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 uninitilized 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)
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
73 static int sysctl_ip_vs_lblc_expiration = 24*60*60*HZ;
77 * for IPVS lblc entry hash table
79 #ifndef CONFIG_IP_VS_LBLC_TAB_BITS
80 #define CONFIG_IP_VS_LBLC_TAB_BITS 10
82 #define IP_VS_LBLC_TAB_BITS CONFIG_IP_VS_LBLC_TAB_BITS
83 #define IP_VS_LBLC_TAB_SIZE (1 << IP_VS_LBLC_TAB_BITS)
84 #define IP_VS_LBLC_TAB_MASK (IP_VS_LBLC_TAB_SIZE - 1)
88 * IPVS lblc entry represents an association between destination
89 * IP address and its destination server
91 struct ip_vs_lblc_entry {
92 struct list_head list;
93 int af; /* address family */
94 union nf_inet_addr addr; /* destination IP address */
95 struct ip_vs_dest *dest; /* real server (cache) */
96 unsigned long lastuse; /* last used time */
101 * IPVS lblc hash table
103 struct ip_vs_lblc_table {
104 struct list_head bucket[IP_VS_LBLC_TAB_SIZE]; /* hash bucket */
105 atomic_t entries; /* number of entries */
106 int max_size; /* maximum size of entries */
107 struct timer_list periodic_timer; /* collect stale entries */
108 int rover; /* rover for expire check */
109 int counter; /* counter for no expire */
114 * IPVS LBLC sysctl table
117 static ctl_table vs_vars_table[] = {
119 .procname = "lblc_expiration",
120 .data = &sysctl_ip_vs_lblc_expiration,
121 .maxlen = sizeof(int),
123 .proc_handler = proc_dointvec_jiffies,
128 static struct ctl_table_header * sysctl_header;
130 static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry *en)
134 * We don't kfree dest because it is refered either by its service
135 * or the trash dest list.
137 atomic_dec(&en->dest->refcnt);
143 * Returns hash value for IPVS LBLC entry
145 static inline unsigned
146 ip_vs_lblc_hashkey(int af, const union nf_inet_addr *addr)
148 __be32 addr_fold = addr->ip;
150 #ifdef CONFIG_IP_VS_IPV6
152 addr_fold = addr->ip6[0]^addr->ip6[1]^
153 addr->ip6[2]^addr->ip6[3];
155 return (ntohl(addr_fold)*2654435761UL) & IP_VS_LBLC_TAB_MASK;
160 * Hash an entry in the ip_vs_lblc_table.
161 * returns bool success.
164 ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en)
166 unsigned hash = ip_vs_lblc_hashkey(en->af, &en->addr);
168 list_add(&en->list, &tbl->bucket[hash]);
169 atomic_inc(&tbl->entries);
174 * Get ip_vs_lblc_entry associated with supplied parameters. Called under read
177 static inline struct ip_vs_lblc_entry *
178 ip_vs_lblc_get(int af, struct ip_vs_lblc_table *tbl,
179 const union nf_inet_addr *addr)
181 unsigned hash = ip_vs_lblc_hashkey(af, addr);
182 struct ip_vs_lblc_entry *en;
184 list_for_each_entry(en, &tbl->bucket[hash], list)
185 if (ip_vs_addr_equal(af, &en->addr, addr))
193 * Create or update an ip_vs_lblc_entry, which is a mapping of a destination IP
194 * address to a server. Called under write lock.
196 static inline struct ip_vs_lblc_entry *
197 ip_vs_lblc_new(struct ip_vs_lblc_table *tbl, const union nf_inet_addr *daddr,
198 struct ip_vs_dest *dest)
200 struct ip_vs_lblc_entry *en;
202 en = ip_vs_lblc_get(dest->af, tbl, daddr);
204 en = kmalloc(sizeof(*en), GFP_ATOMIC);
206 pr_err("%s(): no memory\n", __func__);
211 ip_vs_addr_copy(dest->af, &en->addr, daddr);
212 en->lastuse = jiffies;
214 atomic_inc(&dest->refcnt);
217 ip_vs_lblc_hash(tbl, en);
218 } else if (en->dest != dest) {
219 atomic_dec(&en->dest->refcnt);
220 atomic_inc(&dest->refcnt);
229 * Flush all the entries of the specified table.
231 static void ip_vs_lblc_flush(struct ip_vs_lblc_table *tbl)
233 struct ip_vs_lblc_entry *en, *nxt;
236 for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
237 list_for_each_entry_safe(en, nxt, &tbl->bucket[i], list) {
239 atomic_dec(&tbl->entries);
245 static inline void ip_vs_lblc_full_check(struct ip_vs_service *svc)
247 struct ip_vs_lblc_table *tbl = svc->sched_data;
248 struct ip_vs_lblc_entry *en, *nxt;
249 unsigned long now = jiffies;
252 for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
253 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
255 write_lock(&svc->sched_lock);
256 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
258 en->lastuse + sysctl_ip_vs_lblc_expiration))
262 atomic_dec(&tbl->entries);
264 write_unlock(&svc->sched_lock);
271 * Periodical timer handler for IPVS lblc table
272 * It is used to collect stale entries when the number of entries
273 * exceeds the maximum size of the table.
275 * Fixme: we probably need more complicated algorithm to collect
276 * entries that have not been used for a long time even
277 * if the number of entries doesn't exceed the maximum size
279 * The full expiration check is for this purpose now.
281 static void ip_vs_lblc_check_expire(unsigned long data)
283 struct ip_vs_service *svc = (struct ip_vs_service *) data;
284 struct ip_vs_lblc_table *tbl = svc->sched_data;
285 unsigned long now = jiffies;
288 struct ip_vs_lblc_entry *en, *nxt;
290 if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) {
291 /* do full expiration check */
292 ip_vs_lblc_full_check(svc);
297 if (atomic_read(&tbl->entries) <= tbl->max_size) {
302 goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3;
303 if (goal > tbl->max_size/2)
304 goal = tbl->max_size/2;
306 for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
307 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
309 write_lock(&svc->sched_lock);
310 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
311 if (time_before(now, en->lastuse + ENTRY_TIMEOUT))
315 atomic_dec(&tbl->entries);
318 write_unlock(&svc->sched_lock);
325 mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL);
329 static int ip_vs_lblc_init_svc(struct ip_vs_service *svc)
332 struct ip_vs_lblc_table *tbl;
335 * Allocate the ip_vs_lblc_table for this service
337 tbl = kmalloc(sizeof(*tbl), GFP_ATOMIC);
339 pr_err("%s(): no memory\n", __func__);
342 svc->sched_data = tbl;
343 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for "
344 "current service\n", sizeof(*tbl));
347 * Initialize the hash buckets
349 for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
350 INIT_LIST_HEAD(&tbl->bucket[i]);
352 tbl->max_size = IP_VS_LBLC_TAB_SIZE*16;
357 * Hook periodic timer for garbage collection
359 setup_timer(&tbl->periodic_timer, ip_vs_lblc_check_expire,
361 mod_timer(&tbl->periodic_timer, jiffies + CHECK_EXPIRE_INTERVAL);
367 static int ip_vs_lblc_done_svc(struct ip_vs_service *svc)
369 struct ip_vs_lblc_table *tbl = svc->sched_data;
371 /* remove periodic timer */
372 del_timer_sync(&tbl->periodic_timer);
374 /* got to clean up table entries here */
375 ip_vs_lblc_flush(tbl);
377 /* release the table itself */
379 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n",
386 static inline struct ip_vs_dest *
387 __ip_vs_lblc_schedule(struct ip_vs_service *svc)
389 struct ip_vs_dest *dest, *least;
393 * We think the overhead of processing active connections is fifty
394 * times higher than that of inactive connections in average. (This
395 * fifty times might not be accurate, we will change it later.) We
396 * use the following formula to estimate the overhead:
397 * dest->activeconns*50 + dest->inactconns
399 * (dest overhead) / dest->weight
401 * Remember -- no floats in kernel mode!!!
402 * The comparison of h1*w2 > h2*w1 is equivalent to that of
404 * if every weight is larger than zero.
406 * The server with weight=0 is quiesced and will not receive any
409 list_for_each_entry(dest, &svc->destinations, n_list) {
410 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
412 if (atomic_read(&dest->weight) > 0) {
414 loh = atomic_read(&least->activeconns) * 50
415 + atomic_read(&least->inactconns);
422 * Find the destination with the least load.
425 list_for_each_entry_continue(dest, &svc->destinations, n_list) {
426 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
429 doh = atomic_read(&dest->activeconns) * 50
430 + atomic_read(&dest->inactconns);
431 if (loh * atomic_read(&dest->weight) >
432 doh * atomic_read(&least->weight)) {
438 IP_VS_DBG_BUF(6, "LBLC: server %s:%d "
439 "activeconns %d refcnt %d weight %d overhead %d\n",
440 IP_VS_DBG_ADDR(least->af, &least->addr),
442 atomic_read(&least->activeconns),
443 atomic_read(&least->refcnt),
444 atomic_read(&least->weight), loh);
451 * If this destination server is overloaded and there is a less loaded
452 * server, then return true.
455 is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc)
457 if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) {
458 struct ip_vs_dest *d;
460 list_for_each_entry(d, &svc->destinations, n_list) {
461 if (atomic_read(&d->activeconns)*2
462 < atomic_read(&d->weight)) {
472 * Locality-Based (weighted) Least-Connection scheduling
474 static struct ip_vs_dest *
475 ip_vs_lblc_schedule(struct ip_vs_service *svc, const struct sk_buff *skb)
477 struct ip_vs_lblc_table *tbl = svc->sched_data;
478 struct ip_vs_iphdr iph;
479 struct ip_vs_dest *dest = NULL;
480 struct ip_vs_lblc_entry *en;
482 ip_vs_fill_iphdr(svc->af, skb_network_header(skb), &iph);
484 IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
486 /* First look in our cache */
487 read_lock(&svc->sched_lock);
488 en = ip_vs_lblc_get(svc->af, tbl, &iph.daddr);
490 /* We only hold a read lock, but this is atomic */
491 en->lastuse = jiffies;
494 * If the destination is not available, i.e. it's in the trash,
495 * we must ignore it, as it may be removed from under our feet,
496 * if someone drops our reference count. Our caller only makes
497 * sure that destinations, that are not in the trash, are not
498 * moved to the trash, while we are scheduling. But anyone can
499 * free up entries from the trash at any time.
502 if (en->dest->flags & IP_VS_DEST_F_AVAILABLE)
505 read_unlock(&svc->sched_lock);
507 /* If the destination has a weight and is not overloaded, use it */
508 if (dest && atomic_read(&dest->weight) > 0 && !is_overloaded(dest, svc))
511 /* No cache entry or it is invalid, time to schedule */
512 dest = __ip_vs_lblc_schedule(svc);
514 IP_VS_ERR_RL("LBLC: no destination available\n");
518 /* If we fail to create a cache entry, we'll just use the valid dest */
519 write_lock(&svc->sched_lock);
520 ip_vs_lblc_new(tbl, &iph.daddr, dest);
521 write_unlock(&svc->sched_lock);
524 IP_VS_DBG_BUF(6, "LBLC: destination IP address %s --> server %s:%d\n",
525 IP_VS_DBG_ADDR(svc->af, &iph.daddr),
526 IP_VS_DBG_ADDR(svc->af, &dest->addr), ntohs(dest->port));
533 * IPVS LBLC Scheduler structure
535 static struct ip_vs_scheduler ip_vs_lblc_scheduler =
538 .refcnt = ATOMIC_INIT(0),
539 .module = THIS_MODULE,
540 .n_list = LIST_HEAD_INIT(ip_vs_lblc_scheduler.n_list),
541 .init_service = ip_vs_lblc_init_svc,
542 .done_service = ip_vs_lblc_done_svc,
543 .schedule = ip_vs_lblc_schedule,
547 static int __init ip_vs_lblc_init(void)
551 sysctl_header = register_sysctl_paths(net_vs_ctl_path, vs_vars_table);
552 ret = register_ip_vs_scheduler(&ip_vs_lblc_scheduler);
554 unregister_sysctl_table(sysctl_header);
559 static void __exit ip_vs_lblc_cleanup(void)
561 unregister_sysctl_table(sysctl_header);
562 unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler);
566 module_init(ip_vs_lblc_init);
567 module_exit(ip_vs_lblc_cleanup);
568 MODULE_LICENSE("GPL");