1 /* 2 * IPVS: Locality-Based Least-Connection 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 * Martin Hamilton : fixed the terrible locking bugs 13 * *lock(tbl->lock) ==> *lock(&tbl->lock) 14 * Wensong Zhang : fixed the uninitialized tbl->lock bug 15 * Wensong Zhang : added doing full expiration check to 16 * collect stale entries of 24+ hours when 17 * no partial expire check in a half hour 18 * Julian Anastasov : replaced del_timer call with del_timer_sync 19 * to avoid the possible race between timer 20 * handler and del_timer thread in SMP 21 * 22 */ 23 24 /* 25 * The lblc algorithm is as follows (pseudo code): 26 * 27 * if cachenode[dest_ip] is null then 28 * n, cachenode[dest_ip] <- {weighted least-conn node}; 29 * else 30 * n <- cachenode[dest_ip]; 31 * if (n is dead) OR 32 * (n.conns>n.weight AND 33 * there is a node m with m.conns<m.weight/2) then 34 * n, cachenode[dest_ip] <- {weighted least-conn node}; 35 * 36 * return n; 37 * 38 * Thanks must go to Wenzhuo Zhang for talking WCCP to me and pushing 39 * me to write this module. 40 */ 41 42 #define KMSG_COMPONENT "IPVS" 43 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt 44 45 #include <linux/ip.h> 46 #include <linux/slab.h> 47 #include <linux/module.h> 48 #include <linux/kernel.h> 49 #include <linux/skbuff.h> 50 #include <linux/jiffies.h> 51 52 /* for sysctl */ 53 #include <linux/fs.h> 54 #include <linux/sysctl.h> 55 56 #include <net/ip_vs.h> 57 58 59 /* 60 * It is for garbage collection of stale IPVS lblc entries, 61 * when the table is full. 62 */ 63 #define CHECK_EXPIRE_INTERVAL (60*HZ) 64 #define ENTRY_TIMEOUT (6*60*HZ) 65 66 #define DEFAULT_EXPIRATION (24*60*60*HZ) 67 68 /* 69 * It is for full expiration check. 70 * When there is no partial expiration check (garbage collection) 71 * in a half hour, do a full expiration check to collect stale 72 * entries that haven't been touched for a day. 73 */ 74 #define COUNT_FOR_FULL_EXPIRATION 30 75 76 77 /* 78 * for IPVS lblc entry hash table 79 */ 80 #ifndef CONFIG_IP_VS_LBLC_TAB_BITS 81 #define CONFIG_IP_VS_LBLC_TAB_BITS 10 82 #endif 83 #define IP_VS_LBLC_TAB_BITS CONFIG_IP_VS_LBLC_TAB_BITS 84 #define IP_VS_LBLC_TAB_SIZE (1 << IP_VS_LBLC_TAB_BITS) 85 #define IP_VS_LBLC_TAB_MASK (IP_VS_LBLC_TAB_SIZE - 1) 86 87 88 /* 89 * IPVS lblc entry represents an association between destination 90 * IP address and its destination server 91 */ 92 struct ip_vs_lblc_entry { 93 struct hlist_node list; 94 int af; /* address family */ 95 union nf_inet_addr addr; /* destination IP address */ 96 struct ip_vs_dest __rcu *dest; /* real server (cache) */ 97 unsigned long lastuse; /* last used time */ 98 struct rcu_head rcu_head; 99 }; 100 101 102 /* 103 * IPVS lblc hash table 104 */ 105 struct ip_vs_lblc_table { 106 struct rcu_head rcu_head; 107 struct hlist_head bucket[IP_VS_LBLC_TAB_SIZE]; /* hash bucket */ 108 struct timer_list periodic_timer; /* collect stale entries */ 109 atomic_t entries; /* number of entries */ 110 int max_size; /* maximum size of entries */ 111 int rover; /* rover for expire check */ 112 int counter; /* counter for no expire */ 113 bool dead; 114 }; 115 116 117 /* 118 * IPVS LBLC sysctl table 119 */ 120 #ifdef CONFIG_SYSCTL 121 static struct ctl_table vs_vars_table[] = { 122 { 123 .procname = "lblc_expiration", 124 .data = NULL, 125 .maxlen = sizeof(int), 126 .mode = 0644, 127 .proc_handler = proc_dointvec_jiffies, 128 }, 129 { } 130 }; 131 #endif 132 133 static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry *en) 134 { 135 struct ip_vs_dest *dest; 136 137 hlist_del_rcu(&en->list); 138 /* 139 * We don't kfree dest because it is referred either by its service 140 * or the trash dest list. 141 */ 142 dest = rcu_dereference_protected(en->dest, 1); 143 ip_vs_dest_put(dest); 144 kfree_rcu(en, rcu_head); 145 } 146 147 148 /* 149 * Returns hash value for IPVS LBLC entry 150 */ 151 static inline unsigned int 152 ip_vs_lblc_hashkey(int af, const union nf_inet_addr *addr) 153 { 154 __be32 addr_fold = addr->ip; 155 156 #ifdef CONFIG_IP_VS_IPV6 157 if (af == AF_INET6) 158 addr_fold = addr->ip6[0]^addr->ip6[1]^ 159 addr->ip6[2]^addr->ip6[3]; 160 #endif 161 return (ntohl(addr_fold)*2654435761UL) & IP_VS_LBLC_TAB_MASK; 162 } 163 164 165 /* 166 * Hash an entry in the ip_vs_lblc_table. 167 * returns bool success. 168 */ 169 static void 170 ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en) 171 { 172 unsigned int hash = ip_vs_lblc_hashkey(en->af, &en->addr); 173 174 hlist_add_head_rcu(&en->list, &tbl->bucket[hash]); 175 atomic_inc(&tbl->entries); 176 } 177 178 179 /* Get ip_vs_lblc_entry associated with supplied parameters. */ 180 static inline struct ip_vs_lblc_entry * 181 ip_vs_lblc_get(int af, struct ip_vs_lblc_table *tbl, 182 const union nf_inet_addr *addr) 183 { 184 unsigned int hash = ip_vs_lblc_hashkey(af, addr); 185 struct ip_vs_lblc_entry *en; 186 187 hlist_for_each_entry_rcu(en, &tbl->bucket[hash], list) 188 if (ip_vs_addr_equal(af, &en->addr, addr)) 189 return en; 190 191 return NULL; 192 } 193 194 195 /* 196 * Create or update an ip_vs_lblc_entry, which is a mapping of a destination IP 197 * address to a server. Called under spin lock. 198 */ 199 static inline struct ip_vs_lblc_entry * 200 ip_vs_lblc_new(struct ip_vs_lblc_table *tbl, const union nf_inet_addr *daddr, 201 struct ip_vs_dest *dest) 202 { 203 struct ip_vs_lblc_entry *en; 204 205 en = ip_vs_lblc_get(dest->af, tbl, daddr); 206 if (!en) { 207 en = kmalloc(sizeof(*en), GFP_ATOMIC); 208 if (!en) 209 return NULL; 210 211 en->af = dest->af; 212 ip_vs_addr_copy(dest->af, &en->addr, daddr); 213 en->lastuse = jiffies; 214 215 ip_vs_dest_hold(dest); 216 RCU_INIT_POINTER(en->dest, dest); 217 218 ip_vs_lblc_hash(tbl, en); 219 } else { 220 struct ip_vs_dest *old_dest; 221 222 old_dest = rcu_dereference_protected(en->dest, 1); 223 if (old_dest != dest) { 224 ip_vs_dest_put(old_dest); 225 ip_vs_dest_hold(dest); 226 /* No ordering constraints for refcnt */ 227 RCU_INIT_POINTER(en->dest, dest); 228 } 229 } 230 231 return en; 232 } 233 234 235 /* 236 * Flush all the entries of the specified table. 237 */ 238 static void ip_vs_lblc_flush(struct ip_vs_service *svc) 239 { 240 struct ip_vs_lblc_table *tbl = svc->sched_data; 241 struct ip_vs_lblc_entry *en; 242 struct hlist_node *next; 243 int i; 244 245 spin_lock_bh(&svc->sched_lock); 246 tbl->dead = 1; 247 for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) { 248 hlist_for_each_entry_safe(en, next, &tbl->bucket[i], list) { 249 ip_vs_lblc_free(en); 250 atomic_dec(&tbl->entries); 251 } 252 } 253 spin_unlock_bh(&svc->sched_lock); 254 } 255 256 static int sysctl_lblc_expiration(struct ip_vs_service *svc) 257 { 258 #ifdef CONFIG_SYSCTL 259 struct netns_ipvs *ipvs = net_ipvs(svc->net); 260 return ipvs->sysctl_lblc_expiration; 261 #else 262 return DEFAULT_EXPIRATION; 263 #endif 264 } 265 266 static inline void ip_vs_lblc_full_check(struct ip_vs_service *svc) 267 { 268 struct ip_vs_lblc_table *tbl = svc->sched_data; 269 struct ip_vs_lblc_entry *en; 270 struct hlist_node *next; 271 unsigned long now = jiffies; 272 int i, j; 273 274 for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) { 275 j = (j + 1) & IP_VS_LBLC_TAB_MASK; 276 277 spin_lock(&svc->sched_lock); 278 hlist_for_each_entry_safe(en, next, &tbl->bucket[j], list) { 279 if (time_before(now, 280 en->lastuse + 281 sysctl_lblc_expiration(svc))) 282 continue; 283 284 ip_vs_lblc_free(en); 285 atomic_dec(&tbl->entries); 286 } 287 spin_unlock(&svc->sched_lock); 288 } 289 tbl->rover = j; 290 } 291 292 293 /* 294 * Periodical timer handler for IPVS lblc table 295 * It is used to collect stale entries when the number of entries 296 * exceeds the maximum size of the table. 297 * 298 * Fixme: we probably need more complicated algorithm to collect 299 * entries that have not been used for a long time even 300 * if the number of entries doesn't exceed the maximum size 301 * of the table. 302 * The full expiration check is for this purpose now. 303 */ 304 static void ip_vs_lblc_check_expire(unsigned long data) 305 { 306 struct ip_vs_service *svc = (struct ip_vs_service *) data; 307 struct ip_vs_lblc_table *tbl = svc->sched_data; 308 unsigned long now = jiffies; 309 int goal; 310 int i, j; 311 struct ip_vs_lblc_entry *en; 312 struct hlist_node *next; 313 314 if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) { 315 /* do full expiration check */ 316 ip_vs_lblc_full_check(svc); 317 tbl->counter = 1; 318 goto out; 319 } 320 321 if (atomic_read(&tbl->entries) <= tbl->max_size) { 322 tbl->counter++; 323 goto out; 324 } 325 326 goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3; 327 if (goal > tbl->max_size/2) 328 goal = tbl->max_size/2; 329 330 for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) { 331 j = (j + 1) & IP_VS_LBLC_TAB_MASK; 332 333 spin_lock(&svc->sched_lock); 334 hlist_for_each_entry_safe(en, next, &tbl->bucket[j], list) { 335 if (time_before(now, en->lastuse + ENTRY_TIMEOUT)) 336 continue; 337 338 ip_vs_lblc_free(en); 339 atomic_dec(&tbl->entries); 340 goal--; 341 } 342 spin_unlock(&svc->sched_lock); 343 if (goal <= 0) 344 break; 345 } 346 tbl->rover = j; 347 348 out: 349 mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL); 350 } 351 352 353 static int ip_vs_lblc_init_svc(struct ip_vs_service *svc) 354 { 355 int i; 356 struct ip_vs_lblc_table *tbl; 357 358 /* 359 * Allocate the ip_vs_lblc_table for this service 360 */ 361 tbl = kmalloc(sizeof(*tbl), GFP_KERNEL); 362 if (tbl == NULL) 363 return -ENOMEM; 364 365 svc->sched_data = tbl; 366 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for " 367 "current service\n", sizeof(*tbl)); 368 369 /* 370 * Initialize the hash buckets 371 */ 372 for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) { 373 INIT_HLIST_HEAD(&tbl->bucket[i]); 374 } 375 tbl->max_size = IP_VS_LBLC_TAB_SIZE*16; 376 tbl->rover = 0; 377 tbl->counter = 1; 378 tbl->dead = 0; 379 380 /* 381 * Hook periodic timer for garbage collection 382 */ 383 setup_timer(&tbl->periodic_timer, ip_vs_lblc_check_expire, 384 (unsigned long)svc); 385 mod_timer(&tbl->periodic_timer, jiffies + CHECK_EXPIRE_INTERVAL); 386 387 return 0; 388 } 389 390 391 static void ip_vs_lblc_done_svc(struct ip_vs_service *svc) 392 { 393 struct ip_vs_lblc_table *tbl = svc->sched_data; 394 395 /* remove periodic timer */ 396 del_timer_sync(&tbl->periodic_timer); 397 398 /* got to clean up table entries here */ 399 ip_vs_lblc_flush(svc); 400 401 /* release the table itself */ 402 kfree_rcu(tbl, rcu_head); 403 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n", 404 sizeof(*tbl)); 405 } 406 407 408 static inline struct ip_vs_dest * 409 __ip_vs_lblc_schedule(struct ip_vs_service *svc) 410 { 411 struct ip_vs_dest *dest, *least; 412 int loh, doh; 413 414 /* 415 * We use the following formula to estimate the load: 416 * (dest overhead) / dest->weight 417 * 418 * Remember -- no floats in kernel mode!!! 419 * The comparison of h1*w2 > h2*w1 is equivalent to that of 420 * h1/w1 > h2/w2 421 * if every weight is larger than zero. 422 * 423 * The server with weight=0 is quiesced and will not receive any 424 * new connection. 425 */ 426 list_for_each_entry_rcu(dest, &svc->destinations, n_list) { 427 if (dest->flags & IP_VS_DEST_F_OVERLOAD) 428 continue; 429 if (atomic_read(&dest->weight) > 0) { 430 least = dest; 431 loh = ip_vs_dest_conn_overhead(least); 432 goto nextstage; 433 } 434 } 435 return NULL; 436 437 /* 438 * Find the destination with the least load. 439 */ 440 nextstage: 441 list_for_each_entry_continue_rcu(dest, &svc->destinations, n_list) { 442 if (dest->flags & IP_VS_DEST_F_OVERLOAD) 443 continue; 444 445 doh = ip_vs_dest_conn_overhead(dest); 446 if (loh * atomic_read(&dest->weight) > 447 doh * atomic_read(&least->weight)) { 448 least = dest; 449 loh = doh; 450 } 451 } 452 453 IP_VS_DBG_BUF(6, "LBLC: server %s:%d " 454 "activeconns %d refcnt %d weight %d overhead %d\n", 455 IP_VS_DBG_ADDR(least->af, &least->addr), 456 ntohs(least->port), 457 atomic_read(&least->activeconns), 458 atomic_read(&least->refcnt), 459 atomic_read(&least->weight), loh); 460 461 return least; 462 } 463 464 465 /* 466 * If this destination server is overloaded and there is a less loaded 467 * server, then return true. 468 */ 469 static inline int 470 is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc) 471 { 472 if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) { 473 struct ip_vs_dest *d; 474 475 list_for_each_entry_rcu(d, &svc->destinations, n_list) { 476 if (atomic_read(&d->activeconns)*2 477 < atomic_read(&d->weight)) { 478 return 1; 479 } 480 } 481 } 482 return 0; 483 } 484 485 486 /* 487 * Locality-Based (weighted) Least-Connection scheduling 488 */ 489 static struct ip_vs_dest * 490 ip_vs_lblc_schedule(struct ip_vs_service *svc, const struct sk_buff *skb, 491 struct ip_vs_iphdr *iph) 492 { 493 struct ip_vs_lblc_table *tbl = svc->sched_data; 494 struct ip_vs_dest *dest = NULL; 495 struct ip_vs_lblc_entry *en; 496 497 IP_VS_DBG(6, "%s(): Scheduling...\n", __func__); 498 499 /* First look in our cache */ 500 en = ip_vs_lblc_get(svc->af, tbl, &iph->daddr); 501 if (en) { 502 /* We only hold a read lock, but this is atomic */ 503 en->lastuse = jiffies; 504 505 /* 506 * If the destination is not available, i.e. it's in the trash, 507 * we must ignore it, as it may be removed from under our feet, 508 * if someone drops our reference count. Our caller only makes 509 * sure that destinations, that are not in the trash, are not 510 * moved to the trash, while we are scheduling. But anyone can 511 * free up entries from the trash at any time. 512 */ 513 514 dest = rcu_dereference(en->dest); 515 if ((dest->flags & IP_VS_DEST_F_AVAILABLE) && 516 atomic_read(&dest->weight) > 0 && !is_overloaded(dest, svc)) 517 goto out; 518 } 519 520 /* No cache entry or it is invalid, time to schedule */ 521 dest = __ip_vs_lblc_schedule(svc); 522 if (!dest) { 523 ip_vs_scheduler_err(svc, "no destination available"); 524 return NULL; 525 } 526 527 /* If we fail to create a cache entry, we'll just use the valid dest */ 528 spin_lock_bh(&svc->sched_lock); 529 if (!tbl->dead) 530 ip_vs_lblc_new(tbl, &iph->daddr, dest); 531 spin_unlock_bh(&svc->sched_lock); 532 533 out: 534 IP_VS_DBG_BUF(6, "LBLC: destination IP address %s --> server %s:%d\n", 535 IP_VS_DBG_ADDR(svc->af, &iph->daddr), 536 IP_VS_DBG_ADDR(svc->af, &dest->addr), ntohs(dest->port)); 537 538 return dest; 539 } 540 541 542 /* 543 * IPVS LBLC Scheduler structure 544 */ 545 static struct ip_vs_scheduler ip_vs_lblc_scheduler = 546 { 547 .name = "lblc", 548 .refcnt = ATOMIC_INIT(0), 549 .module = THIS_MODULE, 550 .n_list = LIST_HEAD_INIT(ip_vs_lblc_scheduler.n_list), 551 .init_service = ip_vs_lblc_init_svc, 552 .done_service = ip_vs_lblc_done_svc, 553 .schedule = ip_vs_lblc_schedule, 554 }; 555 556 /* 557 * per netns init. 558 */ 559 #ifdef CONFIG_SYSCTL 560 static int __net_init __ip_vs_lblc_init(struct net *net) 561 { 562 struct netns_ipvs *ipvs = net_ipvs(net); 563 564 if (!ipvs) 565 return -ENOENT; 566 567 if (!net_eq(net, &init_net)) { 568 ipvs->lblc_ctl_table = kmemdup(vs_vars_table, 569 sizeof(vs_vars_table), 570 GFP_KERNEL); 571 if (ipvs->lblc_ctl_table == NULL) 572 return -ENOMEM; 573 574 /* Don't export sysctls to unprivileged users */ 575 if (net->user_ns != &init_user_ns) 576 ipvs->lblc_ctl_table[0].procname = NULL; 577 578 } else 579 ipvs->lblc_ctl_table = vs_vars_table; 580 ipvs->sysctl_lblc_expiration = DEFAULT_EXPIRATION; 581 ipvs->lblc_ctl_table[0].data = &ipvs->sysctl_lblc_expiration; 582 583 ipvs->lblc_ctl_header = 584 register_net_sysctl(net, "net/ipv4/vs", ipvs->lblc_ctl_table); 585 if (!ipvs->lblc_ctl_header) { 586 if (!net_eq(net, &init_net)) 587 kfree(ipvs->lblc_ctl_table); 588 return -ENOMEM; 589 } 590 591 return 0; 592 } 593 594 static void __net_exit __ip_vs_lblc_exit(struct net *net) 595 { 596 struct netns_ipvs *ipvs = net_ipvs(net); 597 598 unregister_net_sysctl_table(ipvs->lblc_ctl_header); 599 600 if (!net_eq(net, &init_net)) 601 kfree(ipvs->lblc_ctl_table); 602 } 603 604 #else 605 606 static int __net_init __ip_vs_lblc_init(struct net *net) { return 0; } 607 static void __net_exit __ip_vs_lblc_exit(struct net *net) { } 608 609 #endif 610 611 static struct pernet_operations ip_vs_lblc_ops = { 612 .init = __ip_vs_lblc_init, 613 .exit = __ip_vs_lblc_exit, 614 }; 615 616 static int __init ip_vs_lblc_init(void) 617 { 618 int ret; 619 620 ret = register_pernet_subsys(&ip_vs_lblc_ops); 621 if (ret) 622 return ret; 623 624 ret = register_ip_vs_scheduler(&ip_vs_lblc_scheduler); 625 if (ret) 626 unregister_pernet_subsys(&ip_vs_lblc_ops); 627 return ret; 628 } 629 630 static void __exit ip_vs_lblc_cleanup(void) 631 { 632 unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler); 633 unregister_pernet_subsys(&ip_vs_lblc_ops); 634 synchronize_rcu(); 635 } 636 637 638 module_init(ip_vs_lblc_init); 639 module_exit(ip_vs_lblc_cleanup); 640 MODULE_LICENSE("GPL"); 641