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)
66 #define DEFAULT_EXPIRATION (24*60*60*HZ)
69 * It is for full expiration check.
70 * When there is no partial expiration check (garbage collection)
71 * in a half hour, do a full expiration check to collect stale
72 * entries that haven't been touched for a day.
74 #define COUNT_FOR_FULL_EXPIRATION 30
78 * for IPVS lblc entry hash table
80 #ifndef CONFIG_IP_VS_LBLC_TAB_BITS
81 #define CONFIG_IP_VS_LBLC_TAB_BITS 10
83 #define IP_VS_LBLC_TAB_BITS CONFIG_IP_VS_LBLC_TAB_BITS
84 #define IP_VS_LBLC_TAB_SIZE (1 << IP_VS_LBLC_TAB_BITS)
85 #define IP_VS_LBLC_TAB_MASK (IP_VS_LBLC_TAB_SIZE - 1)
89 * IPVS lblc entry represents an association between destination
90 * IP address and its destination server
92 struct ip_vs_lblc_entry
{
93 struct hlist_node list
;
94 int af
; /* address family */
95 union nf_inet_addr addr
; /* destination IP address */
96 struct ip_vs_dest
*dest
; /* real server (cache) */
97 unsigned long lastuse
; /* last used time */
98 struct rcu_head rcu_head
;
103 * IPVS lblc hash table
105 struct ip_vs_lblc_table
{
106 struct rcu_head rcu_head
;
107 struct hlist_head bucket
[IP_VS_LBLC_TAB_SIZE
]; /* hash bucket */
108 struct timer_list periodic_timer
; /* collect stale entries */
109 atomic_t entries
; /* number of entries */
110 int max_size
; /* maximum size of entries */
111 int rover
; /* rover for expire check */
112 int counter
; /* counter for no expire */
118 * IPVS LBLC sysctl table
121 static struct ctl_table vs_vars_table
[] = {
123 .procname
= "lblc_expiration",
125 .maxlen
= sizeof(int),
127 .proc_handler
= proc_dointvec_jiffies
,
133 static void ip_vs_lblc_rcu_free(struct rcu_head
*head
)
135 struct ip_vs_lblc_entry
*en
= container_of(head
,
136 struct ip_vs_lblc_entry
,
139 ip_vs_dest_put_and_free(en
->dest
);
143 static inline void ip_vs_lblc_del(struct ip_vs_lblc_entry
*en
)
145 hlist_del_rcu(&en
->list
);
146 call_rcu(&en
->rcu_head
, ip_vs_lblc_rcu_free
);
150 * Returns hash value for IPVS LBLC entry
152 static inline unsigned int
153 ip_vs_lblc_hashkey(int af
, const union nf_inet_addr
*addr
)
155 __be32 addr_fold
= addr
->ip
;
157 #ifdef CONFIG_IP_VS_IPV6
159 addr_fold
= addr
->ip6
[0]^addr
->ip6
[1]^
160 addr
->ip6
[2]^addr
->ip6
[3];
162 return (ntohl(addr_fold
)*2654435761UL) & IP_VS_LBLC_TAB_MASK
;
167 * Hash an entry in the ip_vs_lblc_table.
168 * returns bool success.
171 ip_vs_lblc_hash(struct ip_vs_lblc_table
*tbl
, struct ip_vs_lblc_entry
*en
)
173 unsigned int hash
= ip_vs_lblc_hashkey(en
->af
, &en
->addr
);
175 hlist_add_head_rcu(&en
->list
, &tbl
->bucket
[hash
]);
176 atomic_inc(&tbl
->entries
);
180 /* Get ip_vs_lblc_entry associated with supplied parameters. */
181 static inline struct ip_vs_lblc_entry
*
182 ip_vs_lblc_get(int af
, struct ip_vs_lblc_table
*tbl
,
183 const union nf_inet_addr
*addr
)
185 unsigned int hash
= ip_vs_lblc_hashkey(af
, addr
);
186 struct ip_vs_lblc_entry
*en
;
188 hlist_for_each_entry_rcu(en
, &tbl
->bucket
[hash
], list
)
189 if (ip_vs_addr_equal(af
, &en
->addr
, addr
))
197 * Create or update an ip_vs_lblc_entry, which is a mapping of a destination IP
198 * address to a server. Called under spin lock.
200 static inline struct ip_vs_lblc_entry
*
201 ip_vs_lblc_new(struct ip_vs_lblc_table
*tbl
, const union nf_inet_addr
*daddr
,
202 u16 af
, struct ip_vs_dest
*dest
)
204 struct ip_vs_lblc_entry
*en
;
206 en
= ip_vs_lblc_get(af
, tbl
, daddr
);
208 if (en
->dest
== dest
)
212 en
= kmalloc(sizeof(*en
), GFP_ATOMIC
);
217 ip_vs_addr_copy(af
, &en
->addr
, daddr
);
218 en
->lastuse
= jiffies
;
220 ip_vs_dest_hold(dest
);
223 ip_vs_lblc_hash(tbl
, en
);
230 * Flush all the entries of the specified table.
232 static void ip_vs_lblc_flush(struct ip_vs_service
*svc
)
234 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
235 struct ip_vs_lblc_entry
*en
;
236 struct hlist_node
*next
;
239 spin_lock_bh(&svc
->sched_lock
);
241 for (i
= 0; i
< IP_VS_LBLC_TAB_SIZE
; i
++) {
242 hlist_for_each_entry_safe(en
, next
, &tbl
->bucket
[i
], list
) {
244 atomic_dec(&tbl
->entries
);
247 spin_unlock_bh(&svc
->sched_lock
);
250 static int sysctl_lblc_expiration(struct ip_vs_service
*svc
)
253 return svc
->ipvs
->sysctl_lblc_expiration
;
255 return DEFAULT_EXPIRATION
;
259 static inline void ip_vs_lblc_full_check(struct ip_vs_service
*svc
)
261 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
262 struct ip_vs_lblc_entry
*en
;
263 struct hlist_node
*next
;
264 unsigned long now
= jiffies
;
267 for (i
= 0, j
= tbl
->rover
; i
< IP_VS_LBLC_TAB_SIZE
; i
++) {
268 j
= (j
+ 1) & IP_VS_LBLC_TAB_MASK
;
270 spin_lock(&svc
->sched_lock
);
271 hlist_for_each_entry_safe(en
, next
, &tbl
->bucket
[j
], list
) {
274 sysctl_lblc_expiration(svc
)))
278 atomic_dec(&tbl
->entries
);
280 spin_unlock(&svc
->sched_lock
);
287 * Periodical timer handler for IPVS lblc table
288 * It is used to collect stale entries when the number of entries
289 * exceeds the maximum size of the table.
291 * Fixme: we probably need more complicated algorithm to collect
292 * entries that have not been used for a long time even
293 * if the number of entries doesn't exceed the maximum size
295 * The full expiration check is for this purpose now.
297 static void ip_vs_lblc_check_expire(unsigned long data
)
299 struct ip_vs_service
*svc
= (struct ip_vs_service
*) data
;
300 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
301 unsigned long now
= jiffies
;
304 struct ip_vs_lblc_entry
*en
;
305 struct hlist_node
*next
;
307 if ((tbl
->counter
% COUNT_FOR_FULL_EXPIRATION
) == 0) {
308 /* do full expiration check */
309 ip_vs_lblc_full_check(svc
);
314 if (atomic_read(&tbl
->entries
) <= tbl
->max_size
) {
319 goal
= (atomic_read(&tbl
->entries
) - tbl
->max_size
)*4/3;
320 if (goal
> tbl
->max_size
/2)
321 goal
= tbl
->max_size
/2;
323 for (i
= 0, j
= tbl
->rover
; i
< IP_VS_LBLC_TAB_SIZE
; i
++) {
324 j
= (j
+ 1) & IP_VS_LBLC_TAB_MASK
;
326 spin_lock(&svc
->sched_lock
);
327 hlist_for_each_entry_safe(en
, next
, &tbl
->bucket
[j
], list
) {
328 if (time_before(now
, en
->lastuse
+ ENTRY_TIMEOUT
))
332 atomic_dec(&tbl
->entries
);
335 spin_unlock(&svc
->sched_lock
);
342 mod_timer(&tbl
->periodic_timer
, jiffies
+ CHECK_EXPIRE_INTERVAL
);
346 static int ip_vs_lblc_init_svc(struct ip_vs_service
*svc
)
349 struct ip_vs_lblc_table
*tbl
;
352 * Allocate the ip_vs_lblc_table for this service
354 tbl
= kmalloc(sizeof(*tbl
), GFP_KERNEL
);
358 svc
->sched_data
= tbl
;
359 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for "
360 "current service\n", sizeof(*tbl
));
363 * Initialize the hash buckets
365 for (i
= 0; i
< IP_VS_LBLC_TAB_SIZE
; i
++) {
366 INIT_HLIST_HEAD(&tbl
->bucket
[i
]);
368 tbl
->max_size
= IP_VS_LBLC_TAB_SIZE
*16;
374 * Hook periodic timer for garbage collection
376 setup_timer(&tbl
->periodic_timer
, ip_vs_lblc_check_expire
,
378 mod_timer(&tbl
->periodic_timer
, jiffies
+ CHECK_EXPIRE_INTERVAL
);
384 static void ip_vs_lblc_done_svc(struct ip_vs_service
*svc
)
386 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
388 /* remove periodic timer */
389 del_timer_sync(&tbl
->periodic_timer
);
391 /* got to clean up table entries here */
392 ip_vs_lblc_flush(svc
);
394 /* release the table itself */
395 kfree_rcu(tbl
, rcu_head
);
396 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n",
401 static inline struct ip_vs_dest
*
402 __ip_vs_lblc_schedule(struct ip_vs_service
*svc
)
404 struct ip_vs_dest
*dest
, *least
;
408 * We use the following formula to estimate the load:
409 * (dest overhead) / dest->weight
411 * Remember -- no floats in kernel mode!!!
412 * The comparison of h1*w2 > h2*w1 is equivalent to that of
414 * if every weight is larger than zero.
416 * The server with weight=0 is quiesced and will not receive any
419 list_for_each_entry_rcu(dest
, &svc
->destinations
, n_list
) {
420 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
422 if (atomic_read(&dest
->weight
) > 0) {
424 loh
= ip_vs_dest_conn_overhead(least
);
431 * Find the destination with the least load.
434 list_for_each_entry_continue_rcu(dest
, &svc
->destinations
, n_list
) {
435 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
438 doh
= ip_vs_dest_conn_overhead(dest
);
439 if ((__s64
)loh
* atomic_read(&dest
->weight
) >
440 (__s64
)doh
* atomic_read(&least
->weight
)) {
446 IP_VS_DBG_BUF(6, "LBLC: server %s:%d "
447 "activeconns %d refcnt %d weight %d overhead %d\n",
448 IP_VS_DBG_ADDR(least
->af
, &least
->addr
),
450 atomic_read(&least
->activeconns
),
451 atomic_read(&least
->refcnt
),
452 atomic_read(&least
->weight
), loh
);
459 * If this destination server is overloaded and there is a less loaded
460 * server, then return true.
463 is_overloaded(struct ip_vs_dest
*dest
, struct ip_vs_service
*svc
)
465 if (atomic_read(&dest
->activeconns
) > atomic_read(&dest
->weight
)) {
466 struct ip_vs_dest
*d
;
468 list_for_each_entry_rcu(d
, &svc
->destinations
, n_list
) {
469 if (atomic_read(&d
->activeconns
)*2
470 < atomic_read(&d
->weight
)) {
480 * Locality-Based (weighted) Least-Connection scheduling
482 static struct ip_vs_dest
*
483 ip_vs_lblc_schedule(struct ip_vs_service
*svc
, const struct sk_buff
*skb
,
484 struct ip_vs_iphdr
*iph
)
486 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
487 struct ip_vs_dest
*dest
= NULL
;
488 struct ip_vs_lblc_entry
*en
;
490 IP_VS_DBG(6, "%s(): Scheduling...\n", __func__
);
492 /* First look in our cache */
493 en
= ip_vs_lblc_get(svc
->af
, tbl
, &iph
->daddr
);
495 /* We only hold a read lock, but this is atomic */
496 en
->lastuse
= jiffies
;
499 * If the destination is not available, i.e. it's in the trash,
500 * we must ignore it, as it may be removed from under our feet,
501 * if someone drops our reference count. Our caller only makes
502 * sure that destinations, that are not in the trash, are not
503 * moved to the trash, while we are scheduling. But anyone can
504 * free up entries from the trash at any time.
508 if ((dest
->flags
& IP_VS_DEST_F_AVAILABLE
) &&
509 atomic_read(&dest
->weight
) > 0 && !is_overloaded(dest
, svc
))
513 /* No cache entry or it is invalid, time to schedule */
514 dest
= __ip_vs_lblc_schedule(svc
);
516 ip_vs_scheduler_err(svc
, "no destination available");
520 /* If we fail to create a cache entry, we'll just use the valid dest */
521 spin_lock_bh(&svc
->sched_lock
);
523 ip_vs_lblc_new(tbl
, &iph
->daddr
, svc
->af
, dest
);
524 spin_unlock_bh(&svc
->sched_lock
);
527 IP_VS_DBG_BUF(6, "LBLC: destination IP address %s --> server %s:%d\n",
528 IP_VS_DBG_ADDR(svc
->af
, &iph
->daddr
),
529 IP_VS_DBG_ADDR(dest
->af
, &dest
->addr
), ntohs(dest
->port
));
536 * IPVS LBLC Scheduler structure
538 static struct ip_vs_scheduler ip_vs_lblc_scheduler
= {
540 .refcnt
= ATOMIC_INIT(0),
541 .module
= THIS_MODULE
,
542 .n_list
= LIST_HEAD_INIT(ip_vs_lblc_scheduler
.n_list
),
543 .init_service
= ip_vs_lblc_init_svc
,
544 .done_service
= ip_vs_lblc_done_svc
,
545 .schedule
= ip_vs_lblc_schedule
,
552 static int __net_init
__ip_vs_lblc_init(struct net
*net
)
554 struct netns_ipvs
*ipvs
= net_ipvs(net
);
559 if (!net_eq(net
, &init_net
)) {
560 ipvs
->lblc_ctl_table
= kmemdup(vs_vars_table
,
561 sizeof(vs_vars_table
),
563 if (ipvs
->lblc_ctl_table
== NULL
)
566 /* Don't export sysctls to unprivileged users */
567 if (net
->user_ns
!= &init_user_ns
)
568 ipvs
->lblc_ctl_table
[0].procname
= NULL
;
571 ipvs
->lblc_ctl_table
= vs_vars_table
;
572 ipvs
->sysctl_lblc_expiration
= DEFAULT_EXPIRATION
;
573 ipvs
->lblc_ctl_table
[0].data
= &ipvs
->sysctl_lblc_expiration
;
575 ipvs
->lblc_ctl_header
=
576 register_net_sysctl(net
, "net/ipv4/vs", ipvs
->lblc_ctl_table
);
577 if (!ipvs
->lblc_ctl_header
) {
578 if (!net_eq(net
, &init_net
))
579 kfree(ipvs
->lblc_ctl_table
);
586 static void __net_exit
__ip_vs_lblc_exit(struct net
*net
)
588 struct netns_ipvs
*ipvs
= net_ipvs(net
);
590 unregister_net_sysctl_table(ipvs
->lblc_ctl_header
);
592 if (!net_eq(net
, &init_net
))
593 kfree(ipvs
->lblc_ctl_table
);
598 static int __net_init
__ip_vs_lblc_init(struct net
*net
) { return 0; }
599 static void __net_exit
__ip_vs_lblc_exit(struct net
*net
) { }
603 static struct pernet_operations ip_vs_lblc_ops
= {
604 .init
= __ip_vs_lblc_init
,
605 .exit
= __ip_vs_lblc_exit
,
608 static int __init
ip_vs_lblc_init(void)
612 ret
= register_pernet_subsys(&ip_vs_lblc_ops
);
616 ret
= register_ip_vs_scheduler(&ip_vs_lblc_scheduler
);
618 unregister_pernet_subsys(&ip_vs_lblc_ops
);
622 static void __exit
ip_vs_lblc_cleanup(void)
624 unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler
);
625 unregister_pernet_subsys(&ip_vs_lblc_ops
);
630 module_init(ip_vs_lblc_init
);
631 module_exit(ip_vs_lblc_cleanup
);
632 MODULE_LICENSE("GPL");