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/module.h>
47 #include <linux/kernel.h>
48 #include <linux/skbuff.h>
49 #include <linux/jiffies.h>
53 #include <linux/sysctl.h>
55 #include <net/ip_vs.h>
59 * It is for garbage collection of stale IPVS lblc entries,
60 * when the table is full.
62 #define CHECK_EXPIRE_INTERVAL (60*HZ)
63 #define ENTRY_TIMEOUT (6*60*HZ)
66 * It is for full expiration check.
67 * When there is no partial expiration check (garbage collection)
68 * in a half hour, do a full expiration check to collect stale
69 * entries that haven't been touched for a day.
71 #define COUNT_FOR_FULL_EXPIRATION 30
72 static int sysctl_ip_vs_lblc_expiration
= 24*60*60*HZ
;
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",
119 .data
= &sysctl_ip_vs_lblc_expiration
,
120 .maxlen
= sizeof(int),
122 .proc_handler
= proc_dointvec_jiffies
,
127 static struct ctl_table_header
* sysctl_header
;
129 static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry
*en
)
133 * We don't kfree dest because it is refered either by its service
134 * or the trash dest list.
136 atomic_dec(&en
->dest
->refcnt
);
142 * Returns hash value for IPVS LBLC entry
144 static inline unsigned
145 ip_vs_lblc_hashkey(int af
, const union nf_inet_addr
*addr
)
147 __be32 addr_fold
= addr
->ip
;
149 #ifdef CONFIG_IP_VS_IPV6
151 addr_fold
= addr
->ip6
[0]^addr
->ip6
[1]^
152 addr
->ip6
[2]^addr
->ip6
[3];
154 return (ntohl(addr_fold
)*2654435761UL) & IP_VS_LBLC_TAB_MASK
;
159 * Hash an entry in the ip_vs_lblc_table.
160 * returns bool success.
163 ip_vs_lblc_hash(struct ip_vs_lblc_table
*tbl
, struct ip_vs_lblc_entry
*en
)
165 unsigned hash
= ip_vs_lblc_hashkey(en
->af
, &en
->addr
);
167 list_add(&en
->list
, &tbl
->bucket
[hash
]);
168 atomic_inc(&tbl
->entries
);
173 * Get ip_vs_lblc_entry associated with supplied parameters. Called under read
176 static inline struct ip_vs_lblc_entry
*
177 ip_vs_lblc_get(int af
, struct ip_vs_lblc_table
*tbl
,
178 const union nf_inet_addr
*addr
)
180 unsigned hash
= ip_vs_lblc_hashkey(af
, addr
);
181 struct ip_vs_lblc_entry
*en
;
183 list_for_each_entry(en
, &tbl
->bucket
[hash
], list
)
184 if (ip_vs_addr_equal(af
, &en
->addr
, addr
))
192 * Create or update an ip_vs_lblc_entry, which is a mapping of a destination IP
193 * address to a server. Called under write lock.
195 static inline struct ip_vs_lblc_entry
*
196 ip_vs_lblc_new(struct ip_vs_lblc_table
*tbl
, const union nf_inet_addr
*daddr
,
197 struct ip_vs_dest
*dest
)
199 struct ip_vs_lblc_entry
*en
;
201 en
= ip_vs_lblc_get(dest
->af
, tbl
, daddr
);
203 en
= kmalloc(sizeof(*en
), GFP_ATOMIC
);
205 pr_err("%s(): no memory\n", __func__
);
210 ip_vs_addr_copy(dest
->af
, &en
->addr
, daddr
);
211 en
->lastuse
= jiffies
;
213 atomic_inc(&dest
->refcnt
);
216 ip_vs_lblc_hash(tbl
, en
);
217 } else if (en
->dest
!= dest
) {
218 atomic_dec(&en
->dest
->refcnt
);
219 atomic_inc(&dest
->refcnt
);
228 * Flush all the entries of the specified table.
230 static void ip_vs_lblc_flush(struct ip_vs_lblc_table
*tbl
)
232 struct ip_vs_lblc_entry
*en
, *nxt
;
235 for (i
=0; i
<IP_VS_LBLC_TAB_SIZE
; i
++) {
236 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[i
], list
) {
238 atomic_dec(&tbl
->entries
);
244 static inline void ip_vs_lblc_full_check(struct ip_vs_service
*svc
)
246 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
247 struct ip_vs_lblc_entry
*en
, *nxt
;
248 unsigned long now
= jiffies
;
251 for (i
=0, j
=tbl
->rover
; i
<IP_VS_LBLC_TAB_SIZE
; i
++) {
252 j
= (j
+ 1) & IP_VS_LBLC_TAB_MASK
;
254 write_lock(&svc
->sched_lock
);
255 list_for_each_entry_safe(en
, nxt
, &tbl
->bucket
[j
], list
) {
257 en
->lastuse
+ sysctl_ip_vs_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 think the overhead of processing active connections is fifty
393 * times higher than that of inactive connections in average. (This
394 * fifty times might not be accurate, we will change it later.) We
395 * use the following formula to estimate the overhead:
396 * dest->activeconns*50 + dest->inactconns
398 * (dest overhead) / dest->weight
400 * Remember -- no floats in kernel mode!!!
401 * The comparison of h1*w2 > h2*w1 is equivalent to that of
403 * if every weight is larger than zero.
405 * The server with weight=0 is quiesced and will not receive any
408 list_for_each_entry(dest
, &svc
->destinations
, n_list
) {
409 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
411 if (atomic_read(&dest
->weight
) > 0) {
413 loh
= atomic_read(&least
->activeconns
) * 50
414 + atomic_read(&least
->inactconns
);
421 * Find the destination with the least load.
424 list_for_each_entry_continue(dest
, &svc
->destinations
, n_list
) {
425 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
428 doh
= atomic_read(&dest
->activeconns
) * 50
429 + atomic_read(&dest
->inactconns
);
430 if (loh
* atomic_read(&dest
->weight
) >
431 doh
* atomic_read(&least
->weight
)) {
437 IP_VS_DBG_BUF(6, "LBLC: server %s:%d "
438 "activeconns %d refcnt %d weight %d overhead %d\n",
439 IP_VS_DBG_ADDR(least
->af
, &least
->addr
),
441 atomic_read(&least
->activeconns
),
442 atomic_read(&least
->refcnt
),
443 atomic_read(&least
->weight
), loh
);
450 * If this destination server is overloaded and there is a less loaded
451 * server, then return true.
454 is_overloaded(struct ip_vs_dest
*dest
, struct ip_vs_service
*svc
)
456 if (atomic_read(&dest
->activeconns
) > atomic_read(&dest
->weight
)) {
457 struct ip_vs_dest
*d
;
459 list_for_each_entry(d
, &svc
->destinations
, n_list
) {
460 if (atomic_read(&d
->activeconns
)*2
461 < atomic_read(&d
->weight
)) {
471 * Locality-Based (weighted) Least-Connection scheduling
473 static struct ip_vs_dest
*
474 ip_vs_lblc_schedule(struct ip_vs_service
*svc
, const struct sk_buff
*skb
)
476 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
477 struct ip_vs_iphdr iph
;
478 struct ip_vs_dest
*dest
= NULL
;
479 struct ip_vs_lblc_entry
*en
;
481 ip_vs_fill_iphdr(svc
->af
, skb_network_header(skb
), &iph
);
483 IP_VS_DBG(6, "%s(): Scheduling...\n", __func__
);
485 /* First look in our cache */
486 read_lock(&svc
->sched_lock
);
487 en
= ip_vs_lblc_get(svc
->af
, tbl
, &iph
.daddr
);
489 /* We only hold a read lock, but this is atomic */
490 en
->lastuse
= jiffies
;
493 * If the destination is not available, i.e. it's in the trash,
494 * we must ignore it, as it may be removed from under our feet,
495 * if someone drops our reference count. Our caller only makes
496 * sure that destinations, that are not in the trash, are not
497 * moved to the trash, while we are scheduling. But anyone can
498 * free up entries from the trash at any time.
501 if (en
->dest
->flags
& IP_VS_DEST_F_AVAILABLE
)
504 read_unlock(&svc
->sched_lock
);
506 /* If the destination has a weight and is not overloaded, use it */
507 if (dest
&& atomic_read(&dest
->weight
) > 0 && !is_overloaded(dest
, svc
))
510 /* No cache entry or it is invalid, time to schedule */
511 dest
= __ip_vs_lblc_schedule(svc
);
513 IP_VS_ERR_RL("LBLC: no destination available\n");
517 /* If we fail to create a cache entry, we'll just use the valid dest */
518 write_lock(&svc
->sched_lock
);
519 ip_vs_lblc_new(tbl
, &iph
.daddr
, dest
);
520 write_unlock(&svc
->sched_lock
);
523 IP_VS_DBG_BUF(6, "LBLC: destination IP address %s --> server %s:%d\n",
524 IP_VS_DBG_ADDR(svc
->af
, &iph
.daddr
),
525 IP_VS_DBG_ADDR(svc
->af
, &dest
->addr
), ntohs(dest
->port
));
532 * IPVS LBLC Scheduler structure
534 static struct ip_vs_scheduler ip_vs_lblc_scheduler
=
537 .refcnt
= ATOMIC_INIT(0),
538 .module
= THIS_MODULE
,
539 .n_list
= LIST_HEAD_INIT(ip_vs_lblc_scheduler
.n_list
),
540 .init_service
= ip_vs_lblc_init_svc
,
541 .done_service
= ip_vs_lblc_done_svc
,
542 .schedule
= ip_vs_lblc_schedule
,
546 static int __init
ip_vs_lblc_init(void)
550 sysctl_header
= register_sysctl_paths(net_vs_ctl_path
, vs_vars_table
);
551 ret
= register_ip_vs_scheduler(&ip_vs_lblc_scheduler
);
553 unregister_sysctl_table(sysctl_header
);
558 static void __exit
ip_vs_lblc_cleanup(void)
560 unregister_sysctl_table(sysctl_header
);
561 unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler
);
565 module_init(ip_vs_lblc_init
);
566 module_exit(ip_vs_lblc_cleanup
);
567 MODULE_LICENSE("GPL");