net/netfilter/ipvs/ip_vs_proto.c

   1 /*
   2  * ip_vs_proto.c: transport protocol load balancing support for IPVS
   3  *
   4  * Authors:     Wensong Zhang <wensong@linuxvirtualserver.org>
   5  *              Julian Anastasov <ja@ssi.bg>
   6  *
   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.
  11  *
  12  * Changes:
  13  *
  14  */
  15
  16 #define KMSG_COMPONENT "IPVS"
  17 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  18
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/skbuff.h>
  22 #include <linux/gfp.h>
  23 #include <linux/in.h>
  24 #include <linux/ip.h>
  25 #include <net/protocol.h>
  26 #include <net/tcp.h>
  27 #include <net/udp.h>
  28 #include <linux/stat.h>
  29 #include <linux/proc_fs.h>
  30
  31 #include <net/ip_vs.h>
  32
  33
  34 /*
  35  * IPVS protocols can only be registered/unregistered when the ipvs
  36  * module is loaded/unloaded, so no lock is needed in accessing the
  37  * ipvs protocol table.
  38  */
  39
  40 #define IP_VS_PROTO_TAB_SIZE            32      /* must be power of 2 */
  41 #define IP_VS_PROTO_HASH(proto)         ((proto) & (IP_VS_PROTO_TAB_SIZE-1))
  42
  43 static struct ip_vs_protocol *ip_vs_proto_table[IP_VS_PROTO_TAB_SIZE];
  44
  45
  46 /*
  47  *      register an ipvs protocol
  48  */
  49 static int __used __init register_ip_vs_protocol(struct ip_vs_protocol *pp)
  50 {
  51         unsigned hash = IP_VS_PROTO_HASH(pp->protocol);
  52
  53         pp->next = ip_vs_proto_table[hash];
  54         ip_vs_proto_table[hash] = pp;
  55
  56         if (pp->init != NULL)
  57                 pp->init(pp);
  58
  59         return 0;
  60 }
  61
  62 #if defined(CONFIG_IP_VS_PROTO_TCP) || defined(CONFIG_IP_VS_PROTO_UDP) || \
  63     defined(CONFIG_IP_VS_PROTO_SCTP) || defined(CONFIG_IP_VS_PROTO_AH) || \
  64     defined(CONFIG_IP_VS_PROTO_ESP)
  65 /*
  66  *      register an ipvs protocols netns related data
  67  */
  68 static int
  69 register_ip_vs_proto_netns(struct net *net, struct ip_vs_protocol *pp)
  70 {
  71         struct netns_ipvs *ipvs = net_ipvs(net);
  72         unsigned hash = IP_VS_PROTO_HASH(pp->protocol);
  73         struct ip_vs_proto_data *pd =
  74                         kzalloc(sizeof(struct ip_vs_proto_data), GFP_ATOMIC);
  75
  76         if (!pd)
  77                 return -ENOMEM;
  78
  79         pd->pp = pp;    /* For speed issues */
  80         pd->next = ipvs->proto_data_table[hash];
  81         ipvs->proto_data_table[hash] = pd;
  82         atomic_set(&pd->appcnt, 0);     /* Init app counter */
  83
  84         if (pp->init_netns != NULL)
  85                 pp->init_netns(net, pd);
  86
  87         return 0;
  88 }
  89 #endif
  90
  91 /*
  92  *      unregister an ipvs protocol
  93  */
  94 static int unregister_ip_vs_protocol(struct ip_vs_protocol *pp)
  95 {
  96         struct ip_vs_protocol **pp_p;
  97         unsigned hash = IP_VS_PROTO_HASH(pp->protocol);
  98
  99         pp_p = &ip_vs_proto_table[hash];
 100         for (; *pp_p; pp_p = &(*pp_p)->next) {
 101                 if (*pp_p == pp) {
 102                         *pp_p = pp->next;
 103                         if (pp->exit != NULL)
 104                                 pp->exit(pp);
 105                         return 0;
 106                 }
 107         }
 108
 109         return -ESRCH;
 110 }
 111
 112 /*
 113  *      unregister an ipvs protocols netns data
 114  */
 115 static int
 116 unregister_ip_vs_proto_netns(struct net *net, struct ip_vs_proto_data *pd)
 117 {
 118         struct netns_ipvs *ipvs = net_ipvs(net);
 119         struct ip_vs_proto_data **pd_p;
 120         unsigned hash = IP_VS_PROTO_HASH(pd->pp->protocol);
 121
 122         pd_p = &ipvs->proto_data_table[hash];
 123         for (; *pd_p; pd_p = &(*pd_p)->next) {
 124                 if (*pd_p == pd) {
 125                         *pd_p = pd->next;
 126                         if (pd->pp->exit_netns != NULL)
 127                                 pd->pp->exit_netns(net, pd);
 128                         kfree(pd);
 129                         return 0;
 130                 }
 131         }
 132
 133         return -ESRCH;
 134 }
 135
 136 /*
 137  *      get ip_vs_protocol object by its proto.
 138  */
 139 struct ip_vs_protocol * ip_vs_proto_get(unsigned short proto)
 140 {
 141         struct ip_vs_protocol *pp;
 142         unsigned hash = IP_VS_PROTO_HASH(proto);
 143
 144         for (pp = ip_vs_proto_table[hash]; pp; pp = pp->next) {
 145                 if (pp->protocol == proto)
 146                         return pp;
 147         }
 148
 149         return NULL;
 150 }
 151 EXPORT_SYMBOL(ip_vs_proto_get);
 152
 153 /*
 154  *      get ip_vs_protocol object data by netns and proto
 155  */
 156 struct ip_vs_proto_data *
 157 __ipvs_proto_data_get(struct netns_ipvs *ipvs, unsigned short proto)
 158 {
 159         struct ip_vs_proto_data *pd;
 160         unsigned hash = IP_VS_PROTO_HASH(proto);
 161
 162         for (pd = ipvs->proto_data_table[hash]; pd; pd = pd->next) {
 163                 if (pd->pp->protocol == proto)
 164                         return pd;
 165         }
 166
 167         return NULL;
 168 }
 169
 170 struct ip_vs_proto_data *
 171 ip_vs_proto_data_get(struct net *net, unsigned short proto)
 172 {
 173         struct netns_ipvs *ipvs = net_ipvs(net);
 174
 175         return __ipvs_proto_data_get(ipvs, proto);
 176 }
 177 EXPORT_SYMBOL(ip_vs_proto_data_get);
 178
 179 /*
 180  *      Propagate event for state change to all protocols
 181  */
 182 void ip_vs_protocol_timeout_change(struct netns_ipvs *ipvs, int flags)
 183 {
 184         struct ip_vs_proto_data *pd;
 185         int i;
 186
 187         for (i = 0; i < IP_VS_PROTO_TAB_SIZE; i++) {
 188                 for (pd = ipvs->proto_data_table[i]; pd; pd = pd->next) {
 189                         if (pd->pp->timeout_change)
 190                                 pd->pp->timeout_change(pd, flags);
 191                 }
 192         }
 193 }
 194
 195
 196 int *
 197 ip_vs_create_timeout_table(int *table, int size)
 198 {
 199         return kmemdup(table, size, GFP_ATOMIC);
 200 }
 201
 202
 203 /*
 204  *      Set timeout value for state specified by name
 205  */
 206 int
 207 ip_vs_set_state_timeout(int *table, int num, const char *const *names,
 208                         const char *name, int to)
 209 {
 210         int i;
 211
 212         if (!table || !name || !to)
 213                 return -EINVAL;
 214
 215         for (i = 0; i < num; i++) {
 216                 if (strcmp(names[i], name))
 217                         continue;
 218                 table[i] = to * HZ;
 219                 return 0;
 220         }
 221         return -ENOENT;
 222 }
 223
 224
 225 const char * ip_vs_state_name(__u16 proto, int state)
 226 {
 227         struct ip_vs_protocol *pp = ip_vs_proto_get(proto);
 228
 229         if (pp == NULL || pp->state_name == NULL)
 230                 return (IPPROTO_IP == proto) ? "NONE" : "ERR!";
 231         return pp->state_name(state);
 232 }
 233
 234
 235 static void
 236 ip_vs_tcpudp_debug_packet_v4(struct ip_vs_protocol *pp,
 237                              const struct sk_buff *skb,
 238                              int offset,
 239                              const char *msg)
 240 {
 241         char buf[128];
 242         struct iphdr _iph, *ih;
 243
 244         ih = skb_header_pointer(skb, offset, sizeof(_iph), &_iph);
 245         if (ih == NULL)
 246                 sprintf(buf, "TRUNCATED");
 247         else if (ih->frag_off & htons(IP_OFFSET))
 248                 sprintf(buf, "%pI4->%pI4 frag", &ih->saddr, &ih->daddr);
 249         else {
 250                 __be16 _ports[2], *pptr;
 251
 252                 pptr = skb_header_pointer(skb, offset + ih->ihl*4,
 253                                           sizeof(_ports), _ports);
 254                 if (pptr == NULL)
 255                         sprintf(buf, "TRUNCATED %pI4->%pI4",
 256                                 &ih->saddr, &ih->daddr);
 257                 else
 258                         sprintf(buf, "%pI4:%u->%pI4:%u",
 259                                 &ih->saddr, ntohs(pptr[0]),
 260                                 &ih->daddr, ntohs(pptr[1]));
 261         }
 262
 263         pr_debug("%s: %s %s\n", msg, pp->name, buf);
 264 }
 265
 266 #ifdef CONFIG_IP_VS_IPV6
 267 static void
 268 ip_vs_tcpudp_debug_packet_v6(struct ip_vs_protocol *pp,
 269                              const struct sk_buff *skb,
 270                              int offset,
 271                              const char *msg)
 272 {
 273         char buf[192];
 274         struct ipv6hdr _iph, *ih;
 275
 276         ih = skb_header_pointer(skb, offset, sizeof(_iph), &_iph);
 277         if (ih == NULL)
 278                 sprintf(buf, "TRUNCATED");
 279         else if (ih->nexthdr == IPPROTO_FRAGMENT)
 280                 sprintf(buf, "%pI6->%pI6 frag", &ih->saddr, &ih->daddr);
 281         else {
 282                 __be16 _ports[2], *pptr;
 283
 284                 pptr = skb_header_pointer(skb, offset + sizeof(struct ipv6hdr),
 285                                           sizeof(_ports), _ports);
 286                 if (pptr == NULL)
 287                         sprintf(buf, "TRUNCATED %pI6->%pI6",
 288                                 &ih->saddr, &ih->daddr);
 289                 else
 290                         sprintf(buf, "%pI6:%u->%pI6:%u",
 291                                 &ih->saddr, ntohs(pptr[0]),
 292                                 &ih->daddr, ntohs(pptr[1]));
 293         }
 294
 295         pr_debug("%s: %s %s\n", msg, pp->name, buf);
 296 }
 297 #endif
 298
 299
 300 void
 301 ip_vs_tcpudp_debug_packet(int af, struct ip_vs_protocol *pp,
 302                           const struct sk_buff *skb,
 303                           int offset,
 304                           const char *msg)
 305 {
 306 #ifdef CONFIG_IP_VS_IPV6
 307         if (af == AF_INET6)
 308                 ip_vs_tcpudp_debug_packet_v6(pp, skb, offset, msg);
 309         else
 310 #endif
 311                 ip_vs_tcpudp_debug_packet_v4(pp, skb, offset, msg);
 312 }
 313
 314 /*
 315  * per network name-space init
 316  */
 317 int __net_init ip_vs_protocol_net_init(struct net *net)
 318 {
 319 #ifdef CONFIG_IP_VS_PROTO_TCP
 320         register_ip_vs_proto_netns(net, &ip_vs_protocol_tcp);
 321 #endif
 322 #ifdef CONFIG_IP_VS_PROTO_UDP
 323         register_ip_vs_proto_netns(net, &ip_vs_protocol_udp);
 324 #endif
 325 #ifdef CONFIG_IP_VS_PROTO_SCTP
 326         register_ip_vs_proto_netns(net, &ip_vs_protocol_sctp);
 327 #endif
 328 #ifdef CONFIG_IP_VS_PROTO_AH
 329         register_ip_vs_proto_netns(net, &ip_vs_protocol_ah);
 330 #endif
 331 #ifdef CONFIG_IP_VS_PROTO_ESP
 332         register_ip_vs_proto_netns(net, &ip_vs_protocol_esp);
 333 #endif
 334         return 0;
 335 }
 336
 337 void __net_exit ip_vs_protocol_net_cleanup(struct net *net)
 338 {
 339         struct netns_ipvs *ipvs = net_ipvs(net);
 340         struct ip_vs_proto_data *pd;
 341         int i;
 342
 343         /* unregister all the ipvs proto data for this netns */
 344         for (i = 0; i < IP_VS_PROTO_TAB_SIZE; i++) {
 345                 while ((pd = ipvs->proto_data_table[i]) != NULL)
 346                         unregister_ip_vs_proto_netns(net, pd);
 347         }
 348 }
 349
 350 int __init ip_vs_protocol_init(void)
 351 {
 352         char protocols[64];
 353 #define REGISTER_PROTOCOL(p)                    \
 354         do {                                    \
 355                 register_ip_vs_protocol(p);     \
 356                 strcat(protocols, ", ");        \
 357                 strcat(protocols, (p)->name);   \
 358         } while (0)
 359
 360         protocols[0] = '\0';
 361         protocols[2] = '\0';
 362 #ifdef CONFIG_IP_VS_PROTO_TCP
 363         REGISTER_PROTOCOL(&ip_vs_protocol_tcp);
 364 #endif
 365 #ifdef CONFIG_IP_VS_PROTO_UDP
 366         REGISTER_PROTOCOL(&ip_vs_protocol_udp);
 367 #endif
 368 #ifdef CONFIG_IP_VS_PROTO_SCTP
 369         REGISTER_PROTOCOL(&ip_vs_protocol_sctp);
 370 #endif
 371 #ifdef CONFIG_IP_VS_PROTO_AH
 372         REGISTER_PROTOCOL(&ip_vs_protocol_ah);
 373 #endif
 374 #ifdef CONFIG_IP_VS_PROTO_ESP
 375         REGISTER_PROTOCOL(&ip_vs_protocol_esp);
 376 #endif
 377         pr_info("Registered protocols (%s)\n", &protocols[2]);
 378
 379         return 0;
 380 }
 381
 382
 383 void ip_vs_protocol_cleanup(void)
 384 {
 385         struct ip_vs_protocol *pp;
 386         int i;
 387
 388         /* unregister all the ipvs protocols */
 389         for (i = 0; i < IP_VS_PROTO_TAB_SIZE; i++) {
 390                 while ((pp = ip_vs_proto_table[i]) != NULL)
 391                         unregister_ip_vs_protocol(pp);
 392         }
 393 }