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)
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
76 * for IPVS lblc entry hash table
78 #ifndef CONFIG_IP_VS_LBLC_TAB_BITS
79 #define CONFIG_IP_VS_LBLC_TAB_BITS 10
81 #define IP_VS_LBLC_TAB_BITS CONFIG_IP_VS_LBLC_TAB_BITS
82 #define IP_VS_LBLC_TAB_SIZE (1 << IP_VS_LBLC_TAB_BITS)
83 #define IP_VS_LBLC_TAB_MASK (IP_VS_LBLC_TAB_SIZE - 1)
87 * IPVS lblc entry represents an association between destination
88 * IP address and its destination server
90 struct ip_vs_lblc_entry
{
91 struct list_head list
;
92 int af
; /* address family */
93 union nf_inet_addr addr
; /* destination IP address */
94 struct ip_vs_dest
*dest
; /* real server (cache) */
95 unsigned long lastuse
; /* last used time */
100 * IPVS lblc hash table
102 struct ip_vs_lblc_table
{
103 struct list_head bucket
[IP_VS_LBLC_TAB_SIZE
]; /* hash bucket */
104 atomic_t entries
; /* number of entries */
105 int max_size
; /* maximum size of entries */
106 struct timer_list periodic_timer
; /* collect stale entries */
107 int rover
; /* rover for expire check */
108 int counter
; /* counter for no expire */
113 * IPVS LBLC sysctl table
116 static ctl_table vs_vars_table
[] = {
118 .procname
= "lblc_expiration",
120 .maxlen
= sizeof(int),
122 .proc_handler
= proc_dointvec_jiffies
,
127 static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry
*en
)
131 * We don't kfree dest because it is refered either by its service
132 * or the trash dest list.
134 atomic_dec(&en
->dest
->refcnt
);
140 * Returns hash value for IPVS LBLC entry
142 static inline unsigned
143 ip_vs_lblc_hashkey(int af
, const union nf_inet_addr
*addr
)
145 __be32 addr_fold
= addr
->ip
;
147 #ifdef CONFIG_IP_VS_IPV6
149 addr_fold
= addr
->ip6
[0]^addr
->ip6
[1]^
150 addr
->ip6
[2]^addr
->ip6
[3];
152 return (ntohl(addr_fold
)*2654435761UL) & IP_VS_LBLC_TAB_MASK
;
157 * Hash an entry in the ip_vs_lblc_table.
158 * returns bool success.
161 ip_vs_lblc_hash(struct ip_vs_lblc_table
*tbl
, struct ip_vs_lblc_entry
*en
)
163 unsigned hash
= ip_vs_lblc_hashkey(en
->af
, &en
->addr
);
165 list_add(&en
->list
, &tbl
->bucket
[hash
]);
166 atomic_inc(&tbl
->entries
);
171 * Get ip_vs_lblc_entry associated with supplied parameters. Called under read
174 static inline struct ip_vs_lblc_entry
*
175 ip_vs_lblc_get(int af
, struct ip_vs_lblc_table
*tbl
,
176 const union nf_inet_addr
*addr
)
178 unsigned hash
= ip_vs_lblc_hashkey(af
, addr
);
179 struct ip_vs_lblc_entry
*en
;
181 list_for_each_entry(en
, &tbl
->bucket
[hash
], list
)
182 if (ip_vs_addr_equal(af
, &en
->addr
, addr
))
190 * Create or update an ip_vs_lblc_entry, which is a mapping of a destination IP
191 * address to a server. Called under write lock.
193 static inline struct ip_vs_lblc_entry
*
194 ip_vs_lblc_new(struct ip_vs_lblc_table
*tbl
, const union nf_inet_addr
*daddr
,
195 struct ip_vs_dest
*dest
)
197 struct ip_vs_lblc_entry
*en
;
199 en
= ip_vs_lblc_get(dest
->af
, tbl
, daddr
);
201 en
= kmalloc(sizeof(*en
), GFP_ATOMIC
);
203 pr_err("%s(): no memory\n", __func__
);
208 ip_vs_addr_copy(dest
->af
, &en
->addr
, daddr
);
209 en
->lastuse
= jiffies
;
211 atomic_inc(&dest
->refcnt
);
214 ip_vs_lblc_hash(tbl
, en
);
215 } else if (en
->dest
!= dest
) {
216 atomic_dec(&en
->dest
->refcnt
);
217 atomic_inc(&dest
->refcnt
);
226 * Flush all the entries of the specified table.
228 static void ip_vs_lblc_flush(struct ip_vs_lblc_table
*tbl
)
230 struct ip_vs_lblc_entry
*en
, *nxt
;
233 for (i
=0; i
<IP_VS_LBLC_TAB_SIZE
; i
++) {
234 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[i
], list
) {
236 atomic_dec(&tbl
->entries
);
242 static inline void ip_vs_lblc_full_check(struct ip_vs_service
*svc
)
244 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
245 struct ip_vs_lblc_entry
*en
, *nxt
;
246 unsigned long now
= jiffies
;
248 struct netns_ipvs
*ipvs
= net_ipvs(svc
->net
);
250 for (i
=0, j
=tbl
->rover
; i
<IP_VS_LBLC_TAB_SIZE
; i
++) {
251 j
= (j
+ 1) & IP_VS_LBLC_TAB_MASK
;
253 write_lock(&svc
->sched_lock
);
254 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[j
], list
) {
257 ipvs
->sysctl_lblc_expiration
))
261 atomic_dec(&tbl
->entries
);
263 write_unlock(&svc
->sched_lock
);
270 * Periodical timer handler for IPVS lblc table
271 * It is used to collect stale entries when the number of entries
272 * exceeds the maximum size of the table.
274 * Fixme: we probably need more complicated algorithm to collect
275 * entries that have not been used for a long time even
276 * if the number of entries doesn't exceed the maximum size
278 * The full expiration check is for this purpose now.
280 static void ip_vs_lblc_check_expire(unsigned long data
)
282 struct ip_vs_service
*svc
= (struct ip_vs_service
*) data
;
283 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
284 unsigned long now
= jiffies
;
287 struct ip_vs_lblc_entry
*en
, *nxt
;
289 if ((tbl
->counter
% COUNT_FOR_FULL_EXPIRATION
) == 0) {
290 /* do full expiration check */
291 ip_vs_lblc_full_check(svc
);
296 if (atomic_read(&tbl
->entries
) <= tbl
->max_size
) {
301 goal
= (atomic_read(&tbl
->entries
) - tbl
->max_size
)*4/3;
302 if (goal
> tbl
->max_size
/2)
303 goal
= tbl
->max_size
/2;
305 for (i
=0, j
=tbl
->rover
; i
<IP_VS_LBLC_TAB_SIZE
; i
++) {
306 j
= (j
+ 1) & IP_VS_LBLC_TAB_MASK
;
308 write_lock(&svc
->sched_lock
);
309 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[j
], list
) {
310 if (time_before(now
, en
->lastuse
+ ENTRY_TIMEOUT
))
314 atomic_dec(&tbl
->entries
);
317 write_unlock(&svc
->sched_lock
);
324 mod_timer(&tbl
->periodic_timer
, jiffies
+CHECK_EXPIRE_INTERVAL
);
328 static int ip_vs_lblc_init_svc(struct ip_vs_service
*svc
)
331 struct ip_vs_lblc_table
*tbl
;
334 * Allocate the ip_vs_lblc_table for this service
336 tbl
= kmalloc(sizeof(*tbl
), GFP_ATOMIC
);
338 pr_err("%s(): no memory\n", __func__
);
341 svc
->sched_data
= tbl
;
342 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for "
343 "current service\n", sizeof(*tbl
));
346 * Initialize the hash buckets
348 for (i
=0; i
<IP_VS_LBLC_TAB_SIZE
; i
++) {
349 INIT_LIST_HEAD(&tbl
->bucket
[i
]);
351 tbl
->max_size
= IP_VS_LBLC_TAB_SIZE
*16;
356 * Hook periodic timer for garbage collection
358 setup_timer(&tbl
->periodic_timer
, ip_vs_lblc_check_expire
,
360 mod_timer(&tbl
->periodic_timer
, jiffies
+ CHECK_EXPIRE_INTERVAL
);
366 static int ip_vs_lblc_done_svc(struct ip_vs_service
*svc
)
368 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
370 /* remove periodic timer */
371 del_timer_sync(&tbl
->periodic_timer
);
373 /* got to clean up table entries here */
374 ip_vs_lblc_flush(tbl
);
376 /* release the table itself */
378 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n",
385 static inline struct ip_vs_dest
*
386 __ip_vs_lblc_schedule(struct ip_vs_service
*svc
)
388 struct ip_vs_dest
*dest
, *least
;
392 * We use the following formula to estimate the load:
393 * (dest overhead) / dest->weight
395 * Remember -- no floats in kernel mode!!!
396 * The comparison of h1*w2 > h2*w1 is equivalent to that of
398 * if every weight is larger than zero.
400 * The server with weight=0 is quiesced and will not receive any
403 list_for_each_entry(dest
, &svc
->destinations
, n_list
) {
404 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
406 if (atomic_read(&dest
->weight
) > 0) {
408 loh
= ip_vs_dest_conn_overhead(least
);
415 * Find the destination with the least load.
418 list_for_each_entry_continue(dest
, &svc
->destinations
, n_list
) {
419 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
422 doh
= ip_vs_dest_conn_overhead(dest
);
423 if (loh
* atomic_read(&dest
->weight
) >
424 doh
* atomic_read(&least
->weight
)) {
430 IP_VS_DBG_BUF(6, "LBLC: server %s:%d "
431 "activeconns %d refcnt %d weight %d overhead %d\n",
432 IP_VS_DBG_ADDR(least
->af
, &least
->addr
),
434 atomic_read(&least
->activeconns
),
435 atomic_read(&least
->refcnt
),
436 atomic_read(&least
->weight
), loh
);
443 * If this destination server is overloaded and there is a less loaded
444 * server, then return true.
447 is_overloaded(struct ip_vs_dest
*dest
, struct ip_vs_service
*svc
)
449 if (atomic_read(&dest
->activeconns
) > atomic_read(&dest
->weight
)) {
450 struct ip_vs_dest
*d
;
452 list_for_each_entry(d
, &svc
->destinations
, n_list
) {
453 if (atomic_read(&d
->activeconns
)*2
454 < atomic_read(&d
->weight
)) {
464 * Locality-Based (weighted) Least-Connection scheduling
466 static struct ip_vs_dest
*
467 ip_vs_lblc_schedule(struct ip_vs_service
*svc
, const struct sk_buff
*skb
)
469 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
470 struct ip_vs_iphdr iph
;
471 struct ip_vs_dest
*dest
= NULL
;
472 struct ip_vs_lblc_entry
*en
;
474 ip_vs_fill_iphdr(svc
->af
, skb_network_header(skb
), &iph
);
476 IP_VS_DBG(6, "%s(): Scheduling...\n", __func__
);
478 /* First look in our cache */
479 read_lock(&svc
->sched_lock
);
480 en
= ip_vs_lblc_get(svc
->af
, tbl
, &iph
.daddr
);
482 /* We only hold a read lock, but this is atomic */
483 en
->lastuse
= jiffies
;
486 * If the destination is not available, i.e. it's in the trash,
487 * we must ignore it, as it may be removed from under our feet,
488 * if someone drops our reference count. Our caller only makes
489 * sure that destinations, that are not in the trash, are not
490 * moved to the trash, while we are scheduling. But anyone can
491 * free up entries from the trash at any time.
494 if (en
->dest
->flags
& IP_VS_DEST_F_AVAILABLE
)
497 read_unlock(&svc
->sched_lock
);
499 /* If the destination has a weight and is not overloaded, use it */
500 if (dest
&& atomic_read(&dest
->weight
) > 0 && !is_overloaded(dest
, svc
))
503 /* No cache entry or it is invalid, time to schedule */
504 dest
= __ip_vs_lblc_schedule(svc
);
506 ip_vs_scheduler_err(svc
, "no destination available");
510 /* If we fail to create a cache entry, we'll just use the valid dest */
511 write_lock(&svc
->sched_lock
);
512 ip_vs_lblc_new(tbl
, &iph
.daddr
, dest
);
513 write_unlock(&svc
->sched_lock
);
516 IP_VS_DBG_BUF(6, "LBLC: destination IP address %s --> server %s:%d\n",
517 IP_VS_DBG_ADDR(svc
->af
, &iph
.daddr
),
518 IP_VS_DBG_ADDR(svc
->af
, &dest
->addr
), ntohs(dest
->port
));
525 * IPVS LBLC Scheduler structure
527 static struct ip_vs_scheduler ip_vs_lblc_scheduler
=
530 .refcnt
= ATOMIC_INIT(0),
531 .module
= THIS_MODULE
,
532 .n_list
= LIST_HEAD_INIT(ip_vs_lblc_scheduler
.n_list
),
533 .init_service
= ip_vs_lblc_init_svc
,
534 .done_service
= ip_vs_lblc_done_svc
,
535 .schedule
= ip_vs_lblc_schedule
,
541 static int __net_init
__ip_vs_lblc_init(struct net
*net
)
543 struct netns_ipvs
*ipvs
= net_ipvs(net
);
545 if (!net_eq(net
, &init_net
)) {
546 ipvs
->lblc_ctl_table
= kmemdup(vs_vars_table
,
547 sizeof(vs_vars_table
),
549 if (ipvs
->lblc_ctl_table
== NULL
)
552 ipvs
->lblc_ctl_table
= vs_vars_table
;
553 ipvs
->sysctl_lblc_expiration
= 24*60*60*HZ
;
554 ipvs
->lblc_ctl_table
[0].data
= &ipvs
->sysctl_lblc_expiration
;
557 ipvs
->lblc_ctl_header
=
558 register_net_sysctl_table(net
, net_vs_ctl_path
,
559 ipvs
->lblc_ctl_table
);
560 if (!ipvs
->lblc_ctl_header
) {
561 if (!net_eq(net
, &init_net
))
562 kfree(ipvs
->lblc_ctl_table
);
570 static void __net_exit
__ip_vs_lblc_exit(struct net
*net
)
572 struct netns_ipvs
*ipvs
= net_ipvs(net
);
575 unregister_net_sysctl_table(ipvs
->lblc_ctl_header
);
578 if (!net_eq(net
, &init_net
))
579 kfree(ipvs
->lblc_ctl_table
);
582 static struct pernet_operations ip_vs_lblc_ops
= {
583 .init
= __ip_vs_lblc_init
,
584 .exit
= __ip_vs_lblc_exit
,
587 static int __init
ip_vs_lblc_init(void)
591 ret
= register_pernet_subsys(&ip_vs_lblc_ops
);
595 ret
= register_ip_vs_scheduler(&ip_vs_lblc_scheduler
);
597 unregister_pernet_subsys(&ip_vs_lblc_ops
);
601 static void __exit
ip_vs_lblc_cleanup(void)
603 unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler
);
604 unregister_pernet_subsys(&ip_vs_lblc_ops
);
608 module_init(ip_vs_lblc_init
);
609 module_exit(ip_vs_lblc_cleanup
);
610 MODULE_LICENSE("GPL");