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 static inline struct ip_vs_dest *
 120 ip_vs_sh_get_fallback(struct ip_vs_service *svc, struct ip_vs_sh_state *s,
 121                       const union nf_inet_addr *addr, __be16 port)
 122 {
 123         unsigned int offset;
 124         unsigned int hash;
 125         struct ip_vs_dest *dest;
 126
 127         for (offset = 0; offset < IP_VS_SH_TAB_SIZE; offset++) {
 128                 hash = ip_vs_sh_hashkey(svc->af, addr, port, offset);
 129                 dest = rcu_dereference(s->buckets[hash].dest);
 130                 if (!dest)
 131                         break;
 132                 if (is_unavailable(dest))
 133                         IP_VS_DBG_BUF(6, "SH: selected unavailable server "
 134                                       "%s:%d (offset %d)",
 135                                       IP_VS_DBG_ADDR(svc->af, &dest->addr),
 136                                       ntohs(dest->port), offset);
 137                 else
 138                         return dest;
 139         }
 140
 141         return NULL;
 142 }
 143
 144 /*
 145  *      Assign all the hash buckets of the specified table with the service.
 146  */
 147 static int
 148 ip_vs_sh_reassign(struct ip_vs_sh_state *s, struct ip_vs_service *svc)
 149 {
 150         int i;
 151         struct ip_vs_sh_bucket *b;
 152         struct list_head *p;
 153         struct ip_vs_dest *dest;
 154         int d_count;
 155         bool empty;
 156
 157         b = &s->buckets[0];
 158         p = &svc->destinations;
 159         empty = list_empty(p);
 160         d_count = 0;
 161         for (i=0; i<IP_VS_SH_TAB_SIZE; i++) {
 162                 dest = rcu_dereference_protected(b->dest, 1);
 163                 if (dest)
 164                         ip_vs_dest_put(dest);
 165                 if (empty)
 166                         RCU_INIT_POINTER(b->dest, NULL);
 167                 else {
 168                         if (p == &svc->destinations)
 169                                 p = p->next;
 170
 171                         dest = list_entry(p, struct ip_vs_dest, n_list);
 172                         ip_vs_dest_hold(dest);
 173                         RCU_INIT_POINTER(b->dest, dest);
 174
 175                         IP_VS_DBG_BUF(6, "assigned i: %d dest: %s weight: %d\n",
 176                                       i, IP_VS_DBG_ADDR(svc->af, &dest->addr),
 177                                       atomic_read(&dest->weight));
 178
 179                         /* Don't move to next dest until filling weight */
 180                         if (++d_count >= atomic_read(&dest->weight)) {
 181                                 p = p->next;
 182                                 d_count = 0;
 183                         }
 184
 185                 }
 186                 b++;
 187         }
 188         return 0;
 189 }
 190
 191
 192 /*
 193  *      Flush all the hash buckets of the specified table.
 194  */
 195 static void ip_vs_sh_flush(struct ip_vs_sh_state *s)
 196 {
 197         int i;
 198         struct ip_vs_sh_bucket *b;
 199         struct ip_vs_dest *dest;
 200
 201         b = &s->buckets[0];
 202         for (i=0; i<IP_VS_SH_TAB_SIZE; i++) {
 203                 dest = rcu_dereference_protected(b->dest, 1);
 204                 if (dest) {
 205                         ip_vs_dest_put(dest);
 206                         RCU_INIT_POINTER(b->dest, NULL);
 207                 }
 208                 b++;
 209         }
 210 }
 211
 212
 213 static int ip_vs_sh_init_svc(struct ip_vs_service *svc)
 214 {
 215         struct ip_vs_sh_state *s;
 216
 217         /* allocate the SH table for this service */
 218         s = kzalloc(sizeof(struct ip_vs_sh_state), GFP_KERNEL);
 219         if (s == NULL)
 220                 return -ENOMEM;
 221
 222         svc->sched_data = s;
 223         IP_VS_DBG(6, "SH hash table (memory=%Zdbytes) allocated for "
 224                   "current service\n",
 225                   sizeof(struct ip_vs_sh_bucket)*IP_VS_SH_TAB_SIZE);
 226
 227         /* assign the hash buckets with current dests */
 228         ip_vs_sh_reassign(s, svc);
 229
 230         return 0;
 231 }
 232
 233
 234 static void ip_vs_sh_done_svc(struct ip_vs_service *svc)
 235 {
 236         struct ip_vs_sh_state *s = svc->sched_data;
 237
 238         /* got to clean up hash buckets here */
 239         ip_vs_sh_flush(s);
 240
 241         /* release the table itself */
 242         kfree_rcu(s, rcu_head);
 243         IP_VS_DBG(6, "SH hash table (memory=%Zdbytes) released\n",
 244                   sizeof(struct ip_vs_sh_bucket)*IP_VS_SH_TAB_SIZE);
 245 }
 246
 247
 248 static int ip_vs_sh_dest_changed(struct ip_vs_service *svc,
 249                                  struct ip_vs_dest *dest)
 250 {
 251         struct ip_vs_sh_state *s = svc->sched_data;
 252
 253         /* assign the hash buckets with the updated service */
 254         ip_vs_sh_reassign(s, svc);
 255
 256         return 0;
 257 }
 258
 259
 260 /* Helper function to get port number */
 261 static inline __be16
 262 ip_vs_sh_get_port(const struct sk_buff *skb, struct ip_vs_iphdr *iph)
 263 {
 264         __be16 port;
 265         struct tcphdr _tcph, *th;
 266         struct udphdr _udph, *uh;
 267         sctp_sctphdr_t _sctph, *sh;
 268
 269         switch (iph->protocol) {
 270         case IPPROTO_TCP:
 271                 th = skb_header_pointer(skb, iph->len, sizeof(_tcph), &_tcph);
 272                 if (unlikely(th == NULL))
 273                         return 0;
 274                 port = th->source;
 275                 break;
 276         case IPPROTO_UDP:
 277                 uh = skb_header_pointer(skb, iph->len, sizeof(_udph), &_udph);
 278                 if (unlikely(uh == NULL))
 279                         return 0;
 280                 port = uh->source;
 281                 break;
 282         case IPPROTO_SCTP:
 283                 sh = skb_header_pointer(skb, iph->len, sizeof(_sctph), &_sctph);
 284                 if (unlikely(sh == NULL))
 285                         return 0;
 286                 port = sh->source;
 287                 break;
 288         default:
 289                 port = 0;
 290         }
 291
 292         return port;
 293 }
 294
 295
 296 /*
 297  *      Source Hashing scheduling
 298  */
 299 static struct ip_vs_dest *
 300 ip_vs_sh_schedule(struct ip_vs_service *svc, const struct sk_buff *skb,
 301                   struct ip_vs_iphdr *iph)
 302 {
 303         struct ip_vs_dest *dest;
 304         struct ip_vs_sh_state *s;
 305         __be16 port = 0;
 306
 307         IP_VS_DBG(6, "ip_vs_sh_schedule(): Scheduling...\n");
 308
 309         if (svc->flags & IP_VS_SVC_F_SCHED_SH_PORT)
 310                 port = ip_vs_sh_get_port(skb, iph);
 311
 312         s = (struct ip_vs_sh_state *) svc->sched_data;
 313
 314         if (svc->flags & IP_VS_SVC_F_SCHED_SH_FALLBACK)
 315                 dest = ip_vs_sh_get_fallback(svc, s, &iph->saddr, port);
 316         else
 317                 dest = ip_vs_sh_get(svc, s, &iph->saddr, port);
 318
 319         if (!dest) {
 320                 ip_vs_scheduler_err(svc, "no destination available");
 321                 return NULL;
 322         }
 323
 324         IP_VS_DBG_BUF(6, "SH: source IP address %s --> server %s:%d\n",
 325                       IP_VS_DBG_ADDR(svc->af, &iph->saddr),
 326                       IP_VS_DBG_ADDR(svc->af, &dest->addr),
 327                       ntohs(dest->port));
 328
 329         return dest;
 330 }
 331
 332
 333 /*
 334  *      IPVS SH Scheduler structure
 335  */
 336 static struct ip_vs_scheduler ip_vs_sh_scheduler =
 337 {
 338         .name =                 "sh",
 339         .refcnt =               ATOMIC_INIT(0),
 340         .module =               THIS_MODULE,
 341         .n_list  =              LIST_HEAD_INIT(ip_vs_sh_scheduler.n_list),
 342         .init_service =         ip_vs_sh_init_svc,
 343         .done_service =         ip_vs_sh_done_svc,
 344         .add_dest =             ip_vs_sh_dest_changed,
 345         .del_dest =             ip_vs_sh_dest_changed,
 346         .upd_dest =             ip_vs_sh_dest_changed,
 347         .schedule =             ip_vs_sh_schedule,
 348 };
 349
 350
 351 static int __init ip_vs_sh_init(void)
 352 {
 353         return register_ip_vs_scheduler(&ip_vs_sh_scheduler);
 354 }
 355
 356
 357 static void __exit ip_vs_sh_cleanup(void)
 358 {
 359         unregister_ip_vs_scheduler(&ip_vs_sh_scheduler);
 360         synchronize_rcu();
 361 }
 362
 363
 364 module_init(ip_vs_sh_init);
 365 module_exit(ip_vs_sh_cleanup);
 366 MODULE_LICENSE("GPL");