net/netfilter/ipvs/ip_vs_sh.c

   1 /*
   2  * IPVS:        Source Hashing scheduling module
   3  *
   4  * Authors:     Wensong Zhang <wensong@gnuchina.org>
   5  *
   6  *              This program is free software; you can redistribute it and/or
   7  *              modify it under the terms of the GNU General Public License
   8  *              as published by the Free Software Foundation; either version
   9  *              2 of the License, or (at your option) any later version.
  10  *
  11  * Changes:
  12  *
  13  */
  14
  15 /*
  16  * The sh algorithm is to select server by the hash key of source IP
  17  * address. The pseudo code is as follows:
  18  *
  19  *       n <- servernode[src_ip];
  20  *       if (n is dead) OR
  21  *          (n is overloaded) or (n.weight <= 0) then
  22  *                 return NULL;
  23  *
  24  *       return n;
  25  *
  26  * Notes that servernode is a 256-bucket hash table that maps the hash
  27  * index derived from packet source IP address to the current server
  28  * array. If the sh scheduler is used in cache cluster, it is good to
  29  * combine it with cache_bypass feature. When the statically assigned
  30  * server is dead or overloaded, the load balancer can bypass the cache
  31  * server and send requests to the original server directly.
  32  *
  33  * The weight destination attribute can be used to control the
  34  * distribution of connections to the destinations in servernode. The
  35  * greater the weight, the more connections the destination
  36  * will receive.
  37  *
  38  */
  39
  40 #define KMSG_COMPONENT "IPVS"
  41 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  42
  43 #include <linux/ip.h>
  44 #include <linux/slab.h>
  45 #include <linux/module.h>
  46 #include <linux/kernel.h>
  47 #include <linux/skbuff.h>
  48
  49 #include <net/ip_vs.h>
  50
  51 #include <net/tcp.h>
  52 #include <linux/udp.h>
  53 #include <linux/sctp.h>
  54
  55
  56 /*
  57  *      IPVS SH bucket
  58  */
  59 struct ip_vs_sh_bucket {
  60         struct ip_vs_dest __rcu *dest;  /* real server (cache) */
  61 };
  62
  63 /*
  64  *     for IPVS SH entry hash table
  65  */
  66 #ifndef CONFIG_IP_VS_SH_TAB_BITS
  67 #define CONFIG_IP_VS_SH_TAB_BITS        8
  68 #endif
  69 #define IP_VS_SH_TAB_BITS               CONFIG_IP_VS_SH_TAB_BITS
  70 #define IP_VS_SH_TAB_SIZE               (1 << IP_VS_SH_TAB_BITS)
  71 #define IP_VS_SH_TAB_MASK               (IP_VS_SH_TAB_SIZE - 1)
  72
  73 struct ip_vs_sh_state {
  74         struct rcu_head                 rcu_head;
  75         struct ip_vs_sh_bucket          buckets[IP_VS_SH_TAB_SIZE];
  76 };
  77
  78 /* Helper function to determine if server is unavailable */
  79 static inline bool is_unavailable(struct ip_vs_dest *dest)
  80 {
  81         return atomic_read(&dest->weight) <= 0 ||
  82                dest->flags & IP_VS_DEST_F_OVERLOAD;
  83 }
  84
  85 /*
  86  *      Returns hash value for IPVS SH entry
  87  */
  88 static inline unsigned int
  89 ip_vs_sh_hashkey(int af, const union nf_inet_addr *addr,
  90                  __be16 port, unsigned int offset)
  91 {
  92         __be32 addr_fold = addr->ip;
  93
  94 #ifdef CONFIG_IP_VS_IPV6
  95         if (af == AF_INET6)
  96                 addr_fold = addr->ip6[0]^addr->ip6[1]^
  97                             addr->ip6[2]^addr->ip6[3];
  98 #endif
  99         return (offset + (ntohs(port) + ntohl(addr_fold))*2654435761UL) &
 100                 IP_VS_SH_TAB_MASK;
 101 }
 102
 103
 104 /*
 105  *      Get ip_vs_dest associated with supplied parameters.
 106  */
 107 static inline struct ip_vs_dest *
 108 ip_vs_sh_get(struct ip_vs_service *svc, struct ip_vs_sh_state *s,
 109              const union nf_inet_addr *addr, __be16 port)
 110 {
 111         unsigned int hash = ip_vs_sh_hashkey(svc->af, addr, port, 0);
 112         struct ip_vs_dest *dest = rcu_dereference(s->buckets[hash].dest);
 113
 114         return (!dest || is_unavailable(dest)) ? NULL : dest;
 115 }
 116
 117
 118 /* As ip_vs_sh_get, but with fallback if selected server is unavailable
 119  *
 120  * The fallback strategy loops around the table starting from a "random"
 121  * point (in fact, it is chosen to be the original hash value to make the
 122  * algorithm deterministic) to find a new server.
 123  */
 124 static inline struct ip_vs_dest *
 125 ip_vs_sh_get_fallback(struct ip_vs_service *svc, struct ip_vs_sh_state *s,
 126                       const union nf_inet_addr *addr, __be16 port)
 127 {
 128         unsigned int offset, roffset;
 129         unsigned int hash, ihash;
 130         struct ip_vs_dest *dest;
 131
 132         /* first try the dest it's supposed to go to */
 133         ihash = ip_vs_sh_hashkey(svc->af, addr, port, 0);
 134         dest = rcu_dereference(s->buckets[ihash].dest);
 135         if (!dest)
 136                 return NULL;
 137         if (!is_unavailable(dest))
 138                 return dest;
 139
 140         IP_VS_DBG_BUF(6, "SH: selected unavailable server %s:%d, reselecting",
 141                       IP_VS_DBG_ADDR(dest->af, &dest->addr), ntohs(dest->port));
 142
 143         /* if the original dest is unavailable, loop around the table
 144          * starting from ihash to find a new dest
 145          */
 146         for (offset = 0; offset < IP_VS_SH_TAB_SIZE; offset++) {
 147                 roffset = (offset + ihash) % IP_VS_SH_TAB_SIZE;
 148                 hash = ip_vs_sh_hashkey(svc->af, addr, port, roffset);
 149                 dest = rcu_dereference(s->buckets[hash].dest);
 150                 if (!dest)
 151                         break;
 152                 if (!is_unavailable(dest))
 153                         return dest;
 154                 IP_VS_DBG_BUF(6, "SH: selected unavailable "
 155                               "server %s:%d (offset %d), reselecting",
 156                               IP_VS_DBG_ADDR(dest->af, &dest->addr),
 157                               ntohs(dest->port), roffset);
 158         }
 159
 160         return NULL;
 161 }
 162
 163 /*
 164  *      Assign all the hash buckets of the specified table with the service.
 165  */
 166 static int
 167 ip_vs_sh_reassign(struct ip_vs_sh_state *s, struct ip_vs_service *svc)
 168 {
 169         int i;
 170         struct ip_vs_sh_bucket *b;
 171         struct list_head *p;
 172         struct ip_vs_dest *dest;
 173         int d_count;
 174         bool empty;
 175
 176         b = &s->buckets[0];
 177         p = &svc->destinations;
 178         empty = list_empty(p);
 179         d_count = 0;
 180         for (i=0; i<IP_VS_SH_TAB_SIZE; i++) {
 181                 dest = rcu_dereference_protected(b->dest, 1);
 182                 if (dest)
 183                         ip_vs_dest_put(dest);
 184                 if (empty)
 185                         RCU_INIT_POINTER(b->dest, NULL);
 186                 else {
 187                         if (p == &svc->destinations)
 188                                 p = p->next;
 189
 190                         dest = list_entry(p, struct ip_vs_dest, n_list);
 191                         ip_vs_dest_hold(dest);
 192                         RCU_INIT_POINTER(b->dest, dest);
 193
 194                         IP_VS_DBG_BUF(6, "assigned i: %d dest: %s weight: %d\n",
 195                                       i, IP_VS_DBG_ADDR(dest->af, &dest->addr),
 196                                       atomic_read(&dest->weight));
 197
 198                         /* Don't move to next dest until filling weight */
 199                         if (++d_count >= atomic_read(&dest->weight)) {
 200                                 p = p->next;
 201                                 d_count = 0;
 202                         }
 203
 204                 }
 205                 b++;
 206         }
 207         return 0;
 208 }
 209
 210
 211 /*
 212  *      Flush all the hash buckets of the specified table.
 213  */
 214 static void ip_vs_sh_flush(struct ip_vs_sh_state *s)
 215 {
 216         int i;
 217         struct ip_vs_sh_bucket *b;
 218         struct ip_vs_dest *dest;
 219
 220         b = &s->buckets[0];
 221         for (i=0; i<IP_VS_SH_TAB_SIZE; i++) {
 222                 dest = rcu_dereference_protected(b->dest, 1);
 223                 if (dest) {
 224                         ip_vs_dest_put(dest);
 225                         RCU_INIT_POINTER(b->dest, NULL);
 226                 }
 227                 b++;
 228         }
 229 }
 230
 231
 232 static int ip_vs_sh_init_svc(struct ip_vs_service *svc)
 233 {
 234         struct ip_vs_sh_state *s;
 235
 236         /* allocate the SH table for this service */
 237         s = kzalloc(sizeof(struct ip_vs_sh_state), GFP_KERNEL);
 238         if (s == NULL)
 239                 return -ENOMEM;
 240
 241         svc->sched_data = s;
 242         IP_VS_DBG(6, "SH hash table (memory=%Zdbytes) allocated for "
 243                   "current service\n",
 244                   sizeof(struct ip_vs_sh_bucket)*IP_VS_SH_TAB_SIZE);
 245
 246         /* assign the hash buckets with current dests */
 247         ip_vs_sh_reassign(s, svc);
 248
 249         return 0;
 250 }
 251
 252
 253 static void ip_vs_sh_done_svc(struct ip_vs_service *svc)
 254 {
 255         struct ip_vs_sh_state *s = svc->sched_data;
 256
 257         /* got to clean up hash buckets here */
 258         ip_vs_sh_flush(s);
 259
 260         /* release the table itself */
 261         kfree_rcu(s, rcu_head);
 262         IP_VS_DBG(6, "SH hash table (memory=%Zdbytes) released\n",
 263                   sizeof(struct ip_vs_sh_bucket)*IP_VS_SH_TAB_SIZE);
 264 }
 265
 266
 267 static int ip_vs_sh_dest_changed(struct ip_vs_service *svc,
 268                                  struct ip_vs_dest *dest)
 269 {
 270         struct ip_vs_sh_state *s = svc->sched_data;
 271
 272         /* assign the hash buckets with the updated service */
 273         ip_vs_sh_reassign(s, svc);
 274
 275         return 0;
 276 }
 277
 278
 279 /* Helper function to get port number */
 280 static inline __be16
 281 ip_vs_sh_get_port(const struct sk_buff *skb, struct ip_vs_iphdr *iph)
 282 {
 283         __be16 _ports[2], *ports;
 284
 285         /* At this point we know that we have a valid packet of some kind.
 286          * Because ICMP packets are only guaranteed to have the first 8
 287          * bytes, let's just grab the ports.  Fortunately they're in the
 288          * same position for all three of the protocols we care about.
 289          */
 290         switch (iph->protocol) {
 291         case IPPROTO_TCP:
 292         case IPPROTO_UDP:
 293         case IPPROTO_SCTP:
 294                 ports = skb_header_pointer(skb, iph->len, sizeof(_ports),
 295                                            &_ports);
 296                 if (unlikely(!ports))
 297                         return 0;
 298
 299                 if (likely(!ip_vs_iph_inverse(iph)))
 300                         return ports[0];
 301                 else
 302                         return ports[1];
 303         default:
 304                 return 0;
 305         }
 306 }
 307
 308
 309 /*
 310  *      Source Hashing scheduling
 311  */
 312 static struct ip_vs_dest *
 313 ip_vs_sh_schedule(struct ip_vs_service *svc, const struct sk_buff *skb,
 314                   struct ip_vs_iphdr *iph)
 315 {
 316         struct ip_vs_dest *dest;
 317         struct ip_vs_sh_state *s;
 318         __be16 port = 0;
 319         const union nf_inet_addr *hash_addr;
 320
 321         hash_addr = ip_vs_iph_inverse(iph) ? &iph->daddr : &iph->saddr;
 322
 323         IP_VS_DBG(6, "ip_vs_sh_schedule(): Scheduling...\n");
 324
 325         if (svc->flags & IP_VS_SVC_F_SCHED_SH_PORT)
 326                 port = ip_vs_sh_get_port(skb, iph);
 327
 328         s = (struct ip_vs_sh_state *) svc->sched_data;
 329
 330         if (svc->flags & IP_VS_SVC_F_SCHED_SH_FALLBACK)
 331                 dest = ip_vs_sh_get_fallback(svc, s, hash_addr, port);
 332         else
 333                 dest = ip_vs_sh_get(svc, s, hash_addr, port);
 334
 335         if (!dest) {
 336                 ip_vs_scheduler_err(svc, "no destination available");
 337                 return NULL;
 338         }
 339
 340         IP_VS_DBG_BUF(6, "SH: source IP address %s --> server %s:%d\n",
 341                       IP_VS_DBG_ADDR(svc->af, hash_addr),
 342                       IP_VS_DBG_ADDR(dest->af, &dest->addr),
 343                       ntohs(dest->port));
 344
 345         return dest;
 346 }
 347
 348
 349 /*
 350  *      IPVS SH Scheduler structure
 351  */
 352 static struct ip_vs_scheduler ip_vs_sh_scheduler =
 353 {
 354         .name =                 "sh",
 355         .refcnt =               ATOMIC_INIT(0),
 356         .module =               THIS_MODULE,
 357         .n_list  =              LIST_HEAD_INIT(ip_vs_sh_scheduler.n_list),
 358         .init_service =         ip_vs_sh_init_svc,
 359         .done_service =         ip_vs_sh_done_svc,
 360         .add_dest =             ip_vs_sh_dest_changed,
 361         .del_dest =             ip_vs_sh_dest_changed,
 362         .upd_dest =             ip_vs_sh_dest_changed,
 363         .schedule =             ip_vs_sh_schedule,
 364 };
 365
 366
 367 static int __init ip_vs_sh_init(void)
 368 {
 369         return register_ip_vs_scheduler(&ip_vs_sh_scheduler);
 370 }
 371
 372
 373 static void __exit ip_vs_sh_cleanup(void)
 374 {
 375         unregister_ip_vs_scheduler(&ip_vs_sh_scheduler);
 376         synchronize_rcu();
 377 }
 378
 379
 380 module_init(ip_vs_sh_init);
 381 module_exit(ip_vs_sh_cleanup);
 382 MODULE_LICENSE("GPL");