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;
41 #include <linux/module.h>
42 #include <linux/kernel.h>
43 #include <linux/skbuff.h>
44 #include <linux/jiffies.h>
48 #include <linux/sysctl.h>
49 #include <net/net_namespace.h>
51 #include <net/ip_vs.h>
55 * It is for garbage collection of stale IPVS lblcr entries,
56 * when the table is full.
58 #define CHECK_EXPIRE_INTERVAL (60*HZ)
59 #define ENTRY_TIMEOUT (6*60*HZ)
62 * It is for full expiration check.
63 * When there is no partial expiration check (garbage collection)
64 * in a half hour, do a full expiration check to collect stale
65 * entries that haven't been touched for a day.
67 #define COUNT_FOR_FULL_EXPIRATION 30
68 static int sysctl_ip_vs_lblcr_expiration
= 24*60*60*HZ
;
72 * for IPVS lblcr entry hash table
74 #ifndef CONFIG_IP_VS_LBLCR_TAB_BITS
75 #define CONFIG_IP_VS_LBLCR_TAB_BITS 10
77 #define IP_VS_LBLCR_TAB_BITS CONFIG_IP_VS_LBLCR_TAB_BITS
78 #define IP_VS_LBLCR_TAB_SIZE (1 << IP_VS_LBLCR_TAB_BITS)
79 #define IP_VS_LBLCR_TAB_MASK (IP_VS_LBLCR_TAB_SIZE - 1)
83 * IPVS destination set structure and operations
85 struct ip_vs_dest_list
{
86 struct ip_vs_dest_list
*next
; /* list link */
87 struct ip_vs_dest
*dest
; /* destination server */
90 struct ip_vs_dest_set
{
91 atomic_t size
; /* set size */
92 unsigned long lastmod
; /* last modified time */
93 struct ip_vs_dest_list
*list
; /* destination list */
94 rwlock_t lock
; /* lock for this list */
98 static struct ip_vs_dest_list
*
99 ip_vs_dest_set_insert(struct ip_vs_dest_set
*set
, struct ip_vs_dest
*dest
)
101 struct ip_vs_dest_list
*e
;
103 for (e
=set
->list
; e
!=NULL
; e
=e
->next
) {
105 /* already existed */
109 e
= kmalloc(sizeof(*e
), GFP_ATOMIC
);
111 IP_VS_ERR("ip_vs_dest_set_insert(): no memory\n");
115 atomic_inc(&dest
->refcnt
);
118 /* link it to the list */
121 atomic_inc(&set
->size
);
123 set
->lastmod
= jiffies
;
128 ip_vs_dest_set_erase(struct ip_vs_dest_set
*set
, struct ip_vs_dest
*dest
)
130 struct ip_vs_dest_list
*e
, **ep
;
132 for (ep
=&set
->list
, e
=*ep
; e
!=NULL
; e
=*ep
) {
133 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_list
*e
, **ep
;
150 write_lock(&set
->lock
);
151 for (ep
=&set
->list
, e
=*ep
; e
!=NULL
; e
=*ep
) {
154 * We don't kfree dest because it is refered either
155 * by its service or by the trash dest list.
157 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_list
*e
;
167 struct ip_vs_dest
*dest
, *least
;
173 /* select the first destination server, whose weight > 0 */
174 for (e
=set
->list
; e
!=NULL
; e
=e
->next
) {
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
= atomic_read(&least
->activeconns
) * 50
182 + atomic_read(&least
->inactconns
);
188 /* find the destination with the weighted least load */
190 for (e
=e
->next
; e
!=NULL
; e
=e
->next
) {
192 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
195 doh
= atomic_read(&dest
->activeconns
) * 50
196 + atomic_read(&dest
->inactconns
);
197 if ((loh
* atomic_read(&dest
->weight
) >
198 doh
* atomic_read(&least
->weight
))
199 && (dest
->flags
& IP_VS_DEST_F_AVAILABLE
)) {
205 IP_VS_DBG_BUF(6, "ip_vs_dest_set_min: server %s:%d "
206 "activeconns %d refcnt %d weight %d overhead %d\n",
207 IP_VS_DBG_ADDR(least
->af
, &least
->addr
),
209 atomic_read(&least
->activeconns
),
210 atomic_read(&least
->refcnt
),
211 atomic_read(&least
->weight
), loh
);
216 /* get weighted most-connection node in the destination set */
217 static inline struct ip_vs_dest
*ip_vs_dest_set_max(struct ip_vs_dest_set
*set
)
219 register struct ip_vs_dest_list
*e
;
220 struct ip_vs_dest
*dest
, *most
;
226 /* select the first destination server, whose weight > 0 */
227 for (e
=set
->list
; e
!=NULL
; e
=e
->next
) {
229 if (atomic_read(&most
->weight
) > 0) {
230 moh
= atomic_read(&most
->activeconns
) * 50
231 + atomic_read(&most
->inactconns
);
237 /* find the destination with the weighted most load */
239 for (e
=e
->next
; e
!=NULL
; e
=e
->next
) {
241 doh
= atomic_read(&dest
->activeconns
) * 50
242 + atomic_read(&dest
->inactconns
);
243 /* moh/mw < doh/dw ==> moh*dw < doh*mw, where mw,dw>0 */
244 if ((moh
* atomic_read(&dest
->weight
) <
245 doh
* atomic_read(&most
->weight
))
246 && (atomic_read(&dest
->weight
) > 0)) {
252 IP_VS_DBG_BUF(6, "ip_vs_dest_set_max: server %s:%d "
253 "activeconns %d refcnt %d weight %d overhead %d\n",
254 IP_VS_DBG_ADDR(most
->af
, &most
->addr
), ntohs(most
->port
),
255 atomic_read(&most
->activeconns
),
256 atomic_read(&most
->refcnt
),
257 atomic_read(&most
->weight
), moh
);
263 * IPVS lblcr entry represents an association between destination
264 * IP address and its destination server set
266 struct ip_vs_lblcr_entry
{
267 struct list_head list
;
268 int af
; /* address family */
269 union nf_inet_addr addr
; /* destination IP address */
270 struct ip_vs_dest_set set
; /* destination server set */
271 unsigned long lastuse
; /* last used time */
276 * IPVS lblcr hash table
278 struct ip_vs_lblcr_table
{
279 struct list_head bucket
[IP_VS_LBLCR_TAB_SIZE
]; /* hash bucket */
280 atomic_t entries
; /* number of entries */
281 int max_size
; /* maximum size of entries */
282 struct timer_list periodic_timer
; /* collect stale entries */
283 int rover
; /* rover for expire check */
284 int counter
; /* counter for no expire */
289 * IPVS LBLCR sysctl table
292 static ctl_table vs_vars_table
[] = {
294 .procname
= "lblcr_expiration",
295 .data
= &sysctl_ip_vs_lblcr_expiration
,
296 .maxlen
= sizeof(int),
298 .proc_handler
= proc_dointvec_jiffies
,
303 static struct ctl_table_header
* sysctl_header
;
305 static inline void ip_vs_lblcr_free(struct ip_vs_lblcr_entry
*en
)
308 ip_vs_dest_set_eraseall(&en
->set
);
314 * Returns hash value for IPVS LBLCR entry
316 static inline unsigned
317 ip_vs_lblcr_hashkey(int af
, const union nf_inet_addr
*addr
)
319 __be32 addr_fold
= addr
->ip
;
321 #ifdef CONFIG_IP_VS_IPV6
323 addr_fold
= addr
->ip6
[0]^addr
->ip6
[1]^
324 addr
->ip6
[2]^addr
->ip6
[3];
326 return (ntohl(addr_fold
)*2654435761UL) & IP_VS_LBLCR_TAB_MASK
;
331 * Hash an entry in the ip_vs_lblcr_table.
332 * returns bool success.
335 ip_vs_lblcr_hash(struct ip_vs_lblcr_table
*tbl
, struct ip_vs_lblcr_entry
*en
)
337 unsigned hash
= ip_vs_lblcr_hashkey(en
->af
, &en
->addr
);
339 list_add(&en
->list
, &tbl
->bucket
[hash
]);
340 atomic_inc(&tbl
->entries
);
345 * Get ip_vs_lblcr_entry associated with supplied parameters. Called under
348 static inline struct ip_vs_lblcr_entry
*
349 ip_vs_lblcr_get(int af
, struct ip_vs_lblcr_table
*tbl
,
350 const union nf_inet_addr
*addr
)
352 unsigned hash
= ip_vs_lblcr_hashkey(af
, addr
);
353 struct ip_vs_lblcr_entry
*en
;
355 list_for_each_entry(en
, &tbl
->bucket
[hash
], list
)
356 if (ip_vs_addr_equal(af
, &en
->addr
, addr
))
364 * Create or update an ip_vs_lblcr_entry, which is a mapping of a destination
365 * IP address to a server. Called under write lock.
367 static inline struct ip_vs_lblcr_entry
*
368 ip_vs_lblcr_new(struct ip_vs_lblcr_table
*tbl
, const union nf_inet_addr
*daddr
,
369 struct ip_vs_dest
*dest
)
371 struct ip_vs_lblcr_entry
*en
;
373 en
= ip_vs_lblcr_get(dest
->af
, tbl
, daddr
);
375 en
= kmalloc(sizeof(*en
), GFP_ATOMIC
);
377 IP_VS_ERR("ip_vs_lblcr_new(): no memory\n");
382 ip_vs_addr_copy(dest
->af
, &en
->addr
, daddr
);
383 en
->lastuse
= jiffies
;
385 /* initilize its dest set */
386 atomic_set(&(en
->set
.size
), 0);
388 rwlock_init(&en
->set
.lock
);
390 ip_vs_lblcr_hash(tbl
, en
);
393 write_lock(&en
->set
.lock
);
394 ip_vs_dest_set_insert(&en
->set
, dest
);
395 write_unlock(&en
->set
.lock
);
402 * Flush all the entries of the specified table.
404 static void ip_vs_lblcr_flush(struct ip_vs_lblcr_table
*tbl
)
407 struct ip_vs_lblcr_entry
*en
, *nxt
;
409 /* No locking required, only called during cleanup. */
410 for (i
=0; i
<IP_VS_LBLCR_TAB_SIZE
; i
++) {
411 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[i
], list
) {
412 ip_vs_lblcr_free(en
);
418 static inline void ip_vs_lblcr_full_check(struct ip_vs_service
*svc
)
420 struct ip_vs_lblcr_table
*tbl
= svc
->sched_data
;
421 unsigned long now
= jiffies
;
423 struct ip_vs_lblcr_entry
*en
, *nxt
;
425 for (i
=0, j
=tbl
->rover
; i
<IP_VS_LBLCR_TAB_SIZE
; i
++) {
426 j
= (j
+ 1) & IP_VS_LBLCR_TAB_MASK
;
428 write_lock(&svc
->sched_lock
);
429 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[j
], list
) {
430 if (time_after(en
->lastuse
+sysctl_ip_vs_lblcr_expiration
,
434 ip_vs_lblcr_free(en
);
435 atomic_dec(&tbl
->entries
);
437 write_unlock(&svc
->sched_lock
);
444 * Periodical timer handler for IPVS lblcr table
445 * It is used to collect stale entries when the number of entries
446 * exceeds the maximum size of the table.
448 * Fixme: we probably need more complicated algorithm to collect
449 * entries that have not been used for a long time even
450 * if the number of entries doesn't exceed the maximum size
452 * The full expiration check is for this purpose now.
454 static void ip_vs_lblcr_check_expire(unsigned long data
)
456 struct ip_vs_service
*svc
= (struct ip_vs_service
*) data
;
457 struct ip_vs_lblcr_table
*tbl
= svc
->sched_data
;
458 unsigned long now
= jiffies
;
461 struct ip_vs_lblcr_entry
*en
, *nxt
;
463 if ((tbl
->counter
% COUNT_FOR_FULL_EXPIRATION
) == 0) {
464 /* do full expiration check */
465 ip_vs_lblcr_full_check(svc
);
470 if (atomic_read(&tbl
->entries
) <= tbl
->max_size
) {
475 goal
= (atomic_read(&tbl
->entries
) - tbl
->max_size
)*4/3;
476 if (goal
> tbl
->max_size
/2)
477 goal
= tbl
->max_size
/2;
479 for (i
=0, j
=tbl
->rover
; i
<IP_VS_LBLCR_TAB_SIZE
; i
++) {
480 j
= (j
+ 1) & IP_VS_LBLCR_TAB_MASK
;
482 write_lock(&svc
->sched_lock
);
483 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[j
], list
) {
484 if (time_before(now
, en
->lastuse
+ENTRY_TIMEOUT
))
487 ip_vs_lblcr_free(en
);
488 atomic_dec(&tbl
->entries
);
491 write_unlock(&svc
->sched_lock
);
498 mod_timer(&tbl
->periodic_timer
, jiffies
+CHECK_EXPIRE_INTERVAL
);
501 static int ip_vs_lblcr_init_svc(struct ip_vs_service
*svc
)
504 struct ip_vs_lblcr_table
*tbl
;
507 * Allocate the ip_vs_lblcr_table for this service
509 tbl
= kmalloc(sizeof(*tbl
), GFP_ATOMIC
);
511 IP_VS_ERR("ip_vs_lblcr_init_svc(): no memory\n");
514 svc
->sched_data
= tbl
;
515 IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) allocated for "
516 "current service\n", sizeof(*tbl
));
519 * Initialize the hash buckets
521 for (i
=0; i
<IP_VS_LBLCR_TAB_SIZE
; i
++) {
522 INIT_LIST_HEAD(&tbl
->bucket
[i
]);
524 tbl
->max_size
= IP_VS_LBLCR_TAB_SIZE
*16;
529 * Hook periodic timer for garbage collection
531 setup_timer(&tbl
->periodic_timer
, ip_vs_lblcr_check_expire
,
533 mod_timer(&tbl
->periodic_timer
, jiffies
+ CHECK_EXPIRE_INTERVAL
);
539 static int ip_vs_lblcr_done_svc(struct ip_vs_service
*svc
)
541 struct ip_vs_lblcr_table
*tbl
= svc
->sched_data
;
543 /* remove periodic timer */
544 del_timer_sync(&tbl
->periodic_timer
);
546 /* got to clean up table entries here */
547 ip_vs_lblcr_flush(tbl
);
549 /* release the table itself */
551 IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) released\n",
558 static inline struct ip_vs_dest
*
559 __ip_vs_lblcr_schedule(struct ip_vs_service
*svc
)
561 struct ip_vs_dest
*dest
, *least
;
565 * We think the overhead of processing active connections is fifty
566 * times higher than that of inactive connections in average. (This
567 * fifty times might not be accurate, we will change it later.) We
568 * use the following formula to estimate the overhead:
569 * dest->activeconns*50 + dest->inactconns
571 * (dest overhead) / dest->weight
573 * Remember -- no floats in kernel mode!!!
574 * The comparison of h1*w2 > h2*w1 is equivalent to that of
576 * if every weight is larger than zero.
578 * The server with weight=0 is quiesced and will not receive any
581 list_for_each_entry(dest
, &svc
->destinations
, n_list
) {
582 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
585 if (atomic_read(&dest
->weight
) > 0) {
587 loh
= atomic_read(&least
->activeconns
) * 50
588 + atomic_read(&least
->inactconns
);
595 * Find the destination with the least load.
598 list_for_each_entry_continue(dest
, &svc
->destinations
, n_list
) {
599 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
602 doh
= atomic_read(&dest
->activeconns
) * 50
603 + atomic_read(&dest
->inactconns
);
604 if (loh
* atomic_read(&dest
->weight
) >
605 doh
* atomic_read(&least
->weight
)) {
611 IP_VS_DBG_BUF(6, "LBLCR: server %s:%d "
612 "activeconns %d refcnt %d weight %d overhead %d\n",
613 IP_VS_DBG_ADDR(least
->af
, &least
->addr
),
615 atomic_read(&least
->activeconns
),
616 atomic_read(&least
->refcnt
),
617 atomic_read(&least
->weight
), loh
);
624 * If this destination server is overloaded and there is a less loaded
625 * server, then return true.
628 is_overloaded(struct ip_vs_dest
*dest
, struct ip_vs_service
*svc
)
630 if (atomic_read(&dest
->activeconns
) > atomic_read(&dest
->weight
)) {
631 struct ip_vs_dest
*d
;
633 list_for_each_entry(d
, &svc
->destinations
, n_list
) {
634 if (atomic_read(&d
->activeconns
)*2
635 < atomic_read(&d
->weight
)) {
645 * Locality-Based (weighted) Least-Connection scheduling
647 static struct ip_vs_dest
*
648 ip_vs_lblcr_schedule(struct ip_vs_service
*svc
, const struct sk_buff
*skb
)
650 struct ip_vs_lblcr_table
*tbl
= svc
->sched_data
;
651 struct ip_vs_iphdr iph
;
652 struct ip_vs_dest
*dest
= NULL
;
653 struct ip_vs_lblcr_entry
*en
;
655 ip_vs_fill_iphdr(svc
->af
, skb_network_header(skb
), &iph
);
657 IP_VS_DBG(6, "ip_vs_lblcr_schedule(): Scheduling...\n");
659 /* First look in our cache */
660 read_lock(&svc
->sched_lock
);
661 en
= ip_vs_lblcr_get(svc
->af
, tbl
, &iph
.daddr
);
663 /* We only hold a read lock, but this is atomic */
664 en
->lastuse
= jiffies
;
666 /* Get the least loaded destination */
667 read_lock(&en
->set
.lock
);
668 dest
= ip_vs_dest_set_min(&en
->set
);
669 read_unlock(&en
->set
.lock
);
671 /* More than one destination + enough time passed by, cleanup */
672 if (atomic_read(&en
->set
.size
) > 1 &&
673 time_after(jiffies
, en
->set
.lastmod
+
674 sysctl_ip_vs_lblcr_expiration
)) {
675 struct ip_vs_dest
*m
;
677 write_lock(&en
->set
.lock
);
678 m
= ip_vs_dest_set_max(&en
->set
);
680 ip_vs_dest_set_erase(&en
->set
, m
);
681 write_unlock(&en
->set
.lock
);
684 /* If the destination is not overloaded, use it */
685 if (dest
&& !is_overloaded(dest
, svc
)) {
686 read_unlock(&svc
->sched_lock
);
690 /* The cache entry is invalid, time to schedule */
691 dest
= __ip_vs_lblcr_schedule(svc
);
693 IP_VS_ERR_RL("LBLCR: no destination available\n");
694 read_unlock(&svc
->sched_lock
);
698 /* Update our cache entry */
699 write_lock(&en
->set
.lock
);
700 ip_vs_dest_set_insert(&en
->set
, dest
);
701 write_unlock(&en
->set
.lock
);
703 read_unlock(&svc
->sched_lock
);
708 /* No cache entry, time to schedule */
709 dest
= __ip_vs_lblcr_schedule(svc
);
711 IP_VS_DBG(1, "no destination available\n");
715 /* If we fail to create a cache entry, we'll just use the valid dest */
716 write_lock(&svc
->sched_lock
);
717 ip_vs_lblcr_new(tbl
, &iph
.daddr
, dest
);
718 write_unlock(&svc
->sched_lock
);
721 IP_VS_DBG_BUF(6, "LBLCR: destination IP address %s --> server %s:%d\n",
722 IP_VS_DBG_ADDR(svc
->af
, &iph
.daddr
),
723 IP_VS_DBG_ADDR(svc
->af
, &dest
->addr
), ntohs(dest
->port
));
730 * IPVS LBLCR Scheduler structure
732 static struct ip_vs_scheduler ip_vs_lblcr_scheduler
=
735 .refcnt
= ATOMIC_INIT(0),
736 .module
= THIS_MODULE
,
737 .n_list
= LIST_HEAD_INIT(ip_vs_lblcr_scheduler
.n_list
),
738 .init_service
= ip_vs_lblcr_init_svc
,
739 .done_service
= ip_vs_lblcr_done_svc
,
740 .schedule
= ip_vs_lblcr_schedule
,
744 static int __init
ip_vs_lblcr_init(void)
748 sysctl_header
= register_sysctl_paths(net_vs_ctl_path
, vs_vars_table
);
749 ret
= register_ip_vs_scheduler(&ip_vs_lblcr_scheduler
);
751 unregister_sysctl_table(sysctl_header
);
756 static void __exit
ip_vs_lblcr_cleanup(void)
758 unregister_sysctl_table(sysctl_header
);
759 unregister_ip_vs_scheduler(&ip_vs_lblcr_scheduler
);
763 module_init(ip_vs_lblcr_init
);
764 module_exit(ip_vs_lblcr_cleanup
);
765 MODULE_LICENSE("GPL");