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 + hash_32(ntohs(port) + ntohl(addr_fold), 100 IP_VS_SH_TAB_BITS)) & 101 IP_VS_SH_TAB_MASK; 102 } 103 104 105 /* 106 * Get ip_vs_dest associated with supplied parameters. 107 */ 108 static inline struct ip_vs_dest * 109 ip_vs_sh_get(struct ip_vs_service *svc, struct ip_vs_sh_state *s, 110 const union nf_inet_addr *addr, __be16 port) 111 { 112 unsigned int hash = ip_vs_sh_hashkey(svc->af, addr, port, 0); 113 struct ip_vs_dest *dest = rcu_dereference(s->buckets[hash].dest); 114 115 return (!dest || is_unavailable(dest)) ? NULL : dest; 116 } 117 118 119 /* As ip_vs_sh_get, but with fallback if selected server is unavailable 120 * 121 * The fallback strategy loops around the table starting from a "random" 122 * point (in fact, it is chosen to be the original hash value to make the 123 * algorithm deterministic) to find a new server. 124 */ 125 static inline struct ip_vs_dest * 126 ip_vs_sh_get_fallback(struct ip_vs_service *svc, struct ip_vs_sh_state *s, 127 const union nf_inet_addr *addr, __be16 port) 128 { 129 unsigned int offset, roffset; 130 unsigned int hash, ihash; 131 struct ip_vs_dest *dest; 132 133 /* first try the dest it's supposed to go to */ 134 ihash = ip_vs_sh_hashkey(svc->af, addr, port, 0); 135 dest = rcu_dereference(s->buckets[ihash].dest); 136 if (!dest) 137 return NULL; 138 if (!is_unavailable(dest)) 139 return dest; 140 141 IP_VS_DBG_BUF(6, "SH: selected unavailable server %s:%d, reselecting", 142 IP_VS_DBG_ADDR(dest->af, &dest->addr), ntohs(dest->port)); 143 144 /* if the original dest is unavailable, loop around the table 145 * starting from ihash to find a new dest 146 */ 147 for (offset = 0; offset < IP_VS_SH_TAB_SIZE; offset++) { 148 roffset = (offset + ihash) % IP_VS_SH_TAB_SIZE; 149 hash = ip_vs_sh_hashkey(svc->af, addr, port, roffset); 150 dest = rcu_dereference(s->buckets[hash].dest); 151 if (!dest) 152 break; 153 if (!is_unavailable(dest)) 154 return dest; 155 IP_VS_DBG_BUF(6, "SH: selected unavailable " 156 "server %s:%d (offset %d), reselecting", 157 IP_VS_DBG_ADDR(dest->af, &dest->addr), 158 ntohs(dest->port), roffset); 159 } 160 161 return NULL; 162 } 163 164 /* 165 * Assign all the hash buckets of the specified table with the service. 166 */ 167 static int 168 ip_vs_sh_reassign(struct ip_vs_sh_state *s, struct ip_vs_service *svc) 169 { 170 int i; 171 struct ip_vs_sh_bucket *b; 172 struct list_head *p; 173 struct ip_vs_dest *dest; 174 int d_count; 175 bool empty; 176 177 b = &s->buckets[0]; 178 p = &svc->destinations; 179 empty = list_empty(p); 180 d_count = 0; 181 for (i=0; i<IP_VS_SH_TAB_SIZE; i++) { 182 dest = rcu_dereference_protected(b->dest, 1); 183 if (dest) 184 ip_vs_dest_put(dest); 185 if (empty) 186 RCU_INIT_POINTER(b->dest, NULL); 187 else { 188 if (p == &svc->destinations) 189 p = p->next; 190 191 dest = list_entry(p, struct ip_vs_dest, n_list); 192 ip_vs_dest_hold(dest); 193 RCU_INIT_POINTER(b->dest, dest); 194 195 IP_VS_DBG_BUF(6, "assigned i: %d dest: %s weight: %d\n", 196 i, IP_VS_DBG_ADDR(dest->af, &dest->addr), 197 atomic_read(&dest->weight)); 198 199 /* Don't move to next dest until filling weight */ 200 if (++d_count >= atomic_read(&dest->weight)) { 201 p = p->next; 202 d_count = 0; 203 } 204 205 } 206 b++; 207 } 208 return 0; 209 } 210 211 212 /* 213 * Flush all the hash buckets of the specified table. 214 */ 215 static void ip_vs_sh_flush(struct ip_vs_sh_state *s) 216 { 217 int i; 218 struct ip_vs_sh_bucket *b; 219 struct ip_vs_dest *dest; 220 221 b = &s->buckets[0]; 222 for (i=0; i<IP_VS_SH_TAB_SIZE; i++) { 223 dest = rcu_dereference_protected(b->dest, 1); 224 if (dest) { 225 ip_vs_dest_put(dest); 226 RCU_INIT_POINTER(b->dest, NULL); 227 } 228 b++; 229 } 230 } 231 232 233 static int ip_vs_sh_init_svc(struct ip_vs_service *svc) 234 { 235 struct ip_vs_sh_state *s; 236 237 /* allocate the SH table for this service */ 238 s = kzalloc(sizeof(struct ip_vs_sh_state), GFP_KERNEL); 239 if (s == NULL) 240 return -ENOMEM; 241 242 svc->sched_data = s; 243 IP_VS_DBG(6, "SH hash table (memory=%zdbytes) allocated for " 244 "current service\n", 245 sizeof(struct ip_vs_sh_bucket)*IP_VS_SH_TAB_SIZE); 246 247 /* assign the hash buckets with current dests */ 248 ip_vs_sh_reassign(s, svc); 249 250 return 0; 251 } 252 253 254 static void ip_vs_sh_done_svc(struct ip_vs_service *svc) 255 { 256 struct ip_vs_sh_state *s = svc->sched_data; 257 258 /* got to clean up hash buckets here */ 259 ip_vs_sh_flush(s); 260 261 /* release the table itself */ 262 kfree_rcu(s, rcu_head); 263 IP_VS_DBG(6, "SH hash table (memory=%zdbytes) released\n", 264 sizeof(struct ip_vs_sh_bucket)*IP_VS_SH_TAB_SIZE); 265 } 266 267 268 static int ip_vs_sh_dest_changed(struct ip_vs_service *svc, 269 struct ip_vs_dest *dest) 270 { 271 struct ip_vs_sh_state *s = svc->sched_data; 272 273 /* assign the hash buckets with the updated service */ 274 ip_vs_sh_reassign(s, svc); 275 276 return 0; 277 } 278 279 280 /* Helper function to get port number */ 281 static inline __be16 282 ip_vs_sh_get_port(const struct sk_buff *skb, struct ip_vs_iphdr *iph) 283 { 284 __be16 _ports[2], *ports; 285 286 /* At this point we know that we have a valid packet of some kind. 287 * Because ICMP packets are only guaranteed to have the first 8 288 * bytes, let's just grab the ports. Fortunately they're in the 289 * same position for all three of the protocols we care about. 290 */ 291 switch (iph->protocol) { 292 case IPPROTO_TCP: 293 case IPPROTO_UDP: 294 case IPPROTO_SCTP: 295 ports = skb_header_pointer(skb, iph->len, sizeof(_ports), 296 &_ports); 297 if (unlikely(!ports)) 298 return 0; 299 300 if (likely(!ip_vs_iph_inverse(iph))) 301 return ports[0]; 302 else 303 return ports[1]; 304 default: 305 return 0; 306 } 307 } 308 309 310 /* 311 * Source Hashing scheduling 312 */ 313 static struct ip_vs_dest * 314 ip_vs_sh_schedule(struct ip_vs_service *svc, const struct sk_buff *skb, 315 struct ip_vs_iphdr *iph) 316 { 317 struct ip_vs_dest *dest; 318 struct ip_vs_sh_state *s; 319 __be16 port = 0; 320 const union nf_inet_addr *hash_addr; 321 322 hash_addr = ip_vs_iph_inverse(iph) ? &iph->daddr : &iph->saddr; 323 324 IP_VS_DBG(6, "ip_vs_sh_schedule(): Scheduling...\n"); 325 326 if (svc->flags & IP_VS_SVC_F_SCHED_SH_PORT) 327 port = ip_vs_sh_get_port(skb, iph); 328 329 s = (struct ip_vs_sh_state *) svc->sched_data; 330 331 if (svc->flags & IP_VS_SVC_F_SCHED_SH_FALLBACK) 332 dest = ip_vs_sh_get_fallback(svc, s, hash_addr, port); 333 else 334 dest = ip_vs_sh_get(svc, s, hash_addr, port); 335 336 if (!dest) { 337 ip_vs_scheduler_err(svc, "no destination available"); 338 return NULL; 339 } 340 341 IP_VS_DBG_BUF(6, "SH: source IP address %s --> server %s:%d\n", 342 IP_VS_DBG_ADDR(svc->af, hash_addr), 343 IP_VS_DBG_ADDR(dest->af, &dest->addr), 344 ntohs(dest->port)); 345 346 return dest; 347 } 348 349 350 /* 351 * IPVS SH Scheduler structure 352 */ 353 static struct ip_vs_scheduler ip_vs_sh_scheduler = 354 { 355 .name = "sh", 356 .refcnt = ATOMIC_INIT(0), 357 .module = THIS_MODULE, 358 .n_list = LIST_HEAD_INIT(ip_vs_sh_scheduler.n_list), 359 .init_service = ip_vs_sh_init_svc, 360 .done_service = ip_vs_sh_done_svc, 361 .add_dest = ip_vs_sh_dest_changed, 362 .del_dest = ip_vs_sh_dest_changed, 363 .upd_dest = ip_vs_sh_dest_changed, 364 .schedule = ip_vs_sh_schedule, 365 }; 366 367 368 static int __init ip_vs_sh_init(void) 369 { 370 return register_ip_vs_scheduler(&ip_vs_sh_scheduler); 371 } 372 373 374 static void __exit ip_vs_sh_cleanup(void) 375 { 376 unregister_ip_vs_scheduler(&ip_vs_sh_scheduler); 377 synchronize_rcu(); 378 } 379 380 381 module_init(ip_vs_sh_init); 382 module_exit(ip_vs_sh_cleanup); 383 MODULE_LICENSE("GPL"); 384