2 * ip_vs_proto.c: transport protocol load balancing support for IPVS
4 * Authors: Wensong Zhang <wensong@linuxvirtualserver.org>
5 * Julian Anastasov <ja@ssi.bg>
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version
10 * 2 of the License, or (at your option) any later version.
16 #define KMSG_COMPONENT "IPVS"
17 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
19 #include <linux/module.h>
20 #include <linux/kernel.h>
21 #include <linux/skbuff.h>
22 #include <linux/gfp.h>
25 #include <net/protocol.h>
28 #include <asm/system.h>
29 #include <linux/stat.h>
30 #include <linux/proc_fs.h>
32 #include <net/ip_vs.h>
36 * IPVS protocols can only be registered/unregistered when the ipvs
37 * module is loaded/unloaded, so no lock is needed in accessing the
38 * ipvs protocol table.
41 #define IP_VS_PROTO_TAB_SIZE 32 /* must be power of 2 */
42 #define IP_VS_PROTO_HASH(proto) ((proto) & (IP_VS_PROTO_TAB_SIZE-1))
44 static struct ip_vs_protocol *ip_vs_proto_table[IP_VS_PROTO_TAB_SIZE];
48 * register an ipvs protocol
50 static int __used __init register_ip_vs_protocol(struct ip_vs_protocol *pp)
52 unsigned hash = IP_VS_PROTO_HASH(pp->protocol);
54 pp->next = ip_vs_proto_table[hash];
55 ip_vs_proto_table[hash] = pp;
63 #if defined(CONFIG_IP_VS_PROTO_TCP) || defined(CONFIG_IP_VS_PROTO_UDP) || \
64 defined(CONFIG_IP_VS_PROTO_SCTP) || defined(CONFIG_IP_VS_PROTO_AH) || \
65 defined(CONFIG_IP_VS_PROTO_ESP)
67 * register an ipvs protocols netns related data
70 register_ip_vs_proto_netns(struct net *net, struct ip_vs_protocol *pp)
72 struct netns_ipvs *ipvs = net_ipvs(net);
73 unsigned hash = IP_VS_PROTO_HASH(pp->protocol);
74 struct ip_vs_proto_data *pd =
75 kzalloc(sizeof(struct ip_vs_proto_data), GFP_ATOMIC);
78 pr_err("%s(): no memory.\n", __func__);
81 pd->pp = pp; /* For speed issues */
82 pd->next = ipvs->proto_data_table[hash];
83 ipvs->proto_data_table[hash] = pd;
84 atomic_set(&pd->appcnt, 0); /* Init app counter */
86 if (pp->init_netns != NULL)
87 pp->init_netns(net, pd);
94 * unregister an ipvs protocol
96 static int unregister_ip_vs_protocol(struct ip_vs_protocol *pp)
98 struct ip_vs_protocol **pp_p;
99 unsigned hash = IP_VS_PROTO_HASH(pp->protocol);
101 pp_p = &ip_vs_proto_table[hash];
102 for (; *pp_p; pp_p = &(*pp_p)->next) {
105 if (pp->exit != NULL)
115 * unregister an ipvs protocols netns data
118 unregister_ip_vs_proto_netns(struct net *net, struct ip_vs_proto_data *pd)
120 struct netns_ipvs *ipvs = net_ipvs(net);
121 struct ip_vs_proto_data **pd_p;
122 unsigned hash = IP_VS_PROTO_HASH(pd->pp->protocol);
124 pd_p = &ipvs->proto_data_table[hash];
125 for (; *pd_p; pd_p = &(*pd_p)->next) {
128 if (pd->pp->exit_netns != NULL)
129 pd->pp->exit_netns(net, pd);
139 * get ip_vs_protocol object by its proto.
141 struct ip_vs_protocol * ip_vs_proto_get(unsigned short proto)
143 struct ip_vs_protocol *pp;
144 unsigned hash = IP_VS_PROTO_HASH(proto);
146 for (pp = ip_vs_proto_table[hash]; pp; pp = pp->next) {
147 if (pp->protocol == proto)
153 EXPORT_SYMBOL(ip_vs_proto_get);
156 * get ip_vs_protocol object data by netns and proto
158 struct ip_vs_proto_data *
159 __ipvs_proto_data_get(struct netns_ipvs *ipvs, unsigned short proto)
161 struct ip_vs_proto_data *pd;
162 unsigned hash = IP_VS_PROTO_HASH(proto);
164 for (pd = ipvs->proto_data_table[hash]; pd; pd = pd->next) {
165 if (pd->pp->protocol == proto)
172 struct ip_vs_proto_data *
173 ip_vs_proto_data_get(struct net *net, unsigned short proto)
175 struct netns_ipvs *ipvs = net_ipvs(net);
177 return __ipvs_proto_data_get(ipvs, proto);
179 EXPORT_SYMBOL(ip_vs_proto_data_get);
182 * Propagate event for state change to all protocols
184 void ip_vs_protocol_timeout_change(struct netns_ipvs *ipvs, int flags)
186 struct ip_vs_proto_data *pd;
189 for (i = 0; i < IP_VS_PROTO_TAB_SIZE; i++) {
190 for (pd = ipvs->proto_data_table[i]; pd; pd = pd->next) {
191 if (pd->pp->timeout_change)
192 pd->pp->timeout_change(pd, flags);
199 ip_vs_create_timeout_table(int *table, int size)
201 return kmemdup(table, size, GFP_ATOMIC);
206 * Set timeout value for state specified by name
209 ip_vs_set_state_timeout(int *table, int num, const char *const *names,
210 const char *name, int to)
214 if (!table || !name || !to)
217 for (i = 0; i < num; i++) {
218 if (strcmp(names[i], name))
227 const char * ip_vs_state_name(__u16 proto, int state)
229 struct ip_vs_protocol *pp = ip_vs_proto_get(proto);
231 if (pp == NULL || pp->state_name == NULL)
232 return (IPPROTO_IP == proto) ? "NONE" : "ERR!";
233 return pp->state_name(state);
238 ip_vs_tcpudp_debug_packet_v4(struct ip_vs_protocol *pp,
239 const struct sk_buff *skb,
244 struct iphdr _iph, *ih;
246 ih = skb_header_pointer(skb, offset, sizeof(_iph), &_iph);
248 sprintf(buf, "TRUNCATED");
249 else if (ih->frag_off & htons(IP_OFFSET))
250 sprintf(buf, "%pI4->%pI4 frag", &ih->saddr, &ih->daddr);
252 __be16 _ports[2], *pptr;
254 pptr = skb_header_pointer(skb, offset + ih->ihl*4,
255 sizeof(_ports), _ports);
257 sprintf(buf, "TRUNCATED %pI4->%pI4",
258 &ih->saddr, &ih->daddr);
260 sprintf(buf, "%pI4:%u->%pI4:%u",
261 &ih->saddr, ntohs(pptr[0]),
262 &ih->daddr, ntohs(pptr[1]));
265 pr_debug("%s: %s %s\n", msg, pp->name, buf);
268 #ifdef CONFIG_IP_VS_IPV6
270 ip_vs_tcpudp_debug_packet_v6(struct ip_vs_protocol *pp,
271 const struct sk_buff *skb,
276 struct ipv6hdr _iph, *ih;
278 ih = skb_header_pointer(skb, offset, sizeof(_iph), &_iph);
280 sprintf(buf, "TRUNCATED");
281 else if (ih->nexthdr == IPPROTO_FRAGMENT)
282 sprintf(buf, "%pI6->%pI6 frag", &ih->saddr, &ih->daddr);
284 __be16 _ports[2], *pptr;
286 pptr = skb_header_pointer(skb, offset + sizeof(struct ipv6hdr),
287 sizeof(_ports), _ports);
289 sprintf(buf, "TRUNCATED %pI6->%pI6",
290 &ih->saddr, &ih->daddr);
292 sprintf(buf, "%pI6:%u->%pI6:%u",
293 &ih->saddr, ntohs(pptr[0]),
294 &ih->daddr, ntohs(pptr[1]));
297 pr_debug("%s: %s %s\n", msg, pp->name, buf);
303 ip_vs_tcpudp_debug_packet(int af, struct ip_vs_protocol *pp,
304 const struct sk_buff *skb,
308 #ifdef CONFIG_IP_VS_IPV6
310 ip_vs_tcpudp_debug_packet_v6(pp, skb, offset, msg);
313 ip_vs_tcpudp_debug_packet_v4(pp, skb, offset, msg);
317 * per network name-space init
319 static int __net_init __ip_vs_protocol_init(struct net *net)
321 #ifdef CONFIG_IP_VS_PROTO_TCP
322 register_ip_vs_proto_netns(net, &ip_vs_protocol_tcp);
324 #ifdef CONFIG_IP_VS_PROTO_UDP
325 register_ip_vs_proto_netns(net, &ip_vs_protocol_udp);
327 #ifdef CONFIG_IP_VS_PROTO_SCTP
328 register_ip_vs_proto_netns(net, &ip_vs_protocol_sctp);
330 #ifdef CONFIG_IP_VS_PROTO_AH
331 register_ip_vs_proto_netns(net, &ip_vs_protocol_ah);
333 #ifdef CONFIG_IP_VS_PROTO_ESP
334 register_ip_vs_proto_netns(net, &ip_vs_protocol_esp);
339 static void __net_exit __ip_vs_protocol_cleanup(struct net *net)
341 struct netns_ipvs *ipvs = net_ipvs(net);
342 struct ip_vs_proto_data *pd;
345 /* unregister all the ipvs proto data for this netns */
346 for (i = 0; i < IP_VS_PROTO_TAB_SIZE; i++) {
347 while ((pd = ipvs->proto_data_table[i]) != NULL)
348 unregister_ip_vs_proto_netns(net, pd);
352 static struct pernet_operations ipvs_proto_ops = {
353 .init = __ip_vs_protocol_init,
354 .exit = __ip_vs_protocol_cleanup,
357 int __init ip_vs_protocol_init(void)
360 #define REGISTER_PROTOCOL(p) \
362 register_ip_vs_protocol(p); \
363 strcat(protocols, ", "); \
364 strcat(protocols, (p)->name); \
369 #ifdef CONFIG_IP_VS_PROTO_TCP
370 REGISTER_PROTOCOL(&ip_vs_protocol_tcp);
372 #ifdef CONFIG_IP_VS_PROTO_UDP
373 REGISTER_PROTOCOL(&ip_vs_protocol_udp);
375 #ifdef CONFIG_IP_VS_PROTO_SCTP
376 REGISTER_PROTOCOL(&ip_vs_protocol_sctp);
378 #ifdef CONFIG_IP_VS_PROTO_AH
379 REGISTER_PROTOCOL(&ip_vs_protocol_ah);
381 #ifdef CONFIG_IP_VS_PROTO_ESP
382 REGISTER_PROTOCOL(&ip_vs_protocol_esp);
384 pr_info("Registered protocols (%s)\n", &protocols[2]);
385 return register_pernet_subsys(&ipvs_proto_ops);
391 void ip_vs_protocol_cleanup(void)
393 struct ip_vs_protocol *pp;
396 unregister_pernet_subsys(&ipvs_proto_ops);
397 /* unregister all the ipvs protocols */
398 for (i = 0; i < IP_VS_PROTO_TAB_SIZE; i++) {
399 while ((pp = ip_vs_proto_table[i]) != NULL)
400 unregister_ip_vs_protocol(pp);