2 * IPVS: Locality-Based Least-Connection scheduling module
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 * 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
25 * The lblc algorithm is as follows (pseudo code):
27 * if cachenode[dest_ip] is null then
28 * n, cachenode[dest_ip] <- {weighted least-conn node};
30 * n <- cachenode[dest_ip];
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};
38 * Thanks must go to Wenzhuo Zhang for talking WCCP to me and pushing
39 * me to write this module.
42 #define KMSG_COMPONENT "IPVS"
43 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
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>
54 #include <linux/sysctl.h>
56 #include <net/ip_vs.h>
60 * It is for garbage collection of stale IPVS lblc entries,
61 * when the table is full.
63 #define CHECK_EXPIRE_INTERVAL (60*HZ)
64 #define ENTRY_TIMEOUT (6*60*HZ)
66 #define DEFAULT_EXPIRATION (24*60*60*HZ)
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.
74 #define COUNT_FOR_FULL_EXPIRATION 30
78 * for IPVS lblc entry hash table
80 #ifndef CONFIG_IP_VS_LBLC_TAB_BITS
81 #define CONFIG_IP_VS_LBLC_TAB_BITS 10
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)
89 * IPVS lblc entry represents an association between destination
90 * IP address and its destination server
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;
103 * IPVS lblc hash table
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 */
118 * IPVS LBLC sysctl table
121 static ctl_table vs_vars_table[] = {
123 .procname = "lblc_expiration",
125 .maxlen = sizeof(int),
127 .proc_handler = proc_dointvec_jiffies,
133 static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry *en)
135 struct ip_vs_dest *dest;
137 hlist_del_rcu(&en->list);
139 * We don't kfree dest because it is referred either by its service
140 * or the trash dest list.
142 dest = rcu_dereference_protected(en->dest, 1);
143 ip_vs_dest_put(dest);
144 kfree_rcu(en, rcu_head);
149 * Returns hash value for IPVS LBLC entry
151 static inline unsigned int
152 ip_vs_lblc_hashkey(int af, const union nf_inet_addr *addr)
154 __be32 addr_fold = addr->ip;
156 #ifdef CONFIG_IP_VS_IPV6
158 addr_fold = addr->ip6[0]^addr->ip6[1]^
159 addr->ip6[2]^addr->ip6[3];
161 return (ntohl(addr_fold)*2654435761UL) & IP_VS_LBLC_TAB_MASK;
166 * Hash an entry in the ip_vs_lblc_table.
167 * returns bool success.
170 ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en)
172 unsigned int hash = ip_vs_lblc_hashkey(en->af, &en->addr);
174 hlist_add_head_rcu(&en->list, &tbl->bucket[hash]);
175 atomic_inc(&tbl->entries);
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)
184 unsigned int hash = ip_vs_lblc_hashkey(af, addr);
185 struct ip_vs_lblc_entry *en;
187 hlist_for_each_entry_rcu(en, &tbl->bucket[hash], list)
188 if (ip_vs_addr_equal(af, &en->addr, addr))
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.
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)
203 struct ip_vs_lblc_entry *en;
205 en = ip_vs_lblc_get(dest->af, tbl, daddr);
207 en = kmalloc(sizeof(*en), GFP_ATOMIC);
212 ip_vs_addr_copy(dest->af, &en->addr, daddr);
213 en->lastuse = jiffies;
215 ip_vs_dest_hold(dest);
216 RCU_INIT_POINTER(en->dest, dest);
218 ip_vs_lblc_hash(tbl, en);
220 struct ip_vs_dest *old_dest;
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);
236 * Flush all the entries of the specified table.
238 static void ip_vs_lblc_flush(struct ip_vs_service *svc)
240 struct ip_vs_lblc_table *tbl = svc->sched_data;
241 struct ip_vs_lblc_entry *en;
242 struct hlist_node *next;
245 spin_lock_bh(&svc->sched_lock);
247 for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
248 hlist_for_each_entry_safe(en, next, &tbl->bucket[i], list) {
250 atomic_dec(&tbl->entries);
253 spin_unlock_bh(&svc->sched_lock);
256 static int sysctl_lblc_expiration(struct ip_vs_service *svc)
259 struct netns_ipvs *ipvs = net_ipvs(svc->net);
260 return ipvs->sysctl_lblc_expiration;
262 return DEFAULT_EXPIRATION;
266 static inline void ip_vs_lblc_full_check(struct ip_vs_service *svc)
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;
274 for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
275 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
277 spin_lock(&svc->sched_lock);
278 hlist_for_each_entry_safe(en, next, &tbl->bucket[j], list) {
281 sysctl_lblc_expiration(svc)))
285 atomic_dec(&tbl->entries);
287 spin_unlock(&svc->sched_lock);
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.
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
302 * The full expiration check is for this purpose now.
304 static void ip_vs_lblc_check_expire(unsigned long data)
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;
311 struct ip_vs_lblc_entry *en;
312 struct hlist_node *next;
314 if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) {
315 /* do full expiration check */
316 ip_vs_lblc_full_check(svc);
321 if (atomic_read(&tbl->entries) <= tbl->max_size) {
326 goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3;
327 if (goal > tbl->max_size/2)
328 goal = tbl->max_size/2;
330 for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
331 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
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))
339 atomic_dec(&tbl->entries);
342 spin_unlock(&svc->sched_lock);
349 mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL);
353 static int ip_vs_lblc_init_svc(struct ip_vs_service *svc)
356 struct ip_vs_lblc_table *tbl;
359 * Allocate the ip_vs_lblc_table for this service
361 tbl = kmalloc(sizeof(*tbl), GFP_KERNEL);
365 svc->sched_data = tbl;
366 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for "
367 "current service\n", sizeof(*tbl));
370 * Initialize the hash buckets
372 for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
373 INIT_HLIST_HEAD(&tbl->bucket[i]);
375 tbl->max_size = IP_VS_LBLC_TAB_SIZE*16;
381 * Hook periodic timer for garbage collection
383 setup_timer(&tbl->periodic_timer, ip_vs_lblc_check_expire,
385 mod_timer(&tbl->periodic_timer, jiffies + CHECK_EXPIRE_INTERVAL);
391 static void ip_vs_lblc_done_svc(struct ip_vs_service *svc)
393 struct ip_vs_lblc_table *tbl = svc->sched_data;
395 /* remove periodic timer */
396 del_timer_sync(&tbl->periodic_timer);
398 /* got to clean up table entries here */
399 ip_vs_lblc_flush(svc);
401 /* release the table itself */
402 kfree_rcu(tbl, rcu_head);
403 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n",
408 static inline struct ip_vs_dest *
409 __ip_vs_lblc_schedule(struct ip_vs_service *svc)
411 struct ip_vs_dest *dest, *least;
415 * We use the following formula to estimate the load:
416 * (dest overhead) / dest->weight
418 * Remember -- no floats in kernel mode!!!
419 * The comparison of h1*w2 > h2*w1 is equivalent to that of
421 * if every weight is larger than zero.
423 * The server with weight=0 is quiesced and will not receive any
426 list_for_each_entry_rcu(dest, &svc->destinations, n_list) {
427 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
429 if (atomic_read(&dest->weight) > 0) {
431 loh = ip_vs_dest_conn_overhead(least);
438 * Find the destination with the least load.
441 list_for_each_entry_continue_rcu(dest, &svc->destinations, n_list) {
442 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
445 doh = ip_vs_dest_conn_overhead(dest);
446 if (loh * atomic_read(&dest->weight) >
447 doh * atomic_read(&least->weight)) {
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),
457 atomic_read(&least->activeconns),
458 atomic_read(&least->refcnt),
459 atomic_read(&least->weight), loh);
466 * If this destination server is overloaded and there is a less loaded
467 * server, then return true.
470 is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc)
472 if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) {
473 struct ip_vs_dest *d;
475 list_for_each_entry_rcu(d, &svc->destinations, n_list) {
476 if (atomic_read(&d->activeconns)*2
477 < atomic_read(&d->weight)) {
487 * Locality-Based (weighted) Least-Connection scheduling
489 static struct ip_vs_dest *
490 ip_vs_lblc_schedule(struct ip_vs_service *svc, const struct sk_buff *skb)
492 struct ip_vs_lblc_table *tbl = svc->sched_data;
493 struct ip_vs_iphdr iph;
494 struct ip_vs_dest *dest = NULL;
495 struct ip_vs_lblc_entry *en;
497 ip_vs_fill_iph_addr_only(svc->af, skb, &iph);
499 IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
501 /* First look in our cache */
502 en = ip_vs_lblc_get(svc->af, tbl, &iph.daddr);
504 /* We only hold a read lock, but this is atomic */
505 en->lastuse = jiffies;
508 * If the destination is not available, i.e. it's in the trash,
509 * we must ignore it, as it may be removed from under our feet,
510 * if someone drops our reference count. Our caller only makes
511 * sure that destinations, that are not in the trash, are not
512 * moved to the trash, while we are scheduling. But anyone can
513 * free up entries from the trash at any time.
516 dest = rcu_dereference(en->dest);
517 if ((dest->flags & IP_VS_DEST_F_AVAILABLE) &&
518 atomic_read(&dest->weight) > 0 && !is_overloaded(dest, svc))
522 /* No cache entry or it is invalid, time to schedule */
523 dest = __ip_vs_lblc_schedule(svc);
525 ip_vs_scheduler_err(svc, "no destination available");
529 /* If we fail to create a cache entry, we'll just use the valid dest */
530 spin_lock_bh(&svc->sched_lock);
532 ip_vs_lblc_new(tbl, &iph.daddr, dest);
533 spin_unlock_bh(&svc->sched_lock);
536 IP_VS_DBG_BUF(6, "LBLC: destination IP address %s --> server %s:%d\n",
537 IP_VS_DBG_ADDR(svc->af, &iph.daddr),
538 IP_VS_DBG_ADDR(svc->af, &dest->addr), ntohs(dest->port));
545 * IPVS LBLC Scheduler structure
547 static struct ip_vs_scheduler ip_vs_lblc_scheduler =
550 .refcnt = ATOMIC_INIT(0),
551 .module = THIS_MODULE,
552 .n_list = LIST_HEAD_INIT(ip_vs_lblc_scheduler.n_list),
553 .init_service = ip_vs_lblc_init_svc,
554 .done_service = ip_vs_lblc_done_svc,
555 .schedule = ip_vs_lblc_schedule,
562 static int __net_init __ip_vs_lblc_init(struct net *net)
564 struct netns_ipvs *ipvs = net_ipvs(net);
569 if (!net_eq(net, &init_net)) {
570 ipvs->lblc_ctl_table = kmemdup(vs_vars_table,
571 sizeof(vs_vars_table),
573 if (ipvs->lblc_ctl_table == NULL)
576 /* Don't export sysctls to unprivileged users */
577 if (net->user_ns != &init_user_ns)
578 ipvs->lblc_ctl_table[0].procname = NULL;
581 ipvs->lblc_ctl_table = vs_vars_table;
582 ipvs->sysctl_lblc_expiration = DEFAULT_EXPIRATION;
583 ipvs->lblc_ctl_table[0].data = &ipvs->sysctl_lblc_expiration;
585 ipvs->lblc_ctl_header =
586 register_net_sysctl(net, "net/ipv4/vs", ipvs->lblc_ctl_table);
587 if (!ipvs->lblc_ctl_header) {
588 if (!net_eq(net, &init_net))
589 kfree(ipvs->lblc_ctl_table);
596 static void __net_exit __ip_vs_lblc_exit(struct net *net)
598 struct netns_ipvs *ipvs = net_ipvs(net);
600 unregister_net_sysctl_table(ipvs->lblc_ctl_header);
602 if (!net_eq(net, &init_net))
603 kfree(ipvs->lblc_ctl_table);
608 static int __net_init __ip_vs_lblc_init(struct net *net) { return 0; }
609 static void __net_exit __ip_vs_lblc_exit(struct net *net) { }
613 static struct pernet_operations ip_vs_lblc_ops = {
614 .init = __ip_vs_lblc_init,
615 .exit = __ip_vs_lblc_exit,
618 static int __init ip_vs_lblc_init(void)
622 ret = register_pernet_subsys(&ip_vs_lblc_ops);
626 ret = register_ip_vs_scheduler(&ip_vs_lblc_scheduler);
628 unregister_pernet_subsys(&ip_vs_lblc_ops);
632 static void __exit ip_vs_lblc_cleanup(void)
634 unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler);
635 unregister_pernet_subsys(&ip_vs_lblc_ops);
640 module_init(ip_vs_lblc_init);
641 module_exit(ip_vs_lblc_cleanup);
642 MODULE_LICENSE("GPL");