]>
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 | */ | |
ff67e4e4 | 12 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
1da177e4 | 13 | #include <linux/module.h> |
1da177e4 LT |
14 | #include <linux/proc_fs.h> |
15 | #include <linux/jhash.h> | |
136e92bb | 16 | #include <linux/bitops.h> |
1da177e4 | 17 | #include <linux/skbuff.h> |
5a0e3ad6 | 18 | #include <linux/slab.h> |
1da177e4 LT |
19 | #include <linux/ip.h> |
20 | #include <linux/tcp.h> | |
21 | #include <linux/udp.h> | |
22 | #include <linux/icmp.h> | |
23 | #include <linux/if_arp.h> | |
1da177e4 | 24 | #include <linux/seq_file.h> |
1da177e4 | 25 | #include <linux/netfilter_arp.h> |
6709dbbb | 26 | #include <linux/netfilter/x_tables.h> |
1da177e4 LT |
27 | #include <linux/netfilter_ipv4/ip_tables.h> |
28 | #include <linux/netfilter_ipv4/ipt_CLUSTERIP.h> | |
587aa641 | 29 | #include <net/netfilter/nf_conntrack.h> |
457c4cbc | 30 | #include <net/net_namespace.h> |
ce4ff76c | 31 | #include <net/netns/generic.h> |
587aa641 | 32 | #include <net/checksum.h> |
3d04ebb6 | 33 | #include <net/ip.h> |
1da177e4 | 34 | |
136e92bb | 35 | #define CLUSTERIP_VERSION "0.8" |
1da177e4 | 36 | |
1da177e4 LT |
37 | MODULE_LICENSE("GPL"); |
38 | MODULE_AUTHOR("Harald Welte <laforge@netfilter.org>"); | |
2ae15b64 | 39 | MODULE_DESCRIPTION("Xtables: CLUSTERIP target"); |
1da177e4 LT |
40 | |
41 | struct clusterip_config { | |
42 | struct list_head list; /* list of all configs */ | |
43 | atomic_t refcount; /* reference count */ | |
44513624 KK |
44 | atomic_t entries; /* number of entries/rules |
45 | * referencing us */ | |
1da177e4 | 46 | |
6a19d614 | 47 | __be32 clusterip; /* the IP address */ |
1da177e4 LT |
48 | u_int8_t clustermac[ETH_ALEN]; /* the MAC address */ |
49 | struct net_device *dev; /* device */ | |
50 | u_int16_t num_total_nodes; /* total number of nodes */ | |
136e92bb | 51 | unsigned long local_nodes; /* node number array */ |
1da177e4 LT |
52 | |
53 | #ifdef CONFIG_PROC_FS | |
54 | struct proc_dir_entry *pde; /* proc dir entry */ | |
55 | #endif | |
56 | enum clusterip_hashmode hash_mode; /* which hashing mode */ | |
57 | u_int32_t hash_initval; /* hash initialization */ | |
d73f33b1 | 58 | struct rcu_head rcu; |
1da177e4 LT |
59 | }; |
60 | ||
1da177e4 | 61 | #ifdef CONFIG_PROC_FS |
9a32144e | 62 | static const struct file_operations clusterip_proc_fops; |
1da177e4 | 63 | #endif |
1da177e4 | 64 | |
c7d03a00 | 65 | static unsigned int clusterip_net_id __read_mostly; |
ce4ff76c G |
66 | |
67 | struct clusterip_net { | |
26a89e43 | 68 | struct list_head configs; |
f1e8077f G |
69 | /* lock protects the configs list */ |
70 | spinlock_t lock; | |
1da177e4 LT |
71 | |
72 | #ifdef CONFIG_PROC_FS | |
ce4ff76c | 73 | struct proc_dir_entry *procdir; |
1da177e4 | 74 | #endif |
ce4ff76c | 75 | }; |
1da177e4 LT |
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 | ||
d73f33b1 ED |
83 | |
84 | static void clusterip_config_rcu_free(struct rcu_head *head) | |
85 | { | |
86 | kfree(container_of(head, struct clusterip_config, rcu)); | |
87 | } | |
88 | ||
1da177e4 | 89 | static inline void |
44513624 KK |
90 | clusterip_config_put(struct clusterip_config *c) |
91 | { | |
92 | if (atomic_dec_and_test(&c->refcount)) | |
d73f33b1 | 93 | call_rcu_bh(&c->rcu, clusterip_config_rcu_free); |
44513624 KK |
94 | } |
95 | ||
44513624 KK |
96 | /* decrease the count of entries using/referencing this config. If last |
97 | * entry(rule) is removed, remove the config from lists, but don't free it | |
98 | * yet, since proc-files could still be holding references */ | |
99 | static inline void | |
100 | clusterip_config_entry_put(struct clusterip_config *c) | |
101 | { | |
d86946d2 G |
102 | struct net *net = dev_net(c->dev); |
103 | struct clusterip_net *cn = net_generic(net, clusterip_net_id); | |
f1e8077f | 104 | |
d73f33b1 | 105 | local_bh_disable(); |
f1e8077f | 106 | if (atomic_dec_and_lock(&c->entries, &cn->lock)) { |
d73f33b1 | 107 | list_del_rcu(&c->list); |
f1e8077f | 108 | spin_unlock(&cn->lock); |
d73f33b1 | 109 | local_bh_enable(); |
44513624 | 110 | |
22bedad3 | 111 | dev_mc_del(c->dev, c->clustermac); |
1da177e4 | 112 | dev_put(c->dev); |
44513624 KK |
113 | |
114 | /* In case anyone still accesses the file, the open/close | |
115 | * functions are also incrementing the refcount on their own, | |
116 | * so it's safe to remove the entry even if it's in use. */ | |
117 | #ifdef CONFIG_PROC_FS | |
a8ca16ea | 118 | proc_remove(c->pde); |
44513624 | 119 | #endif |
4dee9597 | 120 | return; |
1da177e4 | 121 | } |
d73f33b1 | 122 | local_bh_enable(); |
1da177e4 LT |
123 | } |
124 | ||
1da177e4 | 125 | static struct clusterip_config * |
b5ef0f85 | 126 | __clusterip_config_find(struct net *net, __be32 clusterip) |
1da177e4 | 127 | { |
4c610979 | 128 | struct clusterip_config *c; |
b5ef0f85 | 129 | struct clusterip_net *cn = net_generic(net, clusterip_net_id); |
1da177e4 | 130 | |
26a89e43 | 131 | list_for_each_entry_rcu(c, &cn->configs, list) { |
7c4e36bc | 132 | if (c->clusterip == clusterip) |
1da177e4 | 133 | return c; |
1da177e4 LT |
134 | } |
135 | ||
136 | return NULL; | |
137 | } | |
138 | ||
139 | static inline struct clusterip_config * | |
b5ef0f85 | 140 | clusterip_config_find_get(struct net *net, __be32 clusterip, int entry) |
1da177e4 LT |
141 | { |
142 | struct clusterip_config *c; | |
143 | ||
d73f33b1 | 144 | rcu_read_lock_bh(); |
b5ef0f85 | 145 | c = __clusterip_config_find(net, clusterip); |
d73f33b1 | 146 | if (c) { |
3fd0b634 AB |
147 | #ifdef CONFIG_PROC_FS |
148 | if (!c->pde) | |
149 | c = NULL; | |
150 | else | |
151 | #endif | |
152 | if (unlikely(!atomic_inc_not_zero(&c->refcount))) | |
d73f33b1 ED |
153 | c = NULL; |
154 | else if (entry) | |
155 | atomic_inc(&c->entries); | |
1da177e4 | 156 | } |
d73f33b1 | 157 | rcu_read_unlock_bh(); |
1da177e4 LT |
158 | |
159 | return c; | |
160 | } | |
161 | ||
136e92bb KK |
162 | static void |
163 | clusterip_config_init_nodelist(struct clusterip_config *c, | |
164 | const struct ipt_clusterip_tgt_info *i) | |
165 | { | |
166 | int n; | |
167 | ||
7c4e36bc | 168 | for (n = 0; n < i->num_local_nodes; n++) |
136e92bb | 169 | set_bit(i->local_nodes[n] - 1, &c->local_nodes); |
136e92bb KK |
170 | } |
171 | ||
1da177e4 | 172 | static struct clusterip_config * |
3cf93c96 | 173 | clusterip_config_init(const struct ipt_clusterip_tgt_info *i, __be32 ip, |
6c5d5cfb | 174 | struct net_device *dev) |
1da177e4 | 175 | { |
6c5d5cfb | 176 | struct net *net = dev_net(dev); |
1da177e4 | 177 | struct clusterip_config *c; |
6c5d5cfb | 178 | struct clusterip_net *cn = net_generic(net, clusterip_net_id); |
1da177e4 | 179 | |
0da974f4 | 180 | c = kzalloc(sizeof(*c), GFP_ATOMIC); |
1da177e4 | 181 | if (!c) |
6c5d5cfb | 182 | return ERR_PTR(-ENOMEM); |
1da177e4 | 183 | |
1da177e4 LT |
184 | c->dev = dev; |
185 | c->clusterip = ip; | |
186 | memcpy(&c->clustermac, &i->clustermac, ETH_ALEN); | |
187 | c->num_total_nodes = i->num_total_nodes; | |
136e92bb | 188 | clusterip_config_init_nodelist(c, i); |
1da177e4 LT |
189 | c->hash_mode = i->hash_mode; |
190 | c->hash_initval = i->hash_initval; | |
191 | atomic_set(&c->refcount, 1); | |
44513624 | 192 | atomic_set(&c->entries, 1); |
1da177e4 | 193 | |
6c5d5cfb XL |
194 | spin_lock_bh(&cn->lock); |
195 | if (__clusterip_config_find(net, ip)) { | |
196 | spin_unlock_bh(&cn->lock); | |
197 | kfree(c); | |
198 | ||
199 | return ERR_PTR(-EBUSY); | |
200 | } | |
201 | ||
202 | list_add_rcu(&c->list, &cn->configs); | |
203 | spin_unlock_bh(&cn->lock); | |
204 | ||
1da177e4 | 205 | #ifdef CONFIG_PROC_FS |
76592584 PM |
206 | { |
207 | char buffer[16]; | |
208 | ||
209 | /* create proc dir entry */ | |
cffee385 | 210 | sprintf(buffer, "%pI4", &ip); |
6e79d85d | 211 | c->pde = proc_create_data(buffer, S_IWUSR|S_IRUSR, |
ce4ff76c | 212 | cn->procdir, |
6e79d85d | 213 | &clusterip_proc_fops, c); |
76592584 | 214 | if (!c->pde) { |
6c5d5cfb XL |
215 | spin_lock_bh(&cn->lock); |
216 | list_del_rcu(&c->list); | |
217 | spin_unlock_bh(&cn->lock); | |
76592584 | 218 | kfree(c); |
6c5d5cfb XL |
219 | |
220 | return ERR_PTR(-ENOMEM); | |
76592584 | 221 | } |
1da177e4 | 222 | } |
1da177e4 LT |
223 | #endif |
224 | ||
1da177e4 LT |
225 | return c; |
226 | } | |
227 | ||
76592584 | 228 | #ifdef CONFIG_PROC_FS |
1da177e4 LT |
229 | static int |
230 | clusterip_add_node(struct clusterip_config *c, u_int16_t nodenum) | |
231 | { | |
1da177e4 | 232 | |
136e92bb KK |
233 | if (nodenum == 0 || |
234 | nodenum > c->num_total_nodes) | |
1da177e4 | 235 | return 1; |
1da177e4 | 236 | |
136e92bb KK |
237 | /* check if we already have this number in our bitfield */ |
238 | if (test_and_set_bit(nodenum - 1, &c->local_nodes)) | |
239 | return 1; | |
1da177e4 | 240 | |
1da177e4 LT |
241 | return 0; |
242 | } | |
243 | ||
e1931b78 | 244 | static bool |
1da177e4 LT |
245 | clusterip_del_node(struct clusterip_config *c, u_int16_t nodenum) |
246 | { | |
136e92bb KK |
247 | if (nodenum == 0 || |
248 | nodenum > c->num_total_nodes) | |
e1931b78 | 249 | return true; |
e905a9ed | 250 | |
136e92bb | 251 | if (test_and_clear_bit(nodenum - 1, &c->local_nodes)) |
e1931b78 | 252 | return false; |
1da177e4 | 253 | |
e1931b78 | 254 | return true; |
1da177e4 | 255 | } |
76592584 | 256 | #endif |
1da177e4 LT |
257 | |
258 | static inline u_int32_t | |
a47362a2 JE |
259 | clusterip_hashfn(const struct sk_buff *skb, |
260 | const struct clusterip_config *config) | |
1da177e4 | 261 | { |
a47362a2 | 262 | const struct iphdr *iph = ip_hdr(skb); |
1da177e4 | 263 | unsigned long hashval; |
3d04ebb6 CG |
264 | u_int16_t sport = 0, dport = 0; |
265 | int poff; | |
266 | ||
267 | poff = proto_ports_offset(iph->protocol); | |
268 | if (poff >= 0) { | |
269 | const u_int16_t *ports; | |
270 | u16 _ports[2]; | |
271 | ||
272 | ports = skb_header_pointer(skb, iph->ihl * 4 + poff, 4, _ports); | |
273 | if (ports) { | |
274 | sport = ports[0]; | |
275 | dport = ports[1]; | |
276 | } | |
277 | } else { | |
e87cc472 | 278 | net_info_ratelimited("unknown protocol %u\n", iph->protocol); |
1da177e4 LT |
279 | } |
280 | ||
281 | switch (config->hash_mode) { | |
282 | case CLUSTERIP_HASHMODE_SIP: | |
283 | hashval = jhash_1word(ntohl(iph->saddr), | |
284 | config->hash_initval); | |
285 | break; | |
286 | case CLUSTERIP_HASHMODE_SIP_SPT: | |
e905a9ed | 287 | hashval = jhash_2words(ntohl(iph->saddr), sport, |
1da177e4 LT |
288 | config->hash_initval); |
289 | break; | |
290 | case CLUSTERIP_HASHMODE_SIP_SPT_DPT: | |
291 | hashval = jhash_3words(ntohl(iph->saddr), sport, dport, | |
292 | config->hash_initval); | |
293 | break; | |
294 | default: | |
295 | /* to make gcc happy */ | |
296 | hashval = 0; | |
297 | /* This cannot happen, unless the check function wasn't called | |
298 | * at rule load time */ | |
ff67e4e4 | 299 | pr_info("unknown mode %u\n", config->hash_mode); |
1da177e4 LT |
300 | BUG(); |
301 | break; | |
302 | } | |
303 | ||
304 | /* node numbers are 1..n, not 0..n */ | |
8fc54f68 | 305 | return reciprocal_scale(hashval, config->num_total_nodes) + 1; |
1da177e4 LT |
306 | } |
307 | ||
308 | static inline int | |
a47362a2 | 309 | clusterip_responsible(const struct clusterip_config *config, u_int32_t hash) |
1da177e4 | 310 | { |
136e92bb | 311 | return test_bit(hash - 1, &config->local_nodes); |
1da177e4 LT |
312 | } |
313 | ||
e905a9ed YH |
314 | /*********************************************************************** |
315 | * IPTABLES TARGET | |
1da177e4 LT |
316 | ***********************************************************************/ |
317 | ||
318 | static unsigned int | |
4b560b44 | 319 | clusterip_tg(struct sk_buff *skb, const struct xt_action_param *par) |
1da177e4 | 320 | { |
7eb35586 | 321 | const struct ipt_clusterip_tgt_info *cipinfo = par->targinfo; |
587aa641 | 322 | struct nf_conn *ct; |
1da177e4 | 323 | enum ip_conntrack_info ctinfo; |
587aa641 | 324 | u_int32_t hash; |
1da177e4 LT |
325 | |
326 | /* don't need to clusterip_config_get() here, since refcount | |
327 | * is only decremented by destroy() - and ip_tables guarantees | |
328 | * that the ->target() function isn't called after ->destroy() */ | |
329 | ||
3db05fea | 330 | ct = nf_ct_get(skb, &ctinfo); |
94d117a1 | 331 | if (ct == NULL) |
1da177e4 | 332 | return NF_DROP; |
1da177e4 LT |
333 | |
334 | /* special case: ICMP error handling. conntrack distinguishes between | |
335 | * error messages (RELATED) and information requests (see below) */ | |
3666ed1c JP |
336 | if (ip_hdr(skb)->protocol == IPPROTO_ICMP && |
337 | (ctinfo == IP_CT_RELATED || | |
fb048833 | 338 | ctinfo == IP_CT_RELATED_REPLY)) |
6709dbbb | 339 | return XT_CONTINUE; |
1da177e4 | 340 | |
e905a9ed | 341 | /* ip_conntrack_icmp guarantees us that we only have ICMP_ECHO, |
1da177e4 LT |
342 | * TIMESTAMP, INFO_REQUEST or ADDRESS type icmp packets from here |
343 | * on, which all have an ID field [relevant for hashing]. */ | |
344 | ||
3db05fea | 345 | hash = clusterip_hashfn(skb, cipinfo->config); |
1da177e4 LT |
346 | |
347 | switch (ctinfo) { | |
181b1e9c JP |
348 | case IP_CT_NEW: |
349 | ct->mark = hash; | |
350 | break; | |
351 | case IP_CT_RELATED: | |
352 | case IP_CT_RELATED_REPLY: | |
353 | /* FIXME: we don't handle expectations at the moment. | |
354 | * They can arrive on a different node than | |
355 | * the master connection (e.g. FTP passive mode) */ | |
356 | case IP_CT_ESTABLISHED: | |
357 | case IP_CT_ESTABLISHED_REPLY: | |
358 | break; | |
359 | default: /* Prevent gcc warnings */ | |
360 | break; | |
1da177e4 LT |
361 | } |
362 | ||
0d53778e | 363 | #ifdef DEBUG |
3c9fba65 | 364 | nf_ct_dump_tuple_ip(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple); |
1da177e4 | 365 | #endif |
0d53778e | 366 | pr_debug("hash=%u ct_hash=%u ", hash, ct->mark); |
1da177e4 | 367 | if (!clusterip_responsible(cipinfo->config, hash)) { |
0d53778e | 368 | pr_debug("not responsible\n"); |
1da177e4 LT |
369 | return NF_DROP; |
370 | } | |
0d53778e | 371 | pr_debug("responsible\n"); |
1da177e4 LT |
372 | |
373 | /* despite being received via linklayer multicast, this is | |
374 | * actually a unicast IP packet. TCP doesn't like PACKET_MULTICAST */ | |
3db05fea | 375 | skb->pkt_type = PACKET_HOST; |
1da177e4 | 376 | |
6709dbbb | 377 | return XT_CONTINUE; |
1da177e4 LT |
378 | } |
379 | ||
135367b8 | 380 | static int clusterip_tg_check(const struct xt_tgchk_param *par) |
1da177e4 | 381 | { |
af5d6dc2 JE |
382 | struct ipt_clusterip_tgt_info *cipinfo = par->targinfo; |
383 | const struct ipt_entry *e = par->entryinfo; | |
1da177e4 | 384 | struct clusterip_config *config; |
4a5a5c73 | 385 | int ret; |
1da177e4 | 386 | |
55917a21 PNA |
387 | if (par->nft_compat) { |
388 | pr_err("cannot use CLUSTERIP target from nftables compat\n"); | |
389 | return -EOPNOTSUPP; | |
390 | } | |
391 | ||
1da177e4 LT |
392 | if (cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP && |
393 | cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP_SPT && | |
394 | cipinfo->hash_mode != CLUSTERIP_HASHMODE_SIP_SPT_DPT) { | |
ff67e4e4 | 395 | pr_info("unknown mode %u\n", cipinfo->hash_mode); |
d6b00a53 | 396 | return -EINVAL; |
1da177e4 LT |
397 | |
398 | } | |
3666ed1c JP |
399 | if (e->ip.dmsk.s_addr != htonl(0xffffffff) || |
400 | e->ip.dst.s_addr == 0) { | |
ff67e4e4 | 401 | pr_info("Please specify destination IP\n"); |
d6b00a53 | 402 | return -EINVAL; |
1da177e4 LT |
403 | } |
404 | ||
405 | /* FIXME: further sanity checks */ | |
406 | ||
d86946d2 | 407 | config = clusterip_config_find_get(par->net, e->ip.dst.s_addr, 1); |
d3c3f424 | 408 | if (!config) { |
1da177e4 | 409 | if (!(cipinfo->flags & CLUSTERIP_FLAG_NEW)) { |
ff67e4e4 JE |
410 | pr_info("no config found for %pI4, need 'new'\n", |
411 | &e->ip.dst.s_addr); | |
d6b00a53 | 412 | return -EINVAL; |
1da177e4 LT |
413 | } else { |
414 | struct net_device *dev; | |
415 | ||
416 | if (e->ip.iniface[0] == '\0') { | |
ff67e4e4 | 417 | pr_info("Please specify an interface name\n"); |
d6b00a53 | 418 | return -EINVAL; |
1da177e4 LT |
419 | } |
420 | ||
d86946d2 | 421 | dev = dev_get_by_name(par->net, e->ip.iniface); |
1da177e4 | 422 | if (!dev) { |
ff67e4e4 JE |
423 | pr_info("no such interface %s\n", |
424 | e->ip.iniface); | |
4a5a5c73 | 425 | return -ENOENT; |
1da177e4 LT |
426 | } |
427 | ||
e905a9ed | 428 | config = clusterip_config_init(cipinfo, |
1da177e4 | 429 | e->ip.dst.s_addr, dev); |
6c5d5cfb | 430 | if (IS_ERR(config)) { |
1da177e4 | 431 | dev_put(dev); |
6c5d5cfb | 432 | return PTR_ERR(config); |
1da177e4 | 433 | } |
22bedad3 | 434 | dev_mc_add(config->dev, config->clustermac); |
1da177e4 LT |
435 | } |
436 | } | |
d3c3f424 | 437 | cipinfo->config = config; |
1da177e4 | 438 | |
ecb2421b | 439 | ret = nf_ct_netns_get(par->net, par->family); |
f95c74e3 | 440 | if (ret < 0) |
ff67e4e4 JE |
441 | pr_info("cannot load conntrack support for proto=%u\n", |
442 | par->family); | |
43270b1b PNA |
443 | |
444 | if (!par->net->xt.clusterip_deprecated_warning) { | |
445 | pr_info("ipt_CLUSTERIP is deprecated and it will removed soon, " | |
446 | "use xt_cluster instead\n"); | |
447 | par->net->xt.clusterip_deprecated_warning = true; | |
448 | } | |
449 | ||
f95c74e3 | 450 | return ret; |
1da177e4 LT |
451 | } |
452 | ||
453 | /* drop reference count of cluster config when rule is deleted */ | |
a2df1648 | 454 | static void clusterip_tg_destroy(const struct xt_tgdtor_param *par) |
1da177e4 | 455 | { |
a2df1648 | 456 | const struct ipt_clusterip_tgt_info *cipinfo = par->targinfo; |
1da177e4 | 457 | |
44513624 KK |
458 | /* if no more entries are referencing the config, remove it |
459 | * from the list and destroy the proc entry */ | |
460 | clusterip_config_entry_put(cipinfo->config); | |
461 | ||
1da177e4 | 462 | clusterip_config_put(cipinfo->config); |
11078c37 | 463 | |
fe50543c | 464 | nf_ct_netns_put(par->net, par->family); |
1da177e4 LT |
465 | } |
466 | ||
d3c3f424 PM |
467 | #ifdef CONFIG_COMPAT |
468 | struct compat_ipt_clusterip_tgt_info | |
469 | { | |
470 | u_int32_t flags; | |
471 | u_int8_t clustermac[6]; | |
472 | u_int16_t num_total_nodes; | |
473 | u_int16_t num_local_nodes; | |
474 | u_int16_t local_nodes[CLUSTERIP_MAX_NODES]; | |
475 | u_int32_t hash_mode; | |
476 | u_int32_t hash_initval; | |
477 | compat_uptr_t config; | |
478 | }; | |
479 | #endif /* CONFIG_COMPAT */ | |
480 | ||
d3c5ee6d | 481 | static struct xt_target clusterip_tg_reg __read_mostly = { |
1d5cd909 | 482 | .name = "CLUSTERIP", |
ee999d8b | 483 | .family = NFPROTO_IPV4, |
d3c5ee6d JE |
484 | .target = clusterip_tg, |
485 | .checkentry = clusterip_tg_check, | |
486 | .destroy = clusterip_tg_destroy, | |
d3c3f424 | 487 | .targetsize = sizeof(struct ipt_clusterip_tgt_info), |
ec231890 | 488 | .usersize = offsetof(struct ipt_clusterip_tgt_info, config), |
d3c3f424 PM |
489 | #ifdef CONFIG_COMPAT |
490 | .compatsize = sizeof(struct compat_ipt_clusterip_tgt_info), | |
491 | #endif /* CONFIG_COMPAT */ | |
1d5cd909 | 492 | .me = THIS_MODULE |
1da177e4 LT |
493 | }; |
494 | ||
495 | ||
e905a9ed YH |
496 | /*********************************************************************** |
497 | * ARP MANGLING CODE | |
1da177e4 LT |
498 | ***********************************************************************/ |
499 | ||
500 | /* hardcoded for 48bit ethernet and 32bit ipv4 addresses */ | |
501 | struct arp_payload { | |
502 | u_int8_t src_hw[ETH_ALEN]; | |
6a19d614 | 503 | __be32 src_ip; |
1da177e4 | 504 | u_int8_t dst_hw[ETH_ALEN]; |
6a19d614 | 505 | __be32 dst_ip; |
3f30fc15 | 506 | } __packed; |
1da177e4 | 507 | |
0d53778e | 508 | #ifdef DEBUG |
e905a9ed | 509 | static void arp_print(struct arp_payload *payload) |
1da177e4 LT |
510 | { |
511 | #define HBUFFERLEN 30 | |
512 | char hbuffer[HBUFFERLEN]; | |
c8d71d08 | 513 | int j, k; |
1da177e4 | 514 | |
c8d71d08 | 515 | for (k = 0, j = 0; k < HBUFFERLEN - 3 && j < ETH_ALEN; j++) { |
6a8341b6 HH |
516 | hbuffer[k++] = hex_asc_hi(payload->src_hw[j]); |
517 | hbuffer[k++] = hex_asc_lo(payload->src_hw[j]); | |
c8d71d08 | 518 | hbuffer[k++] = ':'; |
1da177e4 | 519 | } |
c8d71d08 | 520 | hbuffer[--k] = '\0'; |
1da177e4 | 521 | |
ff67e4e4 JE |
522 | pr_debug("src %pI4@%s, dst %pI4\n", |
523 | &payload->src_ip, hbuffer, &payload->dst_ip); | |
1da177e4 LT |
524 | } |
525 | #endif | |
526 | ||
527 | static unsigned int | |
06198b34 | 528 | arp_mangle(void *priv, |
3db05fea | 529 | struct sk_buff *skb, |
238e54c9 | 530 | const struct nf_hook_state *state) |
1da177e4 | 531 | { |
3db05fea | 532 | struct arphdr *arp = arp_hdr(skb); |
1da177e4 LT |
533 | struct arp_payload *payload; |
534 | struct clusterip_config *c; | |
9dff2c96 | 535 | struct net *net = state->net; |
1da177e4 LT |
536 | |
537 | /* we don't care about non-ethernet and non-ipv4 ARP */ | |
3666ed1c JP |
538 | if (arp->ar_hrd != htons(ARPHRD_ETHER) || |
539 | arp->ar_pro != htons(ETH_P_IP) || | |
540 | arp->ar_pln != 4 || arp->ar_hln != ETH_ALEN) | |
1da177e4 LT |
541 | return NF_ACCEPT; |
542 | ||
4095ebf1 | 543 | /* we only want to mangle arp requests and replies */ |
3666ed1c JP |
544 | if (arp->ar_op != htons(ARPOP_REPLY) && |
545 | arp->ar_op != htons(ARPOP_REQUEST)) | |
1da177e4 LT |
546 | return NF_ACCEPT; |
547 | ||
548 | payload = (void *)(arp+1); | |
549 | ||
e905a9ed | 550 | /* if there is no clusterip configuration for the arp reply's |
1da177e4 | 551 | * source ip, we don't want to mangle it */ |
d86946d2 | 552 | c = clusterip_config_find_get(net, payload->src_ip, 0); |
1da177e4 LT |
553 | if (!c) |
554 | return NF_ACCEPT; | |
555 | ||
e905a9ed | 556 | /* normally the linux kernel always replies to arp queries of |
1da177e4 LT |
557 | * addresses on different interfacs. However, in the CLUSTERIP case |
558 | * this wouldn't work, since we didn't subscribe the mcast group on | |
559 | * other interfaces */ | |
238e54c9 | 560 | if (c->dev != state->out) { |
ff67e4e4 | 561 | pr_debug("not mangling arp reply on different " |
0d53778e | 562 | "interface: cip'%s'-skb'%s'\n", |
238e54c9 | 563 | c->dev->name, state->out->name); |
1da177e4 LT |
564 | clusterip_config_put(c); |
565 | return NF_ACCEPT; | |
566 | } | |
567 | ||
568 | /* mangle reply hardware address */ | |
569 | memcpy(payload->src_hw, c->clustermac, arp->ar_hln); | |
570 | ||
0d53778e | 571 | #ifdef DEBUG |
ff67e4e4 | 572 | pr_debug("mangled arp reply: "); |
1da177e4 LT |
573 | arp_print(payload); |
574 | #endif | |
575 | ||
576 | clusterip_config_put(c); | |
577 | ||
578 | return NF_ACCEPT; | |
579 | } | |
580 | ||
1999414a | 581 | static struct nf_hook_ops cip_arp_ops __read_mostly = { |
1da177e4 | 582 | .hook = arp_mangle, |
ee999d8b | 583 | .pf = NFPROTO_ARP, |
1da177e4 LT |
584 | .hooknum = NF_ARP_OUT, |
585 | .priority = -1 | |
586 | }; | |
587 | ||
e905a9ed YH |
588 | /*********************************************************************** |
589 | * PROC DIR HANDLING | |
1da177e4 LT |
590 | ***********************************************************************/ |
591 | ||
592 | #ifdef CONFIG_PROC_FS | |
593 | ||
136e92bb KK |
594 | struct clusterip_seq_position { |
595 | unsigned int pos; /* position */ | |
596 | unsigned int weight; /* number of bits set == size */ | |
597 | unsigned int bit; /* current bit */ | |
598 | unsigned long val; /* current value */ | |
599 | }; | |
600 | ||
1da177e4 LT |
601 | static void *clusterip_seq_start(struct seq_file *s, loff_t *pos) |
602 | { | |
47778147 | 603 | struct clusterip_config *c = s->private; |
136e92bb KK |
604 | unsigned int weight; |
605 | u_int32_t local_nodes; | |
606 | struct clusterip_seq_position *idx; | |
607 | ||
608 | /* FIXME: possible race */ | |
609 | local_nodes = c->local_nodes; | |
610 | weight = hweight32(local_nodes); | |
611 | if (*pos >= weight) | |
1da177e4 LT |
612 | return NULL; |
613 | ||
136e92bb KK |
614 | idx = kmalloc(sizeof(struct clusterip_seq_position), GFP_KERNEL); |
615 | if (!idx) | |
1da177e4 LT |
616 | return ERR_PTR(-ENOMEM); |
617 | ||
136e92bb KK |
618 | idx->pos = *pos; |
619 | idx->weight = weight; | |
620 | idx->bit = ffs(local_nodes); | |
621 | idx->val = local_nodes; | |
622 | clear_bit(idx->bit - 1, &idx->val); | |
623 | ||
624 | return idx; | |
1da177e4 LT |
625 | } |
626 | ||
627 | static void *clusterip_seq_next(struct seq_file *s, void *v, loff_t *pos) | |
628 | { | |
3cf93c96 | 629 | struct clusterip_seq_position *idx = v; |
1da177e4 | 630 | |
136e92bb KK |
631 | *pos = ++idx->pos; |
632 | if (*pos >= idx->weight) { | |
1da177e4 LT |
633 | kfree(v); |
634 | return NULL; | |
635 | } | |
136e92bb KK |
636 | idx->bit = ffs(idx->val); |
637 | clear_bit(idx->bit - 1, &idx->val); | |
638 | return idx; | |
1da177e4 LT |
639 | } |
640 | ||
641 | static void clusterip_seq_stop(struct seq_file *s, void *v) | |
642 | { | |
902a3dd5 ED |
643 | if (!IS_ERR(v)) |
644 | kfree(v); | |
1da177e4 LT |
645 | } |
646 | ||
647 | static int clusterip_seq_show(struct seq_file *s, void *v) | |
648 | { | |
3cf93c96 | 649 | struct clusterip_seq_position *idx = v; |
1da177e4 | 650 | |
e905a9ed | 651 | if (idx->pos != 0) |
1da177e4 | 652 | seq_putc(s, ','); |
1da177e4 | 653 | |
136e92bb KK |
654 | seq_printf(s, "%u", idx->bit); |
655 | ||
656 | if (idx->pos == idx->weight - 1) | |
1da177e4 LT |
657 | seq_putc(s, '\n'); |
658 | ||
659 | return 0; | |
660 | } | |
661 | ||
56b3d975 | 662 | static const struct seq_operations clusterip_seq_ops = { |
1da177e4 LT |
663 | .start = clusterip_seq_start, |
664 | .next = clusterip_seq_next, | |
665 | .stop = clusterip_seq_stop, | |
666 | .show = clusterip_seq_show, | |
667 | }; | |
668 | ||
669 | static int clusterip_proc_open(struct inode *inode, struct file *file) | |
670 | { | |
671 | int ret = seq_open(file, &clusterip_seq_ops); | |
672 | ||
673 | if (!ret) { | |
674 | struct seq_file *sf = file->private_data; | |
d9dda78b | 675 | struct clusterip_config *c = PDE_DATA(inode); |
1da177e4 | 676 | |
47778147 | 677 | sf->private = c; |
1da177e4 LT |
678 | |
679 | clusterip_config_get(c); | |
680 | } | |
681 | ||
682 | return ret; | |
683 | } | |
684 | ||
685 | static int clusterip_proc_release(struct inode *inode, struct file *file) | |
686 | { | |
d9dda78b | 687 | struct clusterip_config *c = PDE_DATA(inode); |
1da177e4 LT |
688 | int ret; |
689 | ||
690 | ret = seq_release(inode, file); | |
691 | ||
692 | if (!ret) | |
693 | clusterip_config_put(c); | |
694 | ||
695 | return ret; | |
696 | } | |
697 | ||
698 | static ssize_t clusterip_proc_write(struct file *file, const char __user *input, | |
699 | size_t size, loff_t *ofs) | |
700 | { | |
d9dda78b | 701 | struct clusterip_config *c = PDE_DATA(file_inode(file)); |
1da177e4 LT |
702 | #define PROC_WRITELEN 10 |
703 | char buffer[PROC_WRITELEN+1]; | |
1da177e4 | 704 | unsigned long nodenum; |
4b5511eb | 705 | int rc; |
1da177e4 | 706 | |
961ed183 VK |
707 | if (size > PROC_WRITELEN) |
708 | return -EIO; | |
709 | if (copy_from_user(buffer, input, size)) | |
1da177e4 | 710 | return -EFAULT; |
961ed183 | 711 | buffer[size] = 0; |
1da177e4 LT |
712 | |
713 | if (*buffer == '+') { | |
4b5511eb AP |
714 | rc = kstrtoul(buffer+1, 10, &nodenum); |
715 | if (rc) | |
716 | return rc; | |
1da177e4 LT |
717 | if (clusterip_add_node(c, nodenum)) |
718 | return -ENOMEM; | |
719 | } else if (*buffer == '-') { | |
4b5511eb AP |
720 | rc = kstrtoul(buffer+1, 10, &nodenum); |
721 | if (rc) | |
722 | return rc; | |
1da177e4 LT |
723 | if (clusterip_del_node(c, nodenum)) |
724 | return -ENOENT; | |
725 | } else | |
726 | return -EIO; | |
727 | ||
728 | return size; | |
729 | } | |
730 | ||
9a32144e | 731 | static const struct file_operations clusterip_proc_fops = { |
1da177e4 LT |
732 | .owner = THIS_MODULE, |
733 | .open = clusterip_proc_open, | |
734 | .read = seq_read, | |
735 | .write = clusterip_proc_write, | |
736 | .llseek = seq_lseek, | |
737 | .release = clusterip_proc_release, | |
738 | }; | |
739 | ||
740 | #endif /* CONFIG_PROC_FS */ | |
741 | ||
ce4ff76c G |
742 | static int clusterip_net_init(struct net *net) |
743 | { | |
ce4ff76c G |
744 | struct clusterip_net *cn = net_generic(net, clusterip_net_id); |
745 | ||
26a89e43 G |
746 | INIT_LIST_HEAD(&cn->configs); |
747 | ||
f1e8077f G |
748 | spin_lock_init(&cn->lock); |
749 | ||
26a89e43 | 750 | #ifdef CONFIG_PROC_FS |
ce4ff76c G |
751 | cn->procdir = proc_mkdir("ipt_CLUSTERIP", net->proc_net); |
752 | if (!cn->procdir) { | |
753 | pr_err("Unable to proc dir entry\n"); | |
754 | return -ENOMEM; | |
755 | } | |
756 | #endif /* CONFIG_PROC_FS */ | |
757 | ||
758 | return 0; | |
759 | } | |
760 | ||
761 | static void clusterip_net_exit(struct net *net) | |
762 | { | |
763 | #ifdef CONFIG_PROC_FS | |
764 | struct clusterip_net *cn = net_generic(net, clusterip_net_id); | |
765 | proc_remove(cn->procdir); | |
766 | #endif | |
767 | } | |
768 | ||
769 | static struct pernet_operations clusterip_net_ops = { | |
770 | .init = clusterip_net_init, | |
771 | .exit = clusterip_net_exit, | |
772 | .id = &clusterip_net_id, | |
773 | .size = sizeof(struct clusterip_net), | |
774 | }; | |
775 | ||
d3c5ee6d | 776 | static int __init clusterip_tg_init(void) |
1da177e4 LT |
777 | { |
778 | int ret; | |
779 | ||
ce4ff76c | 780 | ret = register_pernet_subsys(&clusterip_net_ops); |
32292a7f PM |
781 | if (ret < 0) |
782 | return ret; | |
1da177e4 | 783 | |
ce4ff76c G |
784 | ret = xt_register_target(&clusterip_tg_reg); |
785 | if (ret < 0) | |
786 | goto cleanup_subsys; | |
787 | ||
32292a7f PM |
788 | ret = nf_register_hook(&cip_arp_ops); |
789 | if (ret < 0) | |
1da177e4 | 790 | goto cleanup_target; |
1da177e4 | 791 | |
ff67e4e4 | 792 | pr_info("ClusterIP Version %s loaded successfully\n", |
1da177e4 | 793 | CLUSTERIP_VERSION); |
ce4ff76c | 794 | |
1da177e4 LT |
795 | return 0; |
796 | ||
1da177e4 | 797 | cleanup_target: |
d3c5ee6d | 798 | xt_unregister_target(&clusterip_tg_reg); |
ce4ff76c G |
799 | cleanup_subsys: |
800 | unregister_pernet_subsys(&clusterip_net_ops); | |
32292a7f | 801 | return ret; |
1da177e4 LT |
802 | } |
803 | ||
d3c5ee6d | 804 | static void __exit clusterip_tg_exit(void) |
1da177e4 | 805 | { |
ff67e4e4 | 806 | pr_info("ClusterIP Version %s unloading\n", CLUSTERIP_VERSION); |
ce4ff76c | 807 | |
32292a7f | 808 | nf_unregister_hook(&cip_arp_ops); |
d3c5ee6d | 809 | xt_unregister_target(&clusterip_tg_reg); |
ce4ff76c | 810 | unregister_pernet_subsys(&clusterip_net_ops); |
d73f33b1 ED |
811 | |
812 | /* Wait for completion of call_rcu_bh()'s (clusterip_config_rcu_free) */ | |
813 | rcu_barrier_bh(); | |
1da177e4 LT |
814 | } |
815 | ||
d3c5ee6d JE |
816 | module_init(clusterip_tg_init); |
817 | module_exit(clusterip_tg_exit); |