2 * IPVS: Locality-Based Least-Connection with Replication scheduler
4 * Version: $Id: ip_vs_lblcr.c,v 1.11 2002/09/15 08:14:08 wensong Exp $
6 * Authors: Wensong Zhang <wensong@gnuchina.org>
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public License
10 * as published by the Free Software Foundation; either version
11 * 2 of the License, or (at your option) any later version.
14 * Julian Anastasov : Added the missing (dest->weight>0)
15 * condition in the ip_vs_dest_set_max.
20 * The lblc/r algorithm is as follows (pseudo code):
22 * if serverSet[dest_ip] is null then
23 * n, serverSet[dest_ip] <- {weighted least-conn node};
25 * n <- {least-conn (alive) node in serverSet[dest_ip]};
27 * (n.conns>n.weight AND
28 * there is a node m with m.conns<m.weight/2) then
29 * n <- {weighted least-conn node};
30 * add n to serverSet[dest_ip];
31 * if |serverSet[dest_ip]| > 1 AND
32 * now - serverSet[dest_ip].lastMod > T then
33 * m <- {most conn node in serverSet[dest_ip]};
34 * remove m from serverSet[dest_ip];
35 * if serverSet[dest_ip] changed then
36 * serverSet[dest_ip].lastMod <- now;
43 #include <linux/module.h>
44 #include <linux/kernel.h>
45 #include <linux/skbuff.h>
46 #include <linux/jiffies.h>
50 #include <linux/sysctl.h>
51 #include <net/net_namespace.h>
53 #include <net/ip_vs.h>
57 * It is for garbage collection of stale IPVS lblcr entries,
58 * when the table is full.
60 #define CHECK_EXPIRE_INTERVAL (60*HZ)
61 #define ENTRY_TIMEOUT (6*60*HZ)
64 * It is for full expiration check.
65 * When there is no partial expiration check (garbage collection)
66 * in a half hour, do a full expiration check to collect stale
67 * entries that haven't been touched for a day.
69 #define COUNT_FOR_FULL_EXPIRATION 30
70 static int sysctl_ip_vs_lblcr_expiration
= 24*60*60*HZ
;
74 * for IPVS lblcr entry hash table
76 #ifndef CONFIG_IP_VS_LBLCR_TAB_BITS
77 #define CONFIG_IP_VS_LBLCR_TAB_BITS 10
79 #define IP_VS_LBLCR_TAB_BITS CONFIG_IP_VS_LBLCR_TAB_BITS
80 #define IP_VS_LBLCR_TAB_SIZE (1 << IP_VS_LBLCR_TAB_BITS)
81 #define IP_VS_LBLCR_TAB_MASK (IP_VS_LBLCR_TAB_SIZE - 1)
85 * IPVS destination set structure and operations
87 struct ip_vs_dest_list
{
88 struct ip_vs_dest_list
*next
; /* list link */
89 struct ip_vs_dest
*dest
; /* destination server */
92 struct ip_vs_dest_set
{
93 atomic_t size
; /* set size */
94 unsigned long lastmod
; /* last modified time */
95 struct ip_vs_dest_list
*list
; /* destination list */
96 rwlock_t lock
; /* lock for this list */
100 static struct ip_vs_dest_list
*
101 ip_vs_dest_set_insert(struct ip_vs_dest_set
*set
, struct ip_vs_dest
*dest
)
103 struct ip_vs_dest_list
*e
;
105 for (e
=set
->list
; e
!=NULL
; e
=e
->next
) {
107 /* already existed */
111 e
= kmalloc(sizeof(struct ip_vs_dest_list
), GFP_ATOMIC
);
113 IP_VS_ERR("ip_vs_dest_set_insert(): no memory\n");
117 atomic_inc(&dest
->refcnt
);
120 /* link it to the list */
121 write_lock(&set
->lock
);
124 atomic_inc(&set
->size
);
125 write_unlock(&set
->lock
);
127 set
->lastmod
= jiffies
;
132 ip_vs_dest_set_erase(struct ip_vs_dest_set
*set
, struct ip_vs_dest
*dest
)
134 struct ip_vs_dest_list
*e
, **ep
;
136 write_lock(&set
->lock
);
137 for (ep
=&set
->list
, e
=*ep
; e
!=NULL
; e
=*ep
) {
138 if (e
->dest
== dest
) {
141 atomic_dec(&set
->size
);
142 set
->lastmod
= jiffies
;
143 atomic_dec(&e
->dest
->refcnt
);
149 write_unlock(&set
->lock
);
152 static void ip_vs_dest_set_eraseall(struct ip_vs_dest_set
*set
)
154 struct ip_vs_dest_list
*e
, **ep
;
156 write_lock(&set
->lock
);
157 for (ep
=&set
->list
, e
=*ep
; e
!=NULL
; e
=*ep
) {
160 * We don't kfree dest because it is refered either
161 * by its service or by the trash dest list.
163 atomic_dec(&e
->dest
->refcnt
);
166 write_unlock(&set
->lock
);
169 /* get weighted least-connection node in the destination set */
170 static inline struct ip_vs_dest
*ip_vs_dest_set_min(struct ip_vs_dest_set
*set
)
172 register struct ip_vs_dest_list
*e
;
173 struct ip_vs_dest
*dest
, *least
;
179 read_lock(&set
->lock
);
180 /* select the first destination server, whose weight > 0 */
181 for (e
=set
->list
; e
!=NULL
; e
=e
->next
) {
183 if (least
->flags
& IP_VS_DEST_F_OVERLOAD
)
186 if ((atomic_read(&least
->weight
) > 0)
187 && (least
->flags
& IP_VS_DEST_F_AVAILABLE
)) {
188 loh
= atomic_read(&least
->activeconns
) * 50
189 + atomic_read(&least
->inactconns
);
193 read_unlock(&set
->lock
);
196 /* find the destination with the weighted least load */
198 for (e
=e
->next
; e
!=NULL
; e
=e
->next
) {
200 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
203 doh
= atomic_read(&dest
->activeconns
) * 50
204 + atomic_read(&dest
->inactconns
);
205 if ((loh
* atomic_read(&dest
->weight
) >
206 doh
* atomic_read(&least
->weight
))
207 && (dest
->flags
& IP_VS_DEST_F_AVAILABLE
)) {
212 read_unlock(&set
->lock
);
214 IP_VS_DBG(6, "ip_vs_dest_set_min: server %d.%d.%d.%d:%d "
215 "activeconns %d refcnt %d weight %d overhead %d\n",
216 NIPQUAD(least
->addr
), ntohs(least
->port
),
217 atomic_read(&least
->activeconns
),
218 atomic_read(&least
->refcnt
),
219 atomic_read(&least
->weight
), loh
);
224 /* get weighted most-connection node in the destination set */
225 static inline struct ip_vs_dest
*ip_vs_dest_set_max(struct ip_vs_dest_set
*set
)
227 register struct ip_vs_dest_list
*e
;
228 struct ip_vs_dest
*dest
, *most
;
234 read_lock(&set
->lock
);
235 /* select the first destination server, whose weight > 0 */
236 for (e
=set
->list
; e
!=NULL
; e
=e
->next
) {
238 if (atomic_read(&most
->weight
) > 0) {
239 moh
= atomic_read(&most
->activeconns
) * 50
240 + atomic_read(&most
->inactconns
);
244 read_unlock(&set
->lock
);
247 /* find the destination with the weighted most load */
249 for (e
=e
->next
; e
!=NULL
; e
=e
->next
) {
251 doh
= atomic_read(&dest
->activeconns
) * 50
252 + atomic_read(&dest
->inactconns
);
253 /* moh/mw < doh/dw ==> moh*dw < doh*mw, where mw,dw>0 */
254 if ((moh
* atomic_read(&dest
->weight
) <
255 doh
* atomic_read(&most
->weight
))
256 && (atomic_read(&dest
->weight
) > 0)) {
261 read_unlock(&set
->lock
);
263 IP_VS_DBG(6, "ip_vs_dest_set_max: server %d.%d.%d.%d:%d "
264 "activeconns %d refcnt %d weight %d overhead %d\n",
265 NIPQUAD(most
->addr
), ntohs(most
->port
),
266 atomic_read(&most
->activeconns
),
267 atomic_read(&most
->refcnt
),
268 atomic_read(&most
->weight
), moh
);
274 * IPVS lblcr entry represents an association between destination
275 * IP address and its destination server set
277 struct ip_vs_lblcr_entry
{
278 struct list_head list
;
279 __be32 addr
; /* destination IP address */
280 struct ip_vs_dest_set set
; /* destination server set */
281 unsigned long lastuse
; /* last used time */
286 * IPVS lblcr hash table
288 struct ip_vs_lblcr_table
{
289 rwlock_t lock
; /* lock for this table */
290 struct list_head bucket
[IP_VS_LBLCR_TAB_SIZE
]; /* hash bucket */
291 atomic_t entries
; /* number of entries */
292 int max_size
; /* maximum size of entries */
293 struct timer_list periodic_timer
; /* collect stale entries */
294 int rover
; /* rover for expire check */
295 int counter
; /* counter for no expire */
300 * IPVS LBLCR sysctl table
303 static ctl_table vs_vars_table
[] = {
305 .ctl_name
= NET_IPV4_VS_LBLCR_EXPIRE
,
306 .procname
= "lblcr_expiration",
307 .data
= &sysctl_ip_vs_lblcr_expiration
,
308 .maxlen
= sizeof(int),
310 .proc_handler
= &proc_dointvec_jiffies
,
315 static ctl_table vs_table
[] = {
317 .ctl_name
= NET_IPV4_VS
,
320 .child
= vs_vars_table
325 static ctl_table ipvs_ipv4_table
[] = {
327 .ctl_name
= NET_IPV4
,
335 static ctl_table lblcr_root_table
[] = {
340 .child
= ipvs_ipv4_table
345 static struct ctl_table_header
* sysctl_header
;
348 * new/free a ip_vs_lblcr_entry, which is a mapping of a destination
349 * IP address to a server.
351 static inline struct ip_vs_lblcr_entry
*ip_vs_lblcr_new(__be32 daddr
)
353 struct ip_vs_lblcr_entry
*en
;
355 en
= kmalloc(sizeof(struct ip_vs_lblcr_entry
), GFP_ATOMIC
);
357 IP_VS_ERR("ip_vs_lblcr_new(): no memory\n");
361 INIT_LIST_HEAD(&en
->list
);
364 /* initilize its dest set */
365 atomic_set(&(en
->set
.size
), 0);
367 rwlock_init(&en
->set
.lock
);
373 static inline void ip_vs_lblcr_free(struct ip_vs_lblcr_entry
*en
)
376 ip_vs_dest_set_eraseall(&en
->set
);
382 * Returns hash value for IPVS LBLCR entry
384 static inline unsigned ip_vs_lblcr_hashkey(__be32 addr
)
386 return (ntohl(addr
)*2654435761UL) & IP_VS_LBLCR_TAB_MASK
;
391 * Hash an entry in the ip_vs_lblcr_table.
392 * returns bool success.
395 ip_vs_lblcr_hash(struct ip_vs_lblcr_table
*tbl
, struct ip_vs_lblcr_entry
*en
)
399 if (!list_empty(&en
->list
)) {
400 IP_VS_ERR("ip_vs_lblcr_hash(): request for already hashed, "
401 "called from %p\n", __builtin_return_address(0));
406 * Hash by destination IP address
408 hash
= ip_vs_lblcr_hashkey(en
->addr
);
410 write_lock(&tbl
->lock
);
411 list_add(&en
->list
, &tbl
->bucket
[hash
]);
412 atomic_inc(&tbl
->entries
);
413 write_unlock(&tbl
->lock
);
420 * Get ip_vs_lblcr_entry associated with supplied parameters.
422 static inline struct ip_vs_lblcr_entry
*
423 ip_vs_lblcr_get(struct ip_vs_lblcr_table
*tbl
, __be32 addr
)
426 struct ip_vs_lblcr_entry
*en
;
428 hash
= ip_vs_lblcr_hashkey(addr
);
430 read_lock(&tbl
->lock
);
432 list_for_each_entry(en
, &tbl
->bucket
[hash
], list
) {
433 if (en
->addr
== addr
) {
435 read_unlock(&tbl
->lock
);
440 read_unlock(&tbl
->lock
);
447 * Flush all the entries of the specified table.
449 static void ip_vs_lblcr_flush(struct ip_vs_lblcr_table
*tbl
)
452 struct ip_vs_lblcr_entry
*en
, *nxt
;
454 for (i
=0; i
<IP_VS_LBLCR_TAB_SIZE
; i
++) {
455 write_lock(&tbl
->lock
);
456 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[i
], list
) {
457 ip_vs_lblcr_free(en
);
458 atomic_dec(&tbl
->entries
);
460 write_unlock(&tbl
->lock
);
465 static inline void ip_vs_lblcr_full_check(struct ip_vs_lblcr_table
*tbl
)
467 unsigned long now
= jiffies
;
469 struct ip_vs_lblcr_entry
*en
, *nxt
;
471 for (i
=0, j
=tbl
->rover
; i
<IP_VS_LBLCR_TAB_SIZE
; i
++) {
472 j
= (j
+ 1) & IP_VS_LBLCR_TAB_MASK
;
474 write_lock(&tbl
->lock
);
475 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[j
], list
) {
476 if (time_after(en
->lastuse
+sysctl_ip_vs_lblcr_expiration
,
480 ip_vs_lblcr_free(en
);
481 atomic_dec(&tbl
->entries
);
483 write_unlock(&tbl
->lock
);
490 * Periodical timer handler for IPVS lblcr table
491 * It is used to collect stale entries when the number of entries
492 * exceeds the maximum size of the table.
494 * Fixme: we probably need more complicated algorithm to collect
495 * entries that have not been used for a long time even
496 * if the number of entries doesn't exceed the maximum size
498 * The full expiration check is for this purpose now.
500 static void ip_vs_lblcr_check_expire(unsigned long data
)
502 struct ip_vs_lblcr_table
*tbl
;
503 unsigned long now
= jiffies
;
506 struct ip_vs_lblcr_entry
*en
, *nxt
;
508 tbl
= (struct ip_vs_lblcr_table
*)data
;
510 if ((tbl
->counter
% COUNT_FOR_FULL_EXPIRATION
) == 0) {
511 /* do full expiration check */
512 ip_vs_lblcr_full_check(tbl
);
517 if (atomic_read(&tbl
->entries
) <= tbl
->max_size
) {
522 goal
= (atomic_read(&tbl
->entries
) - tbl
->max_size
)*4/3;
523 if (goal
> tbl
->max_size
/2)
524 goal
= tbl
->max_size
/2;
526 for (i
=0, j
=tbl
->rover
; i
<IP_VS_LBLCR_TAB_SIZE
; i
++) {
527 j
= (j
+ 1) & IP_VS_LBLCR_TAB_MASK
;
529 write_lock(&tbl
->lock
);
530 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[j
], list
) {
531 if (time_before(now
, en
->lastuse
+ENTRY_TIMEOUT
))
534 ip_vs_lblcr_free(en
);
535 atomic_dec(&tbl
->entries
);
538 write_unlock(&tbl
->lock
);
545 mod_timer(&tbl
->periodic_timer
, jiffies
+CHECK_EXPIRE_INTERVAL
);
548 static int ip_vs_lblcr_init_svc(struct ip_vs_service
*svc
)
551 struct ip_vs_lblcr_table
*tbl
;
554 * Allocate the ip_vs_lblcr_table for this service
556 tbl
= kmalloc(sizeof(struct ip_vs_lblcr_table
), GFP_ATOMIC
);
558 IP_VS_ERR("ip_vs_lblcr_init_svc(): no memory\n");
561 svc
->sched_data
= tbl
;
562 IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) allocated for "
564 sizeof(struct ip_vs_lblcr_table
));
567 * Initialize the hash buckets
569 for (i
=0; i
<IP_VS_LBLCR_TAB_SIZE
; i
++) {
570 INIT_LIST_HEAD(&tbl
->bucket
[i
]);
572 rwlock_init(&tbl
->lock
);
573 tbl
->max_size
= IP_VS_LBLCR_TAB_SIZE
*16;
578 * Hook periodic timer for garbage collection
580 init_timer(&tbl
->periodic_timer
);
581 tbl
->periodic_timer
.data
= (unsigned long)tbl
;
582 tbl
->periodic_timer
.function
= ip_vs_lblcr_check_expire
;
583 tbl
->periodic_timer
.expires
= jiffies
+CHECK_EXPIRE_INTERVAL
;
584 add_timer(&tbl
->periodic_timer
);
590 static int ip_vs_lblcr_done_svc(struct ip_vs_service
*svc
)
592 struct ip_vs_lblcr_table
*tbl
= svc
->sched_data
;
594 /* remove periodic timer */
595 del_timer_sync(&tbl
->periodic_timer
);
597 /* got to clean up table entries here */
598 ip_vs_lblcr_flush(tbl
);
600 /* release the table itself */
601 kfree(svc
->sched_data
);
602 IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) released\n",
603 sizeof(struct ip_vs_lblcr_table
));
609 static int ip_vs_lblcr_update_svc(struct ip_vs_service
*svc
)
615 static inline struct ip_vs_dest
*
616 __ip_vs_wlc_schedule(struct ip_vs_service
*svc
, struct iphdr
*iph
)
618 struct ip_vs_dest
*dest
, *least
;
622 * We think the overhead of processing active connections is fifty
623 * times higher than that of inactive connections in average. (This
624 * fifty times might not be accurate, we will change it later.) We
625 * use the following formula to estimate the overhead:
626 * dest->activeconns*50 + dest->inactconns
628 * (dest overhead) / dest->weight
630 * Remember -- no floats in kernel mode!!!
631 * The comparison of h1*w2 > h2*w1 is equivalent to that of
633 * if every weight is larger than zero.
635 * The server with weight=0 is quiesced and will not receive any
638 list_for_each_entry(dest
, &svc
->destinations
, n_list
) {
639 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
642 if (atomic_read(&dest
->weight
) > 0) {
644 loh
= atomic_read(&least
->activeconns
) * 50
645 + atomic_read(&least
->inactconns
);
652 * Find the destination with the least load.
655 list_for_each_entry_continue(dest
, &svc
->destinations
, n_list
) {
656 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
659 doh
= atomic_read(&dest
->activeconns
) * 50
660 + atomic_read(&dest
->inactconns
);
661 if (loh
* atomic_read(&dest
->weight
) >
662 doh
* atomic_read(&least
->weight
)) {
668 IP_VS_DBG(6, "LBLCR: server %d.%d.%d.%d:%d "
669 "activeconns %d refcnt %d weight %d overhead %d\n",
670 NIPQUAD(least
->addr
), ntohs(least
->port
),
671 atomic_read(&least
->activeconns
),
672 atomic_read(&least
->refcnt
),
673 atomic_read(&least
->weight
), loh
);
680 * If this destination server is overloaded and there is a less loaded
681 * server, then return true.
684 is_overloaded(struct ip_vs_dest
*dest
, struct ip_vs_service
*svc
)
686 if (atomic_read(&dest
->activeconns
) > atomic_read(&dest
->weight
)) {
687 struct ip_vs_dest
*d
;
689 list_for_each_entry(d
, &svc
->destinations
, n_list
) {
690 if (atomic_read(&d
->activeconns
)*2
691 < atomic_read(&d
->weight
)) {
701 * Locality-Based (weighted) Least-Connection scheduling
703 static struct ip_vs_dest
*
704 ip_vs_lblcr_schedule(struct ip_vs_service
*svc
, const struct sk_buff
*skb
)
706 struct ip_vs_dest
*dest
;
707 struct ip_vs_lblcr_table
*tbl
;
708 struct ip_vs_lblcr_entry
*en
;
709 struct iphdr
*iph
= ip_hdr(skb
);
711 IP_VS_DBG(6, "ip_vs_lblcr_schedule(): Scheduling...\n");
713 tbl
= (struct ip_vs_lblcr_table
*)svc
->sched_data
;
714 en
= ip_vs_lblcr_get(tbl
, iph
->daddr
);
716 dest
= __ip_vs_wlc_schedule(svc
, iph
);
718 IP_VS_DBG(1, "no destination available\n");
721 en
= ip_vs_lblcr_new(iph
->daddr
);
725 ip_vs_dest_set_insert(&en
->set
, dest
);
726 ip_vs_lblcr_hash(tbl
, en
);
728 dest
= ip_vs_dest_set_min(&en
->set
);
729 if (!dest
|| is_overloaded(dest
, svc
)) {
730 dest
= __ip_vs_wlc_schedule(svc
, iph
);
732 IP_VS_DBG(1, "no destination available\n");
735 ip_vs_dest_set_insert(&en
->set
, dest
);
737 if (atomic_read(&en
->set
.size
) > 1 &&
738 jiffies
-en
->set
.lastmod
> sysctl_ip_vs_lblcr_expiration
) {
739 struct ip_vs_dest
*m
;
740 m
= ip_vs_dest_set_max(&en
->set
);
742 ip_vs_dest_set_erase(&en
->set
, m
);
745 en
->lastuse
= jiffies
;
747 IP_VS_DBG(6, "LBLCR: destination IP address %u.%u.%u.%u "
748 "--> server %u.%u.%u.%u:%d\n",
758 * IPVS LBLCR Scheduler structure
760 static struct ip_vs_scheduler ip_vs_lblcr_scheduler
=
763 .refcnt
= ATOMIC_INIT(0),
764 .module
= THIS_MODULE
,
765 .init_service
= ip_vs_lblcr_init_svc
,
766 .done_service
= ip_vs_lblcr_done_svc
,
767 .update_service
= ip_vs_lblcr_update_svc
,
768 .schedule
= ip_vs_lblcr_schedule
,
772 static int __init
ip_vs_lblcr_init(void)
774 INIT_LIST_HEAD(&ip_vs_lblcr_scheduler
.n_list
);
775 sysctl_header
= register_sysctl_table(lblcr_root_table
);
776 return register_ip_vs_scheduler(&ip_vs_lblcr_scheduler
);
780 static void __exit
ip_vs_lblcr_cleanup(void)
782 unregister_sysctl_table(sysctl_header
);
783 unregister_ip_vs_scheduler(&ip_vs_lblcr_scheduler
);
787 module_init(ip_vs_lblcr_init
);
788 module_exit(ip_vs_lblcr_cleanup
);
789 MODULE_LICENSE("GPL");