2 * IPVS: Locality-Based Least-Connection with Replication scheduler
4 * Authors: Wensong Zhang <wensong@gnuchina.org>
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.
12 * Julian Anastasov : Added the missing (dest->weight>0)
13 * condition in the ip_vs_dest_set_max.
18 * The lblc/r algorithm is as follows (pseudo code):
20 * if serverSet[dest_ip] is null then
21 * n, serverSet[dest_ip] <- {weighted least-conn node};
23 * n <- {least-conn (alive) node in serverSet[dest_ip]};
25 * (n.conns>n.weight AND
26 * there is a node m with m.conns<m.weight/2) then
27 * n <- {weighted least-conn node};
28 * add n to serverSet[dest_ip];
29 * if |serverSet[dest_ip]| > 1 AND
30 * now - serverSet[dest_ip].lastMod > T then
31 * m <- {most conn node in serverSet[dest_ip]};
32 * remove m from serverSet[dest_ip];
33 * if serverSet[dest_ip] changed then
34 * serverSet[dest_ip].lastMod <- now;
40 #define KMSG_COMPONENT "IPVS"
41 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
44 #include <linux/module.h>
45 #include <linux/kernel.h>
46 #include <linux/skbuff.h>
47 #include <linux/jiffies.h>
48 #include <linux/list.h>
49 #include <linux/slab.h>
53 #include <linux/sysctl.h>
54 #include <net/net_namespace.h>
56 #include <net/ip_vs.h>
60 * It is for garbage collection of stale IPVS lblcr entries,
61 * when the table is full.
63 #define CHECK_EXPIRE_INTERVAL (60*HZ)
64 #define ENTRY_TIMEOUT (6*60*HZ)
67 * It is for full expiration check.
68 * When there is no partial expiration check (garbage collection)
69 * in a half hour, do a full expiration check to collect stale
70 * entries that haven't been touched for a day.
72 #define COUNT_FOR_FULL_EXPIRATION 30
75 * for IPVS lblcr entry hash table
77 #ifndef CONFIG_IP_VS_LBLCR_TAB_BITS
78 #define CONFIG_IP_VS_LBLCR_TAB_BITS 10
80 #define IP_VS_LBLCR_TAB_BITS CONFIG_IP_VS_LBLCR_TAB_BITS
81 #define IP_VS_LBLCR_TAB_SIZE (1 << IP_VS_LBLCR_TAB_BITS)
82 #define IP_VS_LBLCR_TAB_MASK (IP_VS_LBLCR_TAB_SIZE - 1)
86 * IPVS destination set structure and operations
88 struct ip_vs_dest_set_elem {
89 struct list_head list; /* list link */
90 struct ip_vs_dest *dest; /* destination server */
93 struct ip_vs_dest_set {
94 atomic_t size; /* set size */
95 unsigned long lastmod; /* last modified time */
96 struct list_head list; /* destination list */
97 rwlock_t lock; /* lock for this list */
101 static struct ip_vs_dest_set_elem *
102 ip_vs_dest_set_insert(struct ip_vs_dest_set *set, struct ip_vs_dest *dest)
104 struct ip_vs_dest_set_elem *e;
106 list_for_each_entry(e, &set->list, list) {
108 /* already existed */
112 e = kmalloc(sizeof(*e), GFP_ATOMIC);
114 pr_err("%s(): no memory\n", __func__);
118 atomic_inc(&dest->refcnt);
121 list_add(&e->list, &set->list);
122 atomic_inc(&set->size);
124 set->lastmod = jiffies;
129 ip_vs_dest_set_erase(struct ip_vs_dest_set *set, struct ip_vs_dest *dest)
131 struct ip_vs_dest_set_elem *e;
133 list_for_each_entry(e, &set->list, list) {
134 if (e->dest == dest) {
136 atomic_dec(&set->size);
137 set->lastmod = jiffies;
138 atomic_dec(&e->dest->refcnt);
146 static void ip_vs_dest_set_eraseall(struct ip_vs_dest_set *set)
148 struct ip_vs_dest_set_elem *e, *ep;
150 write_lock(&set->lock);
151 list_for_each_entry_safe(e, ep, &set->list, list) {
153 * We don't kfree dest because it is refered either
154 * by its service or by the trash dest list.
156 atomic_dec(&e->dest->refcnt);
160 write_unlock(&set->lock);
163 /* get weighted least-connection node in the destination set */
164 static inline struct ip_vs_dest *ip_vs_dest_set_min(struct ip_vs_dest_set *set)
166 register struct ip_vs_dest_set_elem *e;
167 struct ip_vs_dest *dest, *least;
173 /* select the first destination server, whose weight > 0 */
174 list_for_each_entry(e, &set->list, list) {
176 if (least->flags & IP_VS_DEST_F_OVERLOAD)
179 if ((atomic_read(&least->weight) > 0)
180 && (least->flags & IP_VS_DEST_F_AVAILABLE)) {
181 loh = ip_vs_dest_conn_overhead(least);
187 /* find the destination with the weighted least load */
189 list_for_each_entry(e, &set->list, list) {
191 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
194 doh = ip_vs_dest_conn_overhead(dest);
195 if ((loh * atomic_read(&dest->weight) >
196 doh * atomic_read(&least->weight))
197 && (dest->flags & IP_VS_DEST_F_AVAILABLE)) {
203 IP_VS_DBG_BUF(6, "%s(): server %s:%d "
204 "activeconns %d refcnt %d weight %d overhead %d\n",
206 IP_VS_DBG_ADDR(least->af, &least->addr),
208 atomic_read(&least->activeconns),
209 atomic_read(&least->refcnt),
210 atomic_read(&least->weight), loh);
215 /* get weighted most-connection node in the destination set */
216 static inline struct ip_vs_dest *ip_vs_dest_set_max(struct ip_vs_dest_set *set)
218 register struct ip_vs_dest_set_elem *e;
219 struct ip_vs_dest *dest, *most;
225 /* select the first destination server, whose weight > 0 */
226 list_for_each_entry(e, &set->list, list) {
228 if (atomic_read(&most->weight) > 0) {
229 moh = ip_vs_dest_conn_overhead(most);
235 /* find the destination with the weighted most load */
237 list_for_each_entry(e, &set->list, list) {
239 doh = ip_vs_dest_conn_overhead(dest);
240 /* moh/mw < doh/dw ==> moh*dw < doh*mw, where mw,dw>0 */
241 if ((moh * atomic_read(&dest->weight) <
242 doh * atomic_read(&most->weight))
243 && (atomic_read(&dest->weight) > 0)) {
249 IP_VS_DBG_BUF(6, "%s(): server %s:%d "
250 "activeconns %d refcnt %d weight %d overhead %d\n",
252 IP_VS_DBG_ADDR(most->af, &most->addr), ntohs(most->port),
253 atomic_read(&most->activeconns),
254 atomic_read(&most->refcnt),
255 atomic_read(&most->weight), moh);
261 * IPVS lblcr entry represents an association between destination
262 * IP address and its destination server set
264 struct ip_vs_lblcr_entry {
265 struct list_head list;
266 int af; /* address family */
267 union nf_inet_addr addr; /* destination IP address */
268 struct ip_vs_dest_set set; /* destination server set */
269 unsigned long lastuse; /* last used time */
274 * IPVS lblcr hash table
276 struct ip_vs_lblcr_table {
277 struct list_head bucket[IP_VS_LBLCR_TAB_SIZE]; /* hash bucket */
278 atomic_t entries; /* number of entries */
279 int max_size; /* maximum size of entries */
280 struct timer_list periodic_timer; /* collect stale entries */
281 int rover; /* rover for expire check */
282 int counter; /* counter for no expire */
287 * IPVS LBLCR sysctl table
290 static ctl_table vs_vars_table[] = {
292 .procname = "lblcr_expiration",
294 .maxlen = sizeof(int),
296 .proc_handler = proc_dointvec_jiffies,
301 static inline void ip_vs_lblcr_free(struct ip_vs_lblcr_entry *en)
304 ip_vs_dest_set_eraseall(&en->set);
310 * Returns hash value for IPVS LBLCR entry
312 static inline unsigned
313 ip_vs_lblcr_hashkey(int af, const union nf_inet_addr *addr)
315 __be32 addr_fold = addr->ip;
317 #ifdef CONFIG_IP_VS_IPV6
319 addr_fold = addr->ip6[0]^addr->ip6[1]^
320 addr->ip6[2]^addr->ip6[3];
322 return (ntohl(addr_fold)*2654435761UL) & IP_VS_LBLCR_TAB_MASK;
327 * Hash an entry in the ip_vs_lblcr_table.
328 * returns bool success.
331 ip_vs_lblcr_hash(struct ip_vs_lblcr_table *tbl, struct ip_vs_lblcr_entry *en)
333 unsigned hash = ip_vs_lblcr_hashkey(en->af, &en->addr);
335 list_add(&en->list, &tbl->bucket[hash]);
336 atomic_inc(&tbl->entries);
341 * Get ip_vs_lblcr_entry associated with supplied parameters. Called under
344 static inline struct ip_vs_lblcr_entry *
345 ip_vs_lblcr_get(int af, struct ip_vs_lblcr_table *tbl,
346 const union nf_inet_addr *addr)
348 unsigned hash = ip_vs_lblcr_hashkey(af, addr);
349 struct ip_vs_lblcr_entry *en;
351 list_for_each_entry(en, &tbl->bucket[hash], list)
352 if (ip_vs_addr_equal(af, &en->addr, addr))
360 * Create or update an ip_vs_lblcr_entry, which is a mapping of a destination
361 * IP address to a server. Called under write lock.
363 static inline struct ip_vs_lblcr_entry *
364 ip_vs_lblcr_new(struct ip_vs_lblcr_table *tbl, const union nf_inet_addr *daddr,
365 struct ip_vs_dest *dest)
367 struct ip_vs_lblcr_entry *en;
369 en = ip_vs_lblcr_get(dest->af, tbl, daddr);
371 en = kmalloc(sizeof(*en), GFP_ATOMIC);
373 pr_err("%s(): no memory\n", __func__);
378 ip_vs_addr_copy(dest->af, &en->addr, daddr);
379 en->lastuse = jiffies;
381 /* initialize its dest set */
382 atomic_set(&(en->set.size), 0);
383 INIT_LIST_HEAD(&en->set.list);
384 rwlock_init(&en->set.lock);
386 ip_vs_lblcr_hash(tbl, en);
389 write_lock(&en->set.lock);
390 ip_vs_dest_set_insert(&en->set, dest);
391 write_unlock(&en->set.lock);
398 * Flush all the entries of the specified table.
400 static void ip_vs_lblcr_flush(struct ip_vs_lblcr_table *tbl)
403 struct ip_vs_lblcr_entry *en, *nxt;
405 /* No locking required, only called during cleanup. */
406 for (i=0; i<IP_VS_LBLCR_TAB_SIZE; i++) {
407 list_for_each_entry_safe(en, nxt, &tbl->bucket[i], list) {
408 ip_vs_lblcr_free(en);
414 static inline void ip_vs_lblcr_full_check(struct ip_vs_service *svc)
416 struct ip_vs_lblcr_table *tbl = svc->sched_data;
417 unsigned long now = jiffies;
419 struct ip_vs_lblcr_entry *en, *nxt;
420 struct netns_ipvs *ipvs = net_ipvs(svc->net);
422 for (i=0, j=tbl->rover; i<IP_VS_LBLCR_TAB_SIZE; i++) {
423 j = (j + 1) & IP_VS_LBLCR_TAB_MASK;
425 write_lock(&svc->sched_lock);
426 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
427 if (time_after(en->lastuse
428 + ipvs->sysctl_lblcr_expiration, now))
431 ip_vs_lblcr_free(en);
432 atomic_dec(&tbl->entries);
434 write_unlock(&svc->sched_lock);
441 * Periodical timer handler for IPVS lblcr table
442 * It is used to collect stale entries when the number of entries
443 * exceeds the maximum size of the table.
445 * Fixme: we probably need more complicated algorithm to collect
446 * entries that have not been used for a long time even
447 * if the number of entries doesn't exceed the maximum size
449 * The full expiration check is for this purpose now.
451 static void ip_vs_lblcr_check_expire(unsigned long data)
453 struct ip_vs_service *svc = (struct ip_vs_service *) data;
454 struct ip_vs_lblcr_table *tbl = svc->sched_data;
455 unsigned long now = jiffies;
458 struct ip_vs_lblcr_entry *en, *nxt;
460 if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) {
461 /* do full expiration check */
462 ip_vs_lblcr_full_check(svc);
467 if (atomic_read(&tbl->entries) <= tbl->max_size) {
472 goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3;
473 if (goal > tbl->max_size/2)
474 goal = tbl->max_size/2;
476 for (i=0, j=tbl->rover; i<IP_VS_LBLCR_TAB_SIZE; i++) {
477 j = (j + 1) & IP_VS_LBLCR_TAB_MASK;
479 write_lock(&svc->sched_lock);
480 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
481 if (time_before(now, en->lastuse+ENTRY_TIMEOUT))
484 ip_vs_lblcr_free(en);
485 atomic_dec(&tbl->entries);
488 write_unlock(&svc->sched_lock);
495 mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL);
498 static int ip_vs_lblcr_init_svc(struct ip_vs_service *svc)
501 struct ip_vs_lblcr_table *tbl;
504 * Allocate the ip_vs_lblcr_table for this service
506 tbl = kmalloc(sizeof(*tbl), GFP_ATOMIC);
508 pr_err("%s(): no memory\n", __func__);
511 svc->sched_data = tbl;
512 IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) allocated for "
513 "current service\n", sizeof(*tbl));
516 * Initialize the hash buckets
518 for (i=0; i<IP_VS_LBLCR_TAB_SIZE; i++) {
519 INIT_LIST_HEAD(&tbl->bucket[i]);
521 tbl->max_size = IP_VS_LBLCR_TAB_SIZE*16;
526 * Hook periodic timer for garbage collection
528 setup_timer(&tbl->periodic_timer, ip_vs_lblcr_check_expire,
530 mod_timer(&tbl->periodic_timer, jiffies + CHECK_EXPIRE_INTERVAL);
536 static int ip_vs_lblcr_done_svc(struct ip_vs_service *svc)
538 struct ip_vs_lblcr_table *tbl = svc->sched_data;
540 /* remove periodic timer */
541 del_timer_sync(&tbl->periodic_timer);
543 /* got to clean up table entries here */
544 ip_vs_lblcr_flush(tbl);
546 /* release the table itself */
548 IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) released\n",
555 static inline struct ip_vs_dest *
556 __ip_vs_lblcr_schedule(struct ip_vs_service *svc)
558 struct ip_vs_dest *dest, *least;
562 * We use the following formula to estimate the load:
563 * (dest overhead) / dest->weight
565 * Remember -- no floats in kernel mode!!!
566 * The comparison of h1*w2 > h2*w1 is equivalent to that of
568 * if every weight is larger than zero.
570 * The server with weight=0 is quiesced and will not receive any
573 list_for_each_entry(dest, &svc->destinations, n_list) {
574 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
577 if (atomic_read(&dest->weight) > 0) {
579 loh = ip_vs_dest_conn_overhead(least);
586 * Find the destination with the least load.
589 list_for_each_entry_continue(dest, &svc->destinations, n_list) {
590 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
593 doh = ip_vs_dest_conn_overhead(dest);
594 if (loh * atomic_read(&dest->weight) >
595 doh * atomic_read(&least->weight)) {
601 IP_VS_DBG_BUF(6, "LBLCR: server %s:%d "
602 "activeconns %d refcnt %d weight %d overhead %d\n",
603 IP_VS_DBG_ADDR(least->af, &least->addr),
605 atomic_read(&least->activeconns),
606 atomic_read(&least->refcnt),
607 atomic_read(&least->weight), loh);
614 * If this destination server is overloaded and there is a less loaded
615 * server, then return true.
618 is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc)
620 if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) {
621 struct ip_vs_dest *d;
623 list_for_each_entry(d, &svc->destinations, n_list) {
624 if (atomic_read(&d->activeconns)*2
625 < atomic_read(&d->weight)) {
635 * Locality-Based (weighted) Least-Connection scheduling
637 static struct ip_vs_dest *
638 ip_vs_lblcr_schedule(struct ip_vs_service *svc, const struct sk_buff *skb)
640 struct ip_vs_lblcr_table *tbl = svc->sched_data;
641 struct ip_vs_iphdr iph;
642 struct ip_vs_dest *dest = NULL;
643 struct ip_vs_lblcr_entry *en;
645 ip_vs_fill_iphdr(svc->af, skb_network_header(skb), &iph);
647 IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
649 /* First look in our cache */
650 read_lock(&svc->sched_lock);
651 en = ip_vs_lblcr_get(svc->af, tbl, &iph.daddr);
653 struct netns_ipvs *ipvs = net_ipvs(svc->net);
654 /* We only hold a read lock, but this is atomic */
655 en->lastuse = jiffies;
657 /* Get the least loaded destination */
658 read_lock(&en->set.lock);
659 dest = ip_vs_dest_set_min(&en->set);
660 read_unlock(&en->set.lock);
662 /* More than one destination + enough time passed by, cleanup */
663 if (atomic_read(&en->set.size) > 1 &&
664 time_after(jiffies, en->set.lastmod +
665 ipvs->sysctl_lblcr_expiration)) {
666 struct ip_vs_dest *m;
668 write_lock(&en->set.lock);
669 m = ip_vs_dest_set_max(&en->set);
671 ip_vs_dest_set_erase(&en->set, m);
672 write_unlock(&en->set.lock);
675 /* If the destination is not overloaded, use it */
676 if (dest && !is_overloaded(dest, svc)) {
677 read_unlock(&svc->sched_lock);
681 /* The cache entry is invalid, time to schedule */
682 dest = __ip_vs_lblcr_schedule(svc);
684 ip_vs_scheduler_err(svc, "no destination available");
685 read_unlock(&svc->sched_lock);
689 /* Update our cache entry */
690 write_lock(&en->set.lock);
691 ip_vs_dest_set_insert(&en->set, dest);
692 write_unlock(&en->set.lock);
694 read_unlock(&svc->sched_lock);
699 /* No cache entry, time to schedule */
700 dest = __ip_vs_lblcr_schedule(svc);
702 IP_VS_DBG(1, "no destination available\n");
706 /* If we fail to create a cache entry, we'll just use the valid dest */
707 write_lock(&svc->sched_lock);
708 ip_vs_lblcr_new(tbl, &iph.daddr, dest);
709 write_unlock(&svc->sched_lock);
712 IP_VS_DBG_BUF(6, "LBLCR: destination IP address %s --> server %s:%d\n",
713 IP_VS_DBG_ADDR(svc->af, &iph.daddr),
714 IP_VS_DBG_ADDR(svc->af, &dest->addr), ntohs(dest->port));
721 * IPVS LBLCR Scheduler structure
723 static struct ip_vs_scheduler ip_vs_lblcr_scheduler =
726 .refcnt = ATOMIC_INIT(0),
727 .module = THIS_MODULE,
728 .n_list = LIST_HEAD_INIT(ip_vs_lblcr_scheduler.n_list),
729 .init_service = ip_vs_lblcr_init_svc,
730 .done_service = ip_vs_lblcr_done_svc,
731 .schedule = ip_vs_lblcr_schedule,
737 static int __net_init __ip_vs_lblcr_init(struct net *net)
739 struct netns_ipvs *ipvs = net_ipvs(net);
741 if (!net_eq(net, &init_net)) {
742 ipvs->lblcr_ctl_table = kmemdup(vs_vars_table,
743 sizeof(vs_vars_table),
745 if (ipvs->lblcr_ctl_table == NULL)
748 ipvs->lblcr_ctl_table = vs_vars_table;
749 ipvs->sysctl_lblcr_expiration = 24*60*60*HZ;
750 ipvs->lblcr_ctl_table[0].data = &ipvs->sysctl_lblcr_expiration;
753 ipvs->lblcr_ctl_header =
754 register_net_sysctl_table(net, net_vs_ctl_path,
755 ipvs->lblcr_ctl_table);
756 if (!ipvs->lblcr_ctl_header) {
757 if (!net_eq(net, &init_net))
758 kfree(ipvs->lblcr_ctl_table);
766 static void __net_exit __ip_vs_lblcr_exit(struct net *net)
768 struct netns_ipvs *ipvs = net_ipvs(net);
771 unregister_net_sysctl_table(ipvs->lblcr_ctl_header);
774 if (!net_eq(net, &init_net))
775 kfree(ipvs->lblcr_ctl_table);
778 static struct pernet_operations ip_vs_lblcr_ops = {
779 .init = __ip_vs_lblcr_init,
780 .exit = __ip_vs_lblcr_exit,
783 static int __init ip_vs_lblcr_init(void)
787 ret = register_pernet_subsys(&ip_vs_lblcr_ops);
791 ret = register_ip_vs_scheduler(&ip_vs_lblcr_scheduler);
793 unregister_pernet_subsys(&ip_vs_lblcr_ops);
797 static void __exit ip_vs_lblcr_cleanup(void)
799 unregister_ip_vs_scheduler(&ip_vs_lblcr_scheduler);
800 unregister_pernet_subsys(&ip_vs_lblcr_ops);
804 module_init(ip_vs_lblcr_init);
805 module_exit(ip_vs_lblcr_cleanup);
806 MODULE_LICENSE("GPL");