Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * IPVS: Locality-Based Least-Connection with Replication scheduler |
| 3 | * |
| 4 | * Version: $Id: ip_vs_lblcr.c,v 1.11 2002/09/15 08:14:08 wensong Exp $ |
| 5 | * |
| 6 | * Authors: Wensong Zhang <wensong@gnuchina.org> |
| 7 | * |
| 8 | * This program is free software; you can redistribute it and/or |
| 9 | * modify it under the terms of the GNU General Public License |
| 10 | * as published by the Free Software Foundation; either version |
| 11 | * 2 of the License, or (at your option) any later version. |
| 12 | * |
| 13 | * Changes: |
| 14 | * Julian Anastasov : Added the missing (dest->weight>0) |
| 15 | * condition in the ip_vs_dest_set_max. |
| 16 | * |
| 17 | */ |
| 18 | |
| 19 | /* |
| 20 | * The lblc/r algorithm is as follows (pseudo code): |
| 21 | * |
| 22 | * if serverSet[dest_ip] is null then |
| 23 | * n, serverSet[dest_ip] <- {weighted least-conn node}; |
| 24 | * else |
| 25 | * n <- {least-conn (alive) node in serverSet[dest_ip]}; |
| 26 | * if (n is null) OR |
| 27 | * (n.conns>n.weight AND |
| 28 | * there is a node m with m.conns<m.weight/2) then |
| 29 | * n <- {weighted least-conn node}; |
| 30 | * add n to serverSet[dest_ip]; |
| 31 | * if |serverSet[dest_ip]| > 1 AND |
| 32 | * now - serverSet[dest_ip].lastMod > T then |
| 33 | * m <- {most conn node in serverSet[dest_ip]}; |
| 34 | * remove m from serverSet[dest_ip]; |
| 35 | * if serverSet[dest_ip] changed then |
| 36 | * serverSet[dest_ip].lastMod <- now; |
| 37 | * |
| 38 | * return n; |
| 39 | * |
| 40 | */ |
| 41 | |
Arnaldo Carvalho de Melo | 14c8502 | 2005-12-27 02:43:12 -0200 | [diff] [blame] | 42 | #include <linux/ip.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 43 | #include <linux/module.h> |
| 44 | #include <linux/kernel.h> |
Arnaldo Carvalho de Melo | 14c8502 | 2005-12-27 02:43:12 -0200 | [diff] [blame] | 45 | #include <linux/skbuff.h> |
Al Viro | d7fe0f2 | 2006-12-03 23:15:30 -0500 | [diff] [blame] | 46 | #include <linux/jiffies.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 47 | |
| 48 | /* for sysctl */ |
| 49 | #include <linux/fs.h> |
| 50 | #include <linux/sysctl.h> |
| 51 | /* for proc_net_create/proc_net_remove */ |
| 52 | #include <linux/proc_fs.h> |
Eric W. Biederman | 457c4cb | 2007-09-12 12:01:34 +0200 | [diff] [blame^] | 53 | #include <net/net_namespace.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 54 | |
| 55 | #include <net/ip_vs.h> |
| 56 | |
| 57 | |
| 58 | /* |
| 59 | * It is for garbage collection of stale IPVS lblcr entries, |
| 60 | * when the table is full. |
| 61 | */ |
| 62 | #define CHECK_EXPIRE_INTERVAL (60*HZ) |
| 63 | #define ENTRY_TIMEOUT (6*60*HZ) |
| 64 | |
| 65 | /* |
| 66 | * It is for full expiration check. |
| 67 | * When there is no partial expiration check (garbage collection) |
| 68 | * in a half hour, do a full expiration check to collect stale |
| 69 | * entries that haven't been touched for a day. |
| 70 | */ |
| 71 | #define COUNT_FOR_FULL_EXPIRATION 30 |
| 72 | static int sysctl_ip_vs_lblcr_expiration = 24*60*60*HZ; |
| 73 | |
| 74 | |
| 75 | /* |
| 76 | * for IPVS lblcr entry hash table |
| 77 | */ |
| 78 | #ifndef CONFIG_IP_VS_LBLCR_TAB_BITS |
| 79 | #define CONFIG_IP_VS_LBLCR_TAB_BITS 10 |
| 80 | #endif |
| 81 | #define IP_VS_LBLCR_TAB_BITS CONFIG_IP_VS_LBLCR_TAB_BITS |
| 82 | #define IP_VS_LBLCR_TAB_SIZE (1 << IP_VS_LBLCR_TAB_BITS) |
| 83 | #define IP_VS_LBLCR_TAB_MASK (IP_VS_LBLCR_TAB_SIZE - 1) |
| 84 | |
| 85 | |
| 86 | /* |
| 87 | * IPVS destination set structure and operations |
| 88 | */ |
| 89 | struct ip_vs_dest_list { |
| 90 | struct ip_vs_dest_list *next; /* list link */ |
| 91 | struct ip_vs_dest *dest; /* destination server */ |
| 92 | }; |
| 93 | |
| 94 | struct ip_vs_dest_set { |
| 95 | atomic_t size; /* set size */ |
| 96 | unsigned long lastmod; /* last modified time */ |
| 97 | struct ip_vs_dest_list *list; /* destination list */ |
| 98 | rwlock_t lock; /* lock for this list */ |
| 99 | }; |
| 100 | |
| 101 | |
| 102 | static struct ip_vs_dest_list * |
| 103 | ip_vs_dest_set_insert(struct ip_vs_dest_set *set, struct ip_vs_dest *dest) |
| 104 | { |
| 105 | struct ip_vs_dest_list *e; |
| 106 | |
| 107 | for (e=set->list; e!=NULL; e=e->next) { |
| 108 | if (e->dest == dest) |
| 109 | /* already existed */ |
| 110 | return NULL; |
| 111 | } |
| 112 | |
| 113 | e = kmalloc(sizeof(struct ip_vs_dest_list), GFP_ATOMIC); |
| 114 | if (e == NULL) { |
| 115 | IP_VS_ERR("ip_vs_dest_set_insert(): no memory\n"); |
| 116 | return NULL; |
| 117 | } |
| 118 | |
| 119 | atomic_inc(&dest->refcnt); |
| 120 | e->dest = dest; |
| 121 | |
| 122 | /* link it to the list */ |
| 123 | write_lock(&set->lock); |
| 124 | e->next = set->list; |
| 125 | set->list = e; |
| 126 | atomic_inc(&set->size); |
| 127 | write_unlock(&set->lock); |
| 128 | |
| 129 | set->lastmod = jiffies; |
| 130 | return e; |
| 131 | } |
| 132 | |
| 133 | static void |
| 134 | ip_vs_dest_set_erase(struct ip_vs_dest_set *set, struct ip_vs_dest *dest) |
| 135 | { |
| 136 | struct ip_vs_dest_list *e, **ep; |
| 137 | |
| 138 | write_lock(&set->lock); |
| 139 | for (ep=&set->list, e=*ep; e!=NULL; e=*ep) { |
| 140 | if (e->dest == dest) { |
| 141 | /* HIT */ |
| 142 | *ep = e->next; |
| 143 | atomic_dec(&set->size); |
| 144 | set->lastmod = jiffies; |
| 145 | atomic_dec(&e->dest->refcnt); |
| 146 | kfree(e); |
| 147 | break; |
| 148 | } |
| 149 | ep = &e->next; |
| 150 | } |
| 151 | write_unlock(&set->lock); |
| 152 | } |
| 153 | |
| 154 | static void ip_vs_dest_set_eraseall(struct ip_vs_dest_set *set) |
| 155 | { |
| 156 | struct ip_vs_dest_list *e, **ep; |
| 157 | |
| 158 | write_lock(&set->lock); |
| 159 | for (ep=&set->list, e=*ep; e!=NULL; e=*ep) { |
| 160 | *ep = e->next; |
| 161 | /* |
| 162 | * We don't kfree dest because it is refered either |
| 163 | * by its service or by the trash dest list. |
| 164 | */ |
| 165 | atomic_dec(&e->dest->refcnt); |
| 166 | kfree(e); |
| 167 | } |
| 168 | write_unlock(&set->lock); |
| 169 | } |
| 170 | |
| 171 | /* get weighted least-connection node in the destination set */ |
| 172 | static inline struct ip_vs_dest *ip_vs_dest_set_min(struct ip_vs_dest_set *set) |
| 173 | { |
| 174 | register struct ip_vs_dest_list *e; |
| 175 | struct ip_vs_dest *dest, *least; |
| 176 | int loh, doh; |
| 177 | |
| 178 | if (set == NULL) |
| 179 | return NULL; |
| 180 | |
| 181 | read_lock(&set->lock); |
| 182 | /* select the first destination server, whose weight > 0 */ |
| 183 | for (e=set->list; e!=NULL; e=e->next) { |
| 184 | least = e->dest; |
| 185 | if (least->flags & IP_VS_DEST_F_OVERLOAD) |
| 186 | continue; |
| 187 | |
| 188 | if ((atomic_read(&least->weight) > 0) |
| 189 | && (least->flags & IP_VS_DEST_F_AVAILABLE)) { |
| 190 | loh = atomic_read(&least->activeconns) * 50 |
| 191 | + atomic_read(&least->inactconns); |
| 192 | goto nextstage; |
| 193 | } |
| 194 | } |
| 195 | read_unlock(&set->lock); |
| 196 | return NULL; |
| 197 | |
| 198 | /* find the destination with the weighted least load */ |
| 199 | nextstage: |
| 200 | for (e=e->next; e!=NULL; e=e->next) { |
| 201 | dest = e->dest; |
| 202 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
| 203 | continue; |
| 204 | |
| 205 | doh = atomic_read(&dest->activeconns) * 50 |
| 206 | + atomic_read(&dest->inactconns); |
| 207 | if ((loh * atomic_read(&dest->weight) > |
| 208 | doh * atomic_read(&least->weight)) |
| 209 | && (dest->flags & IP_VS_DEST_F_AVAILABLE)) { |
| 210 | least = dest; |
| 211 | loh = doh; |
| 212 | } |
| 213 | } |
| 214 | read_unlock(&set->lock); |
| 215 | |
| 216 | IP_VS_DBG(6, "ip_vs_dest_set_min: server %d.%d.%d.%d:%d " |
| 217 | "activeconns %d refcnt %d weight %d overhead %d\n", |
| 218 | NIPQUAD(least->addr), ntohs(least->port), |
| 219 | atomic_read(&least->activeconns), |
| 220 | atomic_read(&least->refcnt), |
| 221 | atomic_read(&least->weight), loh); |
| 222 | return least; |
| 223 | } |
| 224 | |
| 225 | |
| 226 | /* get weighted most-connection node in the destination set */ |
| 227 | static inline struct ip_vs_dest *ip_vs_dest_set_max(struct ip_vs_dest_set *set) |
| 228 | { |
| 229 | register struct ip_vs_dest_list *e; |
| 230 | struct ip_vs_dest *dest, *most; |
| 231 | int moh, doh; |
| 232 | |
| 233 | if (set == NULL) |
| 234 | return NULL; |
| 235 | |
| 236 | read_lock(&set->lock); |
| 237 | /* select the first destination server, whose weight > 0 */ |
| 238 | for (e=set->list; e!=NULL; e=e->next) { |
| 239 | most = e->dest; |
| 240 | if (atomic_read(&most->weight) > 0) { |
| 241 | moh = atomic_read(&most->activeconns) * 50 |
| 242 | + atomic_read(&most->inactconns); |
| 243 | goto nextstage; |
| 244 | } |
| 245 | } |
| 246 | read_unlock(&set->lock); |
| 247 | return NULL; |
| 248 | |
| 249 | /* find the destination with the weighted most load */ |
| 250 | nextstage: |
| 251 | for (e=e->next; e!=NULL; e=e->next) { |
| 252 | dest = e->dest; |
| 253 | doh = atomic_read(&dest->activeconns) * 50 |
| 254 | + atomic_read(&dest->inactconns); |
| 255 | /* moh/mw < doh/dw ==> moh*dw < doh*mw, where mw,dw>0 */ |
| 256 | if ((moh * atomic_read(&dest->weight) < |
| 257 | doh * atomic_read(&most->weight)) |
| 258 | && (atomic_read(&dest->weight) > 0)) { |
| 259 | most = dest; |
| 260 | moh = doh; |
| 261 | } |
| 262 | } |
| 263 | read_unlock(&set->lock); |
| 264 | |
| 265 | IP_VS_DBG(6, "ip_vs_dest_set_max: server %d.%d.%d.%d:%d " |
| 266 | "activeconns %d refcnt %d weight %d overhead %d\n", |
| 267 | NIPQUAD(most->addr), ntohs(most->port), |
| 268 | atomic_read(&most->activeconns), |
| 269 | atomic_read(&most->refcnt), |
| 270 | atomic_read(&most->weight), moh); |
| 271 | return most; |
| 272 | } |
| 273 | |
| 274 | |
| 275 | /* |
| 276 | * IPVS lblcr entry represents an association between destination |
| 277 | * IP address and its destination server set |
| 278 | */ |
| 279 | struct ip_vs_lblcr_entry { |
| 280 | struct list_head list; |
Al Viro | 014d730 | 2006-09-28 14:29:52 -0700 | [diff] [blame] | 281 | __be32 addr; /* destination IP address */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 282 | struct ip_vs_dest_set set; /* destination server set */ |
| 283 | unsigned long lastuse; /* last used time */ |
| 284 | }; |
| 285 | |
| 286 | |
| 287 | /* |
| 288 | * IPVS lblcr hash table |
| 289 | */ |
| 290 | struct ip_vs_lblcr_table { |
| 291 | rwlock_t lock; /* lock for this table */ |
| 292 | struct list_head bucket[IP_VS_LBLCR_TAB_SIZE]; /* hash bucket */ |
| 293 | atomic_t entries; /* number of entries */ |
| 294 | int max_size; /* maximum size of entries */ |
| 295 | struct timer_list periodic_timer; /* collect stale entries */ |
| 296 | int rover; /* rover for expire check */ |
| 297 | int counter; /* counter for no expire */ |
| 298 | }; |
| 299 | |
| 300 | |
| 301 | /* |
| 302 | * IPVS LBLCR sysctl table |
| 303 | */ |
| 304 | |
| 305 | static ctl_table vs_vars_table[] = { |
| 306 | { |
| 307 | .ctl_name = NET_IPV4_VS_LBLCR_EXPIRE, |
| 308 | .procname = "lblcr_expiration", |
| 309 | .data = &sysctl_ip_vs_lblcr_expiration, |
| 310 | .maxlen = sizeof(int), |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 311 | .mode = 0644, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 312 | .proc_handler = &proc_dointvec_jiffies, |
| 313 | }, |
| 314 | { .ctl_name = 0 } |
| 315 | }; |
| 316 | |
| 317 | static ctl_table vs_table[] = { |
| 318 | { |
| 319 | .ctl_name = NET_IPV4_VS, |
| 320 | .procname = "vs", |
| 321 | .mode = 0555, |
| 322 | .child = vs_vars_table |
| 323 | }, |
| 324 | { .ctl_name = 0 } |
| 325 | }; |
| 326 | |
David S. Miller | bf0ff9e | 2005-08-19 16:37:30 -0700 | [diff] [blame] | 327 | static ctl_table ipvs_ipv4_table[] = { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 328 | { |
| 329 | .ctl_name = NET_IPV4, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 330 | .procname = "ipv4", |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 331 | .mode = 0555, |
| 332 | .child = vs_table |
| 333 | }, |
| 334 | { .ctl_name = 0 } |
| 335 | }; |
| 336 | |
| 337 | static ctl_table lblcr_root_table[] = { |
| 338 | { |
| 339 | .ctl_name = CTL_NET, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 340 | .procname = "net", |
| 341 | .mode = 0555, |
David S. Miller | bf0ff9e | 2005-08-19 16:37:30 -0700 | [diff] [blame] | 342 | .child = ipvs_ipv4_table |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 343 | }, |
| 344 | { .ctl_name = 0 } |
| 345 | }; |
| 346 | |
| 347 | static struct ctl_table_header * sysctl_header; |
| 348 | |
| 349 | /* |
| 350 | * new/free a ip_vs_lblcr_entry, which is a mapping of a destination |
| 351 | * IP address to a server. |
| 352 | */ |
Al Viro | 014d730 | 2006-09-28 14:29:52 -0700 | [diff] [blame] | 353 | static inline struct ip_vs_lblcr_entry *ip_vs_lblcr_new(__be32 daddr) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 354 | { |
| 355 | struct ip_vs_lblcr_entry *en; |
| 356 | |
| 357 | en = kmalloc(sizeof(struct ip_vs_lblcr_entry), GFP_ATOMIC); |
| 358 | if (en == NULL) { |
| 359 | IP_VS_ERR("ip_vs_lblcr_new(): no memory\n"); |
| 360 | return NULL; |
| 361 | } |
| 362 | |
| 363 | INIT_LIST_HEAD(&en->list); |
| 364 | en->addr = daddr; |
| 365 | |
| 366 | /* initilize its dest set */ |
| 367 | atomic_set(&(en->set.size), 0); |
| 368 | en->set.list = NULL; |
| 369 | rwlock_init(&en->set.lock); |
| 370 | |
| 371 | return en; |
| 372 | } |
| 373 | |
| 374 | |
| 375 | static inline void ip_vs_lblcr_free(struct ip_vs_lblcr_entry *en) |
| 376 | { |
| 377 | list_del(&en->list); |
| 378 | ip_vs_dest_set_eraseall(&en->set); |
| 379 | kfree(en); |
| 380 | } |
| 381 | |
| 382 | |
| 383 | /* |
| 384 | * Returns hash value for IPVS LBLCR entry |
| 385 | */ |
Al Viro | 014d730 | 2006-09-28 14:29:52 -0700 | [diff] [blame] | 386 | static inline unsigned ip_vs_lblcr_hashkey(__be32 addr) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 387 | { |
| 388 | return (ntohl(addr)*2654435761UL) & IP_VS_LBLCR_TAB_MASK; |
| 389 | } |
| 390 | |
| 391 | |
| 392 | /* |
| 393 | * Hash an entry in the ip_vs_lblcr_table. |
| 394 | * returns bool success. |
| 395 | */ |
| 396 | static int |
| 397 | ip_vs_lblcr_hash(struct ip_vs_lblcr_table *tbl, struct ip_vs_lblcr_entry *en) |
| 398 | { |
| 399 | unsigned hash; |
| 400 | |
| 401 | if (!list_empty(&en->list)) { |
| 402 | IP_VS_ERR("ip_vs_lblcr_hash(): request for already hashed, " |
| 403 | "called from %p\n", __builtin_return_address(0)); |
| 404 | return 0; |
| 405 | } |
| 406 | |
| 407 | /* |
| 408 | * Hash by destination IP address |
| 409 | */ |
| 410 | hash = ip_vs_lblcr_hashkey(en->addr); |
| 411 | |
| 412 | write_lock(&tbl->lock); |
| 413 | list_add(&en->list, &tbl->bucket[hash]); |
| 414 | atomic_inc(&tbl->entries); |
| 415 | write_unlock(&tbl->lock); |
| 416 | |
| 417 | return 1; |
| 418 | } |
| 419 | |
| 420 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 421 | /* |
| 422 | * Get ip_vs_lblcr_entry associated with supplied parameters. |
| 423 | */ |
| 424 | static inline struct ip_vs_lblcr_entry * |
Al Viro | 014d730 | 2006-09-28 14:29:52 -0700 | [diff] [blame] | 425 | ip_vs_lblcr_get(struct ip_vs_lblcr_table *tbl, __be32 addr) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 426 | { |
| 427 | unsigned hash; |
| 428 | struct ip_vs_lblcr_entry *en; |
| 429 | |
| 430 | hash = ip_vs_lblcr_hashkey(addr); |
| 431 | |
| 432 | read_lock(&tbl->lock); |
| 433 | |
| 434 | list_for_each_entry(en, &tbl->bucket[hash], list) { |
| 435 | if (en->addr == addr) { |
| 436 | /* HIT */ |
| 437 | read_unlock(&tbl->lock); |
| 438 | return en; |
| 439 | } |
| 440 | } |
| 441 | |
| 442 | read_unlock(&tbl->lock); |
| 443 | |
| 444 | return NULL; |
| 445 | } |
| 446 | |
| 447 | |
| 448 | /* |
| 449 | * Flush all the entries of the specified table. |
| 450 | */ |
| 451 | static void ip_vs_lblcr_flush(struct ip_vs_lblcr_table *tbl) |
| 452 | { |
| 453 | int i; |
| 454 | struct ip_vs_lblcr_entry *en, *nxt; |
| 455 | |
| 456 | for (i=0; i<IP_VS_LBLCR_TAB_SIZE; i++) { |
| 457 | write_lock(&tbl->lock); |
| 458 | list_for_each_entry_safe(en, nxt, &tbl->bucket[i], list) { |
| 459 | ip_vs_lblcr_free(en); |
| 460 | atomic_dec(&tbl->entries); |
| 461 | } |
| 462 | write_unlock(&tbl->lock); |
| 463 | } |
| 464 | } |
| 465 | |
| 466 | |
| 467 | static inline void ip_vs_lblcr_full_check(struct ip_vs_lblcr_table *tbl) |
| 468 | { |
| 469 | unsigned long now = jiffies; |
| 470 | int i, j; |
| 471 | struct ip_vs_lblcr_entry *en, *nxt; |
| 472 | |
| 473 | for (i=0, j=tbl->rover; i<IP_VS_LBLCR_TAB_SIZE; i++) { |
| 474 | j = (j + 1) & IP_VS_LBLCR_TAB_MASK; |
| 475 | |
| 476 | write_lock(&tbl->lock); |
| 477 | list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) { |
| 478 | if (time_after(en->lastuse+sysctl_ip_vs_lblcr_expiration, |
| 479 | now)) |
| 480 | continue; |
| 481 | |
| 482 | ip_vs_lblcr_free(en); |
| 483 | atomic_dec(&tbl->entries); |
| 484 | } |
| 485 | write_unlock(&tbl->lock); |
| 486 | } |
| 487 | tbl->rover = j; |
| 488 | } |
| 489 | |
| 490 | |
| 491 | /* |
| 492 | * Periodical timer handler for IPVS lblcr table |
| 493 | * It is used to collect stale entries when the number of entries |
| 494 | * exceeds the maximum size of the table. |
| 495 | * |
| 496 | * Fixme: we probably need more complicated algorithm to collect |
| 497 | * entries that have not been used for a long time even |
| 498 | * if the number of entries doesn't exceed the maximum size |
| 499 | * of the table. |
| 500 | * The full expiration check is for this purpose now. |
| 501 | */ |
| 502 | static void ip_vs_lblcr_check_expire(unsigned long data) |
| 503 | { |
| 504 | struct ip_vs_lblcr_table *tbl; |
| 505 | unsigned long now = jiffies; |
| 506 | int goal; |
| 507 | int i, j; |
| 508 | struct ip_vs_lblcr_entry *en, *nxt; |
| 509 | |
| 510 | tbl = (struct ip_vs_lblcr_table *)data; |
| 511 | |
| 512 | if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) { |
| 513 | /* do full expiration check */ |
| 514 | ip_vs_lblcr_full_check(tbl); |
| 515 | tbl->counter = 1; |
| 516 | goto out; |
| 517 | } |
| 518 | |
| 519 | if (atomic_read(&tbl->entries) <= tbl->max_size) { |
| 520 | tbl->counter++; |
| 521 | goto out; |
| 522 | } |
| 523 | |
| 524 | goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3; |
| 525 | if (goal > tbl->max_size/2) |
| 526 | goal = tbl->max_size/2; |
| 527 | |
| 528 | for (i=0, j=tbl->rover; i<IP_VS_LBLCR_TAB_SIZE; i++) { |
| 529 | j = (j + 1) & IP_VS_LBLCR_TAB_MASK; |
| 530 | |
| 531 | write_lock(&tbl->lock); |
| 532 | list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) { |
| 533 | if (time_before(now, en->lastuse+ENTRY_TIMEOUT)) |
| 534 | continue; |
| 535 | |
| 536 | ip_vs_lblcr_free(en); |
| 537 | atomic_dec(&tbl->entries); |
| 538 | goal--; |
| 539 | } |
| 540 | write_unlock(&tbl->lock); |
| 541 | if (goal <= 0) |
| 542 | break; |
| 543 | } |
| 544 | tbl->rover = j; |
| 545 | |
| 546 | out: |
| 547 | mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL); |
| 548 | } |
| 549 | |
| 550 | |
| 551 | #ifdef CONFIG_IP_VS_LBLCR_DEBUG |
| 552 | static struct ip_vs_lblcr_table *lblcr_table_list; |
| 553 | |
| 554 | /* |
| 555 | * /proc/net/ip_vs_lblcr to display the mappings of |
| 556 | * destination IP address <==> its serverSet |
| 557 | */ |
| 558 | static int |
| 559 | ip_vs_lblcr_getinfo(char *buffer, char **start, off_t offset, int length) |
| 560 | { |
| 561 | off_t pos=0, begin; |
| 562 | int len=0, size; |
| 563 | struct ip_vs_lblcr_table *tbl; |
| 564 | unsigned long now = jiffies; |
| 565 | int i; |
| 566 | struct ip_vs_lblcr_entry *en; |
| 567 | |
| 568 | tbl = lblcr_table_list; |
| 569 | |
| 570 | size = sprintf(buffer, "LastTime Dest IP address Server set\n"); |
| 571 | pos += size; |
| 572 | len += size; |
| 573 | |
| 574 | for (i=0; i<IP_VS_LBLCR_TAB_SIZE; i++) { |
| 575 | read_lock_bh(&tbl->lock); |
| 576 | list_for_each_entry(en, &tbl->bucket[i], list) { |
| 577 | char tbuf[16]; |
| 578 | struct ip_vs_dest_list *d; |
| 579 | |
| 580 | sprintf(tbuf, "%u.%u.%u.%u", NIPQUAD(en->addr)); |
| 581 | size = sprintf(buffer+len, "%8lu %-16s ", |
| 582 | now-en->lastuse, tbuf); |
| 583 | |
| 584 | read_lock(&en->set.lock); |
| 585 | for (d=en->set.list; d!=NULL; d=d->next) { |
| 586 | size += sprintf(buffer+len+size, |
| 587 | "%u.%u.%u.%u ", |
| 588 | NIPQUAD(d->dest->addr)); |
| 589 | } |
| 590 | read_unlock(&en->set.lock); |
| 591 | size += sprintf(buffer+len+size, "\n"); |
| 592 | len += size; |
| 593 | pos += size; |
| 594 | if (pos <= offset) |
| 595 | len=0; |
| 596 | if (pos >= offset+length) { |
| 597 | read_unlock_bh(&tbl->lock); |
| 598 | goto done; |
| 599 | } |
| 600 | } |
| 601 | read_unlock_bh(&tbl->lock); |
| 602 | } |
| 603 | |
| 604 | done: |
| 605 | begin = len - (pos - offset); |
| 606 | *start = buffer + begin; |
| 607 | len -= begin; |
| 608 | if(len>length) |
| 609 | len = length; |
| 610 | return len; |
| 611 | } |
| 612 | #endif |
| 613 | |
| 614 | |
| 615 | static int ip_vs_lblcr_init_svc(struct ip_vs_service *svc) |
| 616 | { |
| 617 | int i; |
| 618 | struct ip_vs_lblcr_table *tbl; |
| 619 | |
| 620 | /* |
| 621 | * Allocate the ip_vs_lblcr_table for this service |
| 622 | */ |
| 623 | tbl = kmalloc(sizeof(struct ip_vs_lblcr_table), GFP_ATOMIC); |
| 624 | if (tbl == NULL) { |
| 625 | IP_VS_ERR("ip_vs_lblcr_init_svc(): no memory\n"); |
| 626 | return -ENOMEM; |
| 627 | } |
| 628 | svc->sched_data = tbl; |
| 629 | IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) allocated for " |
| 630 | "current service\n", |
| 631 | sizeof(struct ip_vs_lblcr_table)); |
| 632 | |
| 633 | /* |
| 634 | * Initialize the hash buckets |
| 635 | */ |
| 636 | for (i=0; i<IP_VS_LBLCR_TAB_SIZE; i++) { |
| 637 | INIT_LIST_HEAD(&tbl->bucket[i]); |
| 638 | } |
| 639 | rwlock_init(&tbl->lock); |
| 640 | tbl->max_size = IP_VS_LBLCR_TAB_SIZE*16; |
| 641 | tbl->rover = 0; |
| 642 | tbl->counter = 1; |
| 643 | |
| 644 | /* |
| 645 | * Hook periodic timer for garbage collection |
| 646 | */ |
| 647 | init_timer(&tbl->periodic_timer); |
| 648 | tbl->periodic_timer.data = (unsigned long)tbl; |
| 649 | tbl->periodic_timer.function = ip_vs_lblcr_check_expire; |
| 650 | tbl->periodic_timer.expires = jiffies+CHECK_EXPIRE_INTERVAL; |
| 651 | add_timer(&tbl->periodic_timer); |
| 652 | |
| 653 | #ifdef CONFIG_IP_VS_LBLCR_DEBUG |
| 654 | lblcr_table_list = tbl; |
| 655 | #endif |
| 656 | return 0; |
| 657 | } |
| 658 | |
| 659 | |
| 660 | static int ip_vs_lblcr_done_svc(struct ip_vs_service *svc) |
| 661 | { |
| 662 | struct ip_vs_lblcr_table *tbl = svc->sched_data; |
| 663 | |
| 664 | /* remove periodic timer */ |
| 665 | del_timer_sync(&tbl->periodic_timer); |
| 666 | |
| 667 | /* got to clean up table entries here */ |
| 668 | ip_vs_lblcr_flush(tbl); |
| 669 | |
| 670 | /* release the table itself */ |
| 671 | kfree(svc->sched_data); |
| 672 | IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) released\n", |
| 673 | sizeof(struct ip_vs_lblcr_table)); |
| 674 | |
| 675 | return 0; |
| 676 | } |
| 677 | |
| 678 | |
| 679 | static int ip_vs_lblcr_update_svc(struct ip_vs_service *svc) |
| 680 | { |
| 681 | return 0; |
| 682 | } |
| 683 | |
| 684 | |
| 685 | static inline struct ip_vs_dest * |
| 686 | __ip_vs_wlc_schedule(struct ip_vs_service *svc, struct iphdr *iph) |
| 687 | { |
| 688 | struct ip_vs_dest *dest, *least; |
| 689 | int loh, doh; |
| 690 | |
| 691 | /* |
| 692 | * We think the overhead of processing active connections is fifty |
| 693 | * times higher than that of inactive connections in average. (This |
| 694 | * fifty times might not be accurate, we will change it later.) We |
| 695 | * use the following formula to estimate the overhead: |
| 696 | * dest->activeconns*50 + dest->inactconns |
| 697 | * and the load: |
| 698 | * (dest overhead) / dest->weight |
| 699 | * |
| 700 | * Remember -- no floats in kernel mode!!! |
| 701 | * The comparison of h1*w2 > h2*w1 is equivalent to that of |
| 702 | * h1/w1 > h2/w2 |
| 703 | * if every weight is larger than zero. |
| 704 | * |
| 705 | * The server with weight=0 is quiesced and will not receive any |
| 706 | * new connection. |
| 707 | */ |
| 708 | list_for_each_entry(dest, &svc->destinations, n_list) { |
| 709 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
| 710 | continue; |
| 711 | |
| 712 | if (atomic_read(&dest->weight) > 0) { |
| 713 | least = dest; |
| 714 | loh = atomic_read(&least->activeconns) * 50 |
| 715 | + atomic_read(&least->inactconns); |
| 716 | goto nextstage; |
| 717 | } |
| 718 | } |
| 719 | return NULL; |
| 720 | |
| 721 | /* |
| 722 | * Find the destination with the least load. |
| 723 | */ |
| 724 | nextstage: |
| 725 | list_for_each_entry_continue(dest, &svc->destinations, n_list) { |
| 726 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
| 727 | continue; |
| 728 | |
| 729 | doh = atomic_read(&dest->activeconns) * 50 |
| 730 | + atomic_read(&dest->inactconns); |
| 731 | if (loh * atomic_read(&dest->weight) > |
| 732 | doh * atomic_read(&least->weight)) { |
| 733 | least = dest; |
| 734 | loh = doh; |
| 735 | } |
| 736 | } |
| 737 | |
| 738 | IP_VS_DBG(6, "LBLCR: server %d.%d.%d.%d:%d " |
| 739 | "activeconns %d refcnt %d weight %d overhead %d\n", |
| 740 | NIPQUAD(least->addr), ntohs(least->port), |
| 741 | atomic_read(&least->activeconns), |
| 742 | atomic_read(&least->refcnt), |
| 743 | atomic_read(&least->weight), loh); |
| 744 | |
| 745 | return least; |
| 746 | } |
| 747 | |
| 748 | |
| 749 | /* |
| 750 | * If this destination server is overloaded and there is a less loaded |
| 751 | * server, then return true. |
| 752 | */ |
| 753 | static inline int |
| 754 | is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc) |
| 755 | { |
| 756 | if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) { |
| 757 | struct ip_vs_dest *d; |
| 758 | |
| 759 | list_for_each_entry(d, &svc->destinations, n_list) { |
| 760 | if (atomic_read(&d->activeconns)*2 |
| 761 | < atomic_read(&d->weight)) { |
| 762 | return 1; |
| 763 | } |
| 764 | } |
| 765 | } |
| 766 | return 0; |
| 767 | } |
| 768 | |
| 769 | |
| 770 | /* |
| 771 | * Locality-Based (weighted) Least-Connection scheduling |
| 772 | */ |
| 773 | static struct ip_vs_dest * |
| 774 | ip_vs_lblcr_schedule(struct ip_vs_service *svc, const struct sk_buff *skb) |
| 775 | { |
| 776 | struct ip_vs_dest *dest; |
| 777 | struct ip_vs_lblcr_table *tbl; |
| 778 | struct ip_vs_lblcr_entry *en; |
Arnaldo Carvalho de Melo | eddc9ec | 2007-04-20 22:47:35 -0700 | [diff] [blame] | 779 | struct iphdr *iph = ip_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 780 | |
| 781 | IP_VS_DBG(6, "ip_vs_lblcr_schedule(): Scheduling...\n"); |
| 782 | |
| 783 | tbl = (struct ip_vs_lblcr_table *)svc->sched_data; |
| 784 | en = ip_vs_lblcr_get(tbl, iph->daddr); |
| 785 | if (en == NULL) { |
| 786 | dest = __ip_vs_wlc_schedule(svc, iph); |
| 787 | if (dest == NULL) { |
| 788 | IP_VS_DBG(1, "no destination available\n"); |
| 789 | return NULL; |
| 790 | } |
| 791 | en = ip_vs_lblcr_new(iph->daddr); |
| 792 | if (en == NULL) { |
| 793 | return NULL; |
| 794 | } |
| 795 | ip_vs_dest_set_insert(&en->set, dest); |
| 796 | ip_vs_lblcr_hash(tbl, en); |
| 797 | } else { |
| 798 | dest = ip_vs_dest_set_min(&en->set); |
| 799 | if (!dest || is_overloaded(dest, svc)) { |
| 800 | dest = __ip_vs_wlc_schedule(svc, iph); |
| 801 | if (dest == NULL) { |
| 802 | IP_VS_DBG(1, "no destination available\n"); |
| 803 | return NULL; |
| 804 | } |
| 805 | ip_vs_dest_set_insert(&en->set, dest); |
| 806 | } |
| 807 | if (atomic_read(&en->set.size) > 1 && |
| 808 | jiffies-en->set.lastmod > sysctl_ip_vs_lblcr_expiration) { |
| 809 | struct ip_vs_dest *m; |
| 810 | m = ip_vs_dest_set_max(&en->set); |
| 811 | if (m) |
| 812 | ip_vs_dest_set_erase(&en->set, m); |
| 813 | } |
| 814 | } |
| 815 | en->lastuse = jiffies; |
| 816 | |
| 817 | IP_VS_DBG(6, "LBLCR: destination IP address %u.%u.%u.%u " |
| 818 | "--> server %u.%u.%u.%u:%d\n", |
| 819 | NIPQUAD(en->addr), |
| 820 | NIPQUAD(dest->addr), |
| 821 | ntohs(dest->port)); |
| 822 | |
| 823 | return dest; |
| 824 | } |
| 825 | |
| 826 | |
| 827 | /* |
| 828 | * IPVS LBLCR Scheduler structure |
| 829 | */ |
| 830 | static struct ip_vs_scheduler ip_vs_lblcr_scheduler = |
| 831 | { |
| 832 | .name = "lblcr", |
| 833 | .refcnt = ATOMIC_INIT(0), |
| 834 | .module = THIS_MODULE, |
| 835 | .init_service = ip_vs_lblcr_init_svc, |
| 836 | .done_service = ip_vs_lblcr_done_svc, |
| 837 | .update_service = ip_vs_lblcr_update_svc, |
| 838 | .schedule = ip_vs_lblcr_schedule, |
| 839 | }; |
| 840 | |
| 841 | |
| 842 | static int __init ip_vs_lblcr_init(void) |
| 843 | { |
| 844 | INIT_LIST_HEAD(&ip_vs_lblcr_scheduler.n_list); |
Eric W. Biederman | 0b4d414 | 2007-02-14 00:34:09 -0800 | [diff] [blame] | 845 | sysctl_header = register_sysctl_table(lblcr_root_table); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 846 | #ifdef CONFIG_IP_VS_LBLCR_DEBUG |
Eric W. Biederman | 457c4cb | 2007-09-12 12:01:34 +0200 | [diff] [blame^] | 847 | proc_net_create(&init_net, "ip_vs_lblcr", 0, ip_vs_lblcr_getinfo); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 848 | #endif |
| 849 | return register_ip_vs_scheduler(&ip_vs_lblcr_scheduler); |
| 850 | } |
| 851 | |
| 852 | |
| 853 | static void __exit ip_vs_lblcr_cleanup(void) |
| 854 | { |
| 855 | #ifdef CONFIG_IP_VS_LBLCR_DEBUG |
Eric W. Biederman | 457c4cb | 2007-09-12 12:01:34 +0200 | [diff] [blame^] | 856 | proc_net_remove(&init_net, "ip_vs_lblcr"); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 857 | #endif |
| 858 | unregister_sysctl_table(sysctl_header); |
| 859 | unregister_ip_vs_scheduler(&ip_vs_lblcr_scheduler); |
| 860 | } |
| 861 | |
| 862 | |
| 863 | module_init(ip_vs_lblcr_init); |
| 864 | module_exit(ip_vs_lblcr_cleanup); |
| 865 | MODULE_LICENSE("GPL"); |