d0f952 | 239 | case IPPROTO_UDPLITE: |
957dc80a PM |
240 | case IPPROTO_SCTP: |
241 | case IPPROTO_DCCP: | |
1da177e4 | 242 | case IPPROTO_ICMP: |
a47362a2 | 243 | ports = (const void *)iph+iph->ihl*4; |
957dc80a PM |
244 | sport = ports[0]; |
245 | dport = ports[1]; | |
1da177e4 LT |
246 | break; |
247 | default: | |
7c4e36bc | 248 | if (net_ratelimit()) |
1da177e4 LT |
249 | printk(KERN_NOTICE "CLUSTERIP: unknown protocol `%u'\n", |
250 | iph->protocol); | |
1da177e4 LT |
251 | sport = dport = 0; |
252 | } | |
253 | ||
254 | switch (config->hash_mode) { | |
255 | case CLUSTERIP_HASHMODE_SIP: | |
256 | hashval = jhash_1word(ntohl(iph->saddr), | |
257 | config->hash_initval); | |
258 | break; | |
259 | case CLUSTERIP_HASHMODE_SIP_SPT: | |
e905a9ed | 260 | hashval = jhash_2words(ntohl(iph->saddr), sport, |
1da177e4 LT |
261 | config->hash_initval); |
262 | break; | |
263 | case CLUSTERIP_HASHMODE_SIP_SPT_DPT: | |
264 | hashval = jhash_3words(ntohl(iph->saddr), sport, dport, | |
265 | config->hash_initval); | |
266 | break; | |
267 | default: | |
268 | /* to make gcc happy */ | |
269 | hashval = 0; | |
270 | /* This cannot happen, unless the check function wasn't called | |
271 | * at rule load time */ | |
272 | printk("CLUSTERIP: unknown mode `%u'\n", config->hash_mode); | |
273 | BUG(); | |
274 | break; | |
275 | } | |
276 | ||
277 | /* node numbers are 1..n, not 0..n */ | |
7c4e36bc | 278 | return (hashval % config->num_total_nodes) + 1; |
1da177e4 LT |
279 | } |
280 | ||
281 | static inline int | |
a47362a2 | 282 | clusterip_responsible(const struct clusterip_config *config, u_int32_t hash) |
1da177e4 | 283 | { |
136e92bb | 284 | return test_bit(hash - 1, &config->local_nodes); |
1da177e4 LT |
285 | } |
286 | ||
e905a9ed YH |
287 | /*********************************************************************** |
288 | * IPTABLES TARGET | |
1da177e4 LT |
289 | ***********************************************************************/ |
290 | ||
291 | static unsigned int | |
292 | target(struct sk_buff **pskb, | |
293 | const struct net_device *in, | |
294 | const struct net_device *out, | |
295 | unsigned int hooknum, | |
c4986734 | 296 | const struct xt_target *target, |
fe1cb108 | 297 | const void *targinfo) |
1da177e4 LT |
298 | { |
299 | const struct ipt_clusterip_tgt_info *cipinfo = targinfo; | |
587aa641 | 300 | struct nf_conn *ct; |
1da177e4 | 301 | enum ip_conntrack_info ctinfo; |
587aa641 | 302 | u_int32_t hash; |
1da177e4 LT |
303 | |
304 | /* don't need to clusterip_config_get() here, since refcount | |
305 | * is only decremented by destroy() - and ip_tables guarantees | |
306 | * that the ->target() function isn't called after ->destroy() */ | |
307 | ||
587aa641 PM |
308 | ct = nf_ct_get(*pskb, &ctinfo); |
309 | if (ct == NULL) { | |
1da177e4 LT |
310 | printk(KERN_ERR "CLUSTERIP: no conntrack!\n"); |
311 | /* FIXME: need to drop invalid ones, since replies | |
e905a9ed | 312 | * to outgoing connections of other nodes will be |
1da177e4 LT |
313 | * marked as INVALID */ |
314 | return NF_DROP; | |
315 | } | |
316 | ||
317 | /* special case: ICMP error handling. conntrack distinguishes between | |
318 | * error messages (RELATED) and information requests (see below) */ | |
eddc9ec5 | 319 | if (ip_hdr(*pskb)->protocol == IPPROTO_ICMP |
e905a9ed | 320 | && (ctinfo == IP_CT_RELATED |
5d927eb0 | 321 | || ctinfo == IP_CT_RELATED+IP_CT_IS_REPLY)) |
6709dbbb | 322 | return XT_CONTINUE; |
1da177e4 | 323 | |
e905a9ed | 324 | /* ip_conntrack_icmp guarantees us that we only have ICMP_ECHO, |
1da177e4 LT |
325 | * TIMESTAMP, INFO_REQUEST or ADDRESS type icmp packets from here |
326 | * on, which all have an ID field [relevant for hashing]. */ | |
327 | ||
328 | hash = clusterip_hashfn(*pskb, cipinfo->config); | |
329 | ||
330 | switch (ctinfo) { | |
331 | case IP_CT_NEW: | |
587aa641 | 332 | ct->mark = hash; |
1da177e4 LT |
333 | break; |
334 | case IP_CT_RELATED: | |
335 | case IP_CT_RELATED+IP_CT_IS_REPLY: | |
336 | /* FIXME: we don't handle expectations at the | |
337 | * moment. they can arrive on a different node than | |
338 | * the master connection (e.g. FTP passive mode) */ | |
339 | case IP_CT_ESTABLISHED: | |
340 | case IP_CT_ESTABLISHED+IP_CT_IS_REPLY: | |
341 | break; | |
342 | default: | |
343 | break; | |
344 | } | |
345 | ||
0d53778e | 346 | #ifdef DEBUG |
1da177e4 LT |
347 | DUMP_TUPLE(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple); |
348 | #endif | |
0d53778e | 349 | pr_debug("hash=%u ct_hash=%u ", hash, ct->mark); |
1da177e4 | 350 | if (!clusterip_responsible(cipinfo->config, hash)) { |
0d53778e | 351 | pr_debug("not responsible\n"); |
1da177e4 LT |
352 | return NF_DROP; |
353 | } | |
0d53778e | 354 | pr_debug("responsible\n"); |
1da177e4 LT |
355 | |
356 | /* despite being received via linklayer multicast, this is | |
357 | * actually a unicast IP packet. TCP doesn't like PACKET_MULTICAST */ | |
358 | (*pskb)->pkt_type = PACKET_HOST; | |
359 | ||
6709dbbb | 360 | return XT_CONTINUE; |
1da177e4 LT |
361 | } |
362 | ||
e1931b78 | 363 | static bool |
1da177e4 | 364 | checkentry(const char *tablename, |
2e4e6a17 | 365 | const void *e_void, |
c4986734 | 366 | const struct xt_target *target, |
e905a9ed YH |
367 | void *targinfo, |
368 | unsigned int hook_mask) | |
1da177e4 LT |
369 | { |
370 | struct ipt_clusterip_tgt_info *cipinfo = targinfo; | |
2e4e6a17 | 371 | const struct ipt_entry *e = e_void; |
1da177e4 LT |
372 | |
373 | struct clusterip_config *config; | |
374 | ||
1da177e4 LT |
375 | if (cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP && |
376 | cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP_SPT && | |
377 | cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP_SPT_DPT) { | |
378 | printk(KERN_WARNING "CLUSTERIP: unknown mode `%u'\n", | |
379 | cipinfo->hash_mode); | |
e1931b78 | 380 | return false; |
1da177e4 LT |
381 | |
382 | } | |
6a19d614 | 383 | if (e->ip.dmsk.s_addr != htonl(0xffffffff) |
1da177e4 LT |
384 | || e->ip.dst.s_addr == 0) { |
385 | printk(KERN_ERR "CLUSTERIP: Please specify destination IP\n"); | |
e1931b78 | 386 | return false; |
1da177e4 LT |
387 | } |
388 | ||
389 | /* FIXME: further sanity checks */ | |
390 | ||
44513624 | 391 | config = clusterip_config_find_get(e->ip.dst.s_addr, 1); |
d3c3f424 | 392 | if (!config) { |
1da177e4 LT |
393 | if (!(cipinfo->flags & CLUSTERIP_FLAG_NEW)) { |
394 | printk(KERN_WARNING "CLUSTERIP: no config found for %u.%u.%u.%u, need 'new'\n", NIPQUAD(e->ip.dst.s_addr)); | |
e1931b78 | 395 | return false; |
1da177e4 LT |
396 | } else { |
397 | struct net_device *dev; | |
398 | ||
399 | if (e->ip.iniface[0] == '\0') { | |
400 | printk(KERN_WARNING "CLUSTERIP: Please specify an interface name\n"); | |
e1931b78 | 401 | return false; |
1da177e4 LT |
402 | } |
403 | ||
404 | dev = dev_get_by_name(e->ip.iniface); | |
405 | if (!dev) { | |
406 | printk(KERN_WARNING "CLUSTERIP: no such interface %s\n", e->ip.iniface); | |
e1931b78 | 407 | return false; |
1da177e4 LT |
408 | } |
409 | ||
e905a9ed | 410 | config = clusterip_config_init(cipinfo, |
1da177e4 LT |
411 | e->ip.dst.s_addr, dev); |
412 | if (!config) { | |
413 | printk(KERN_WARNING "CLUSTERIP: cannot allocate config\n"); | |
414 | dev_put(dev); | |
e1931b78 | 415 | return false; |
1da177e4 LT |
416 | } |
417 | dev_mc_add(config->dev,config->clustermac, ETH_ALEN, 0); | |
418 | } | |
419 | } | |
d3c3f424 | 420 | cipinfo->config = config; |
1da177e4 | 421 | |
11078c37 YK |
422 | if (nf_ct_l3proto_try_module_get(target->family) < 0) { |
423 | printk(KERN_WARNING "can't load conntrack support for " | |
424 | "proto=%d\n", target->family); | |
e1931b78 | 425 | return false; |
11078c37 YK |
426 | } |
427 | ||
e1931b78 | 428 | return true; |
1da177e4 LT |
429 | } |
430 | ||
431 | /* drop reference count of cluster config when rule is deleted */ | |
efa74165 | 432 | static void destroy(const struct xt_target *target, void *targinfo) |
1da177e4 | 433 | { |
c4986734 | 434 | struct ipt_clusterip_tgt_info *cipinfo = targinfo; |
1da177e4 | 435 | |
44513624 KK |
436 | /* if no more entries are referencing the config, remove it |
437 | * from the list and destroy the proc entry */ | |
438 | clusterip_config_entry_put(cipinfo->config); | |
439 | ||
1da177e4 | 440 | clusterip_config_put(cipinfo->config); |
11078c37 YK |
441 | |
442 | nf_ct_l3proto_module_put(target->family); | |
1da177e4 LT |
443 | } |
444 | ||
d3c3f424 PM |
445 | #ifdef CONFIG_COMPAT |
446 | struct compat_ipt_clusterip_tgt_info | |
447 | { | |
448 | u_int32_t flags; | |
449 | u_int8_t clustermac[6]; | |
450 | u_int16_t num_total_nodes; | |
451 | u_int16_t num_local_nodes; | |
452 | u_int16_t local_nodes[CLUSTERIP_MAX_NODES]; | |
453 | u_int32_t hash_mode; | |
454 | u_int32_t hash_initval; | |
455 | compat_uptr_t config; | |
456 | }; | |
457 | #endif /* CONFIG_COMPAT */ | |
458 | ||
9f15c530 | 459 | static struct xt_target clusterip_tgt __read_mostly = { |
1d5cd909 | 460 | .name = "CLUSTERIP", |
6709dbbb | 461 | .family = AF_INET, |
1d5cd909 | 462 | .target = target, |
1d5cd909 PM |
463 | .checkentry = checkentry, |
464 | .destroy = destroy, | |
d3c3f424 PM |
465 | .targetsize = sizeof(struct ipt_clusterip_tgt_info), |
466 | #ifdef CONFIG_COMPAT | |
467 | .compatsize = sizeof(struct compat_ipt_clusterip_tgt_info), | |
468 | #endif /* CONFIG_COMPAT */ | |
1d5cd909 | 469 | .me = THIS_MODULE |
1da177e4 LT |
470 | }; |
471 | ||
472 | ||
e905a9ed YH |
473 | /*********************************************************************** |
474 | * ARP MANGLING CODE | |
1da177e4 LT |
475 | ***********************************************************************/ |
476 | ||
477 | /* hardcoded for 48bit ethernet and 32bit ipv4 addresses */ | |
478 | struct arp_payload { | |
479 | u_int8_t src_hw[ETH_ALEN]; | |
6a19d614 | 480 | __be32 src_ip; |
1da177e4 | 481 | u_int8_t dst_hw[ETH_ALEN]; |
6a19d614 | 482 | __be32 dst_ip; |
1da177e4 LT |
483 | } __attribute__ ((packed)); |
484 | ||
0d53778e | 485 | #ifdef DEBUG |
e905a9ed | 486 | static void arp_print(struct arp_payload *payload) |
1da177e4 LT |
487 | { |
488 | #define HBUFFERLEN 30 | |
489 | char hbuffer[HBUFFERLEN]; | |
490 | int j,k; | |
491 | const char hexbuf[]= "0123456789abcdef"; | |
492 | ||
493 | for (k=0, j=0; k < HBUFFERLEN-3 && j < ETH_ALEN; j++) { | |
494 | hbuffer[k++]=hexbuf[(payload->src_hw[j]>>4)&15]; | |
495 | hbuffer[k++]=hexbuf[payload->src_hw[j]&15]; | |
496 | hbuffer[k++]=':'; | |
497 | } | |
498 | hbuffer[--k]='\0'; | |
499 | ||
e905a9ed | 500 | printk("src %u.%u.%u.%u@%s, dst %u.%u.%u.%u\n", |
1da177e4 LT |
501 | NIPQUAD(payload->src_ip), hbuffer, |
502 | NIPQUAD(payload->dst_ip)); | |
503 | } | |
504 | #endif | |
505 | ||
506 | static unsigned int | |
507 | arp_mangle(unsigned int hook, | |
508 | struct sk_buff **pskb, | |
509 | const struct net_device *in, | |
510 | const struct net_device *out, | |
511 | int (*okfn)(struct sk_buff *)) | |
512 | { | |
d0a92be0 | 513 | struct arphdr *arp = arp_hdr(*pskb); |
1da177e4 LT |
514 | struct arp_payload *payload; |
515 | struct clusterip_config *c; | |
516 | ||
517 | /* we don't care about non-ethernet and non-ipv4 ARP */ | |
518 | if (arp->ar_hrd != htons(ARPHRD_ETHER) | |
519 | || arp->ar_pro != htons(ETH_P_IP) | |
520 | || arp->ar_pln != 4 || arp->ar_hln != ETH_ALEN) | |
521 | return NF_ACCEPT; | |
522 | ||
4095ebf1 HW |
523 | /* we only want to mangle arp requests and replies */ |
524 | if (arp->ar_op != htons(ARPOP_REPLY) | |
525 | && arp->ar_op != htons(ARPOP_REQUEST)) | |
1da177e4 LT |
526 | return NF_ACCEPT; |
527 | ||
528 | payload = (void *)(arp+1); | |
529 | ||
e905a9ed | 530 | /* if there is no clusterip configuration for the arp reply's |
1da177e4 | 531 | * source ip, we don't want to mangle it */ |
44513624 | 532 | c = clusterip_config_find_get(payload->src_ip, 0); |
1da177e4 LT |
533 | if (!c) |
534 | return NF_ACCEPT; | |
535 | ||
e905a9ed | 536 | /* normally the linux kernel always replies to arp queries of |
1da177e4 LT |
537 | * addresses on different interfacs. However, in the CLUSTERIP case |
538 | * this wouldn't work, since we didn't subscribe the mcast group on | |
539 | * other interfaces */ | |
540 | if (c->dev != out) { | |
0d53778e PM |
541 | pr_debug("CLUSTERIP: not mangling arp reply on different " |
542 | "interface: cip'%s'-skb'%s'\n", | |
543 | c->dev->name, out->name); | |
1da177e4 LT |
544 | clusterip_config_put(c); |
545 | return NF_ACCEPT; | |
546 | } | |
547 | ||
548 | /* mangle reply hardware address */ | |
549 | memcpy(payload->src_hw, c->clustermac, arp->ar_hln); | |
550 | ||
0d53778e PM |
551 | #ifdef DEBUG |
552 | pr_debug(KERN_DEBUG "CLUSTERIP mangled arp reply: "); | |
1da177e4 LT |
553 | arp_print(payload); |
554 | #endif | |
555 | ||
556 | clusterip_config_put(c); | |
557 | ||
558 | return NF_ACCEPT; | |
559 | } | |
560 | ||
561 | static struct nf_hook_ops cip_arp_ops = { | |
562 | .hook = arp_mangle, | |
563 | .pf = NF_ARP, | |
564 | .hooknum = NF_ARP_OUT, | |
565 | .priority = -1 | |
566 | }; | |
567 | ||
e905a9ed YH |
568 | /*********************************************************************** |
569 | * PROC DIR HANDLING | |
1da177e4 LT |
570 | ***********************************************************************/ |
571 | ||
572 | #ifdef CONFIG_PROC_FS | |
573 | ||
136e92bb KK |
574 | struct clusterip_seq_position { |
575 | unsigned int pos; /* position */ | |
576 | unsigned int weight; /* number of bits set == size */ | |
577 | unsigned int bit; /* current bit */ | |
578 | unsigned long val; /* current value */ | |
579 | }; | |
580 | ||
1da177e4 LT |
581 | static void *clusterip_seq_start(struct seq_file *s, loff_t *pos) |
582 | { | |
583 | struct proc_dir_entry *pde = s->private; | |
584 | struct clusterip_config *c = pde->data; | |
136e92bb KK |
585 | unsigned int weight; |
586 | u_int32_t local_nodes; | |
587 | struct clusterip_seq_position *idx; | |
588 | ||
589 | /* FIXME: possible race */ | |
590 | local_nodes = c->local_nodes; | |
591 | weight = hweight32(local_nodes); | |
592 | if (*pos >= weight) | |
1da177e4 LT |
593 | return NULL; |
594 | ||
136e92bb KK |
595 | idx = kmalloc(sizeof(struct clusterip_seq_position), GFP_KERNEL); |
596 | if (!idx) | |
1da177e4 LT |
597 | return ERR_PTR(-ENOMEM); |
598 | ||
136e92bb KK |
599 | idx->pos = *pos; |
600 | idx->weight = weight; | |
601 | idx->bit = ffs(local_nodes); | |
602 | idx->val = local_nodes; | |
603 | clear_bit(idx->bit - 1, &idx->val); | |
604 | ||
605 | return idx; | |
1da177e4 LT |
606 | } |
607 | ||
608 | static void *clusterip_seq_next(struct seq_file *s, void *v, loff_t *pos) | |
609 | { | |
136e92bb | 610 | struct clusterip_seq_position *idx = (struct clusterip_seq_position *)v; |
1da177e4 | 611 | |
136e92bb KK |
612 | *pos = ++idx->pos; |
613 | if (*pos >= idx->weight) { | |
1da177e4 LT |
614 | kfree(v); |
615 | return NULL; | |
616 | } | |
136e92bb KK |
617 | idx->bit = ffs(idx->val); |
618 | clear_bit(idx->bit - 1, &idx->val); | |
619 | return idx; | |
1da177e4 LT |
620 | } |
621 | ||
622 | static void clusterip_seq_stop(struct seq_file *s, void *v) | |
623 | { | |
624 | kfree(v); | |
1da177e4 LT |
625 | } |
626 | ||
627 | static int clusterip_seq_show(struct seq_file *s, void *v) | |
628 | { | |
136e92bb | 629 | struct clusterip_seq_position *idx = (struct clusterip_seq_position *)v; |
1da177e4 | 630 | |
e905a9ed | 631 | if (idx->pos != 0) |
1da177e4 | 632 | seq_putc(s, ','); |
1da177e4 | 633 | |
136e92bb KK |
634 | seq_printf(s, "%u", idx->bit); |
635 | ||
636 | if (idx->pos == idx->weight - 1) | |
1da177e4 LT |
637 | seq_putc(s, '\n'); |
638 | ||
639 | return 0; | |
640 | } | |
641 | ||
642 | static struct seq_operations clusterip_seq_ops = { | |
643 | .start = clusterip_seq_start, | |
644 | .next = clusterip_seq_next, | |
645 | .stop = clusterip_seq_stop, | |
646 | .show = clusterip_seq_show, | |
647 | }; | |
648 | ||
649 | static int clusterip_proc_open(struct inode *inode, struct file *file) | |
650 | { | |
651 | int ret = seq_open(file, &clusterip_seq_ops); | |
652 | ||
653 | if (!ret) { | |
654 | struct seq_file *sf = file->private_data; | |
655 | struct proc_dir_entry *pde = PDE(inode); | |
656 | struct clusterip_config *c = pde->data; | |
657 | ||
658 | sf->private = pde; | |
659 | ||
660 | clusterip_config_get(c); | |
661 | } | |
662 | ||
663 | return ret; | |
664 | } | |
665 | ||
666 | static int clusterip_proc_release(struct inode *inode, struct file *file) | |
667 | { | |
668 | struct proc_dir_entry *pde = PDE(inode); | |
669 | struct clusterip_config *c = pde->data; | |
670 | int ret; | |
671 | ||
672 | ret = seq_release(inode, file); | |
673 | ||
674 | if (!ret) | |
675 | clusterip_config_put(c); | |
676 | ||
677 | return ret; | |
678 | } | |
679 | ||
680 | static ssize_t clusterip_proc_write(struct file *file, const char __user *input, | |
681 | size_t size, loff_t *ofs) | |
682 | { | |
683 | #define PROC_WRITELEN 10 | |
684 | char buffer[PROC_WRITELEN+1]; | |
6df81ab2 | 685 | struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode); |
1da177e4 LT |
686 | struct clusterip_config *c = pde->data; |
687 | unsigned long nodenum; | |
688 | ||
689 | if (copy_from_user(buffer, input, PROC_WRITELEN)) | |
690 | return -EFAULT; | |
691 | ||
692 | if (*buffer == '+') { | |
693 | nodenum = simple_strtoul(buffer+1, NULL, 10); | |
694 | if (clusterip_add_node(c, nodenum)) | |
695 | return -ENOMEM; | |
696 | } else if (*buffer == '-') { | |
697 | nodenum = simple_strtoul(buffer+1, NULL,10); | |
698 | if (clusterip_del_node(c, nodenum)) | |
699 | return -ENOENT; | |
700 | } else | |
701 | return -EIO; | |
702 | ||
703 | return size; | |
704 | } | |
705 | ||
9a32144e | 706 | static const struct file_operations clusterip_proc_fops = { |
1da177e4 LT |
707 | .owner = THIS_MODULE, |
708 | .open = clusterip_proc_open, | |
709 | .read = seq_read, | |
710 | .write = clusterip_proc_write, | |
711 | .llseek = seq_lseek, | |
712 | .release = clusterip_proc_release, | |
713 | }; | |
714 | ||
715 | #endif /* CONFIG_PROC_FS */ | |
716 | ||
32292a7f | 717 | static int __init ipt_clusterip_init(void) |
1da177e4 LT |
718 | { |
719 | int ret; | |
720 | ||
6709dbbb | 721 | ret = xt_register_target(&clusterip_tgt); |
32292a7f PM |
722 | if (ret < 0) |
723 | return ret; | |
1da177e4 | 724 | |
32292a7f PM |
725 | ret = nf_register_hook(&cip_arp_ops); |
726 | if (ret < 0) | |
1da177e4 | 727 | goto cleanup_target; |
1da177e4 LT |
728 | |
729 | #ifdef CONFIG_PROC_FS | |
730 | clusterip_procdir = proc_mkdir("ipt_CLUSTERIP", proc_net); | |
731 | if (!clusterip_procdir) { | |
732 | printk(KERN_ERR "CLUSTERIP: Unable to proc dir entry\n"); | |
733 | ret = -ENOMEM; | |
734 | goto cleanup_hook; | |
735 | } | |
736 | #endif /* CONFIG_PROC_FS */ | |
737 | ||
738 | printk(KERN_NOTICE "ClusterIP Version %s loaded successfully\n", | |
739 | CLUSTERIP_VERSION); | |
1da177e4 LT |
740 | return 0; |
741 | ||
76592584 | 742 | #ifdef CONFIG_PROC_FS |
1da177e4 LT |
743 | cleanup_hook: |
744 | nf_unregister_hook(&cip_arp_ops); | |
76592584 | 745 | #endif /* CONFIG_PROC_FS */ |
1da177e4 | 746 | cleanup_target: |
6709dbbb | 747 | xt_unregister_target(&clusterip_tgt); |
32292a7f | 748 | return ret; |
1da177e4 LT |
749 | } |
750 | ||
65b4b4e8 | 751 | static void __exit ipt_clusterip_fini(void) |
1da177e4 | 752 | { |
32292a7f PM |
753 | printk(KERN_NOTICE "ClusterIP Version %s unloading\n", |
754 | CLUSTERIP_VERSION); | |
755 | #ifdef CONFIG_PROC_FS | |
756 | remove_proc_entry(clusterip_procdir->name, clusterip_procdir->parent); | |
757 | #endif | |
758 | nf_unregister_hook(&cip_arp_ops); | |
6709dbbb | 759 | xt_unregister_target(&clusterip_tgt); |
1da177e4 LT |
760 | } |
761 | ||
65b4b4e8 AM |
762 | module_init(ipt_clusterip_init); |
763 | module_exit(ipt_clusterip_fini); |