1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * IPVS: Locality-Based Least-Connection scheduling module
5 * Authors: Wensong Zhang <wensong@gnuchina.org>
8 * Martin Hamilton : fixed the terrible locking bugs
9 * *lock(tbl->lock) ==> *lock(&tbl->lock)
10 * Wensong Zhang : fixed the uninitialized tbl->lock bug
11 * Wensong Zhang : added doing full expiration check to
12 * collect stale entries of 24+ hours when
13 * no partial expire check in a half hour
14 * Julian Anastasov : replaced del_timer call with del_timer_sync
15 * to avoid the possible race between timer
16 * handler and del_timer thread in SMP
20 * The lblc algorithm is as follows (pseudo code):
22 * if cachenode[dest_ip] is null then
23 * n, cachenode[dest_ip] <- {weighted least-conn node};
25 * n <- cachenode[dest_ip];
27 * (n.conns>n.weight AND
28 * there is a node m with m.conns<m.weight/2) then
29 * n, cachenode[dest_ip] <- {weighted least-conn node};
33 * Thanks must go to Wenzhuo Zhang for talking WCCP to me and pushing
34 * me to write this module.
37 #define KMSG_COMPONENT "IPVS"
38 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
41 #include <linux/slab.h>
42 #include <linux/module.h>
43 #include <linux/kernel.h>
44 #include <linux/skbuff.h>
45 #include <linux/jiffies.h>
46 #include <linux/hash.h>
50 #include <linux/sysctl.h>
52 #include <net/ip_vs.h>
56 * It is for garbage collection of stale IPVS lblc entries,
57 * when the table is full.
59 #define CHECK_EXPIRE_INTERVAL (60*HZ)
60 #define ENTRY_TIMEOUT (6*60*HZ)
62 #define DEFAULT_EXPIRATION (24*60*60*HZ)
65 * It is for full expiration check.
66 * When there is no partial expiration check (garbage collection)
67 * in a half hour, do a full expiration check to collect stale
68 * entries that haven't been touched for a day.
70 #define COUNT_FOR_FULL_EXPIRATION 30
74 * for IPVS lblc entry hash table
76 #ifndef CONFIG_IP_VS_LBLC_TAB_BITS
77 #define CONFIG_IP_VS_LBLC_TAB_BITS 10
79 #define IP_VS_LBLC_TAB_BITS CONFIG_IP_VS_LBLC_TAB_BITS
80 #define IP_VS_LBLC_TAB_SIZE (1 << IP_VS_LBLC_TAB_BITS)
81 #define IP_VS_LBLC_TAB_MASK (IP_VS_LBLC_TAB_SIZE - 1)
85 * IPVS lblc entry represents an association between destination
86 * IP address and its destination server
88 struct ip_vs_lblc_entry
{
89 struct hlist_node list
;
90 int af
; /* address family */
91 union nf_inet_addr addr
; /* destination IP address */
92 struct ip_vs_dest
*dest
; /* real server (cache) */
93 unsigned long lastuse
; /* last used time */
94 struct rcu_head rcu_head
;
99 * IPVS lblc hash table
101 struct ip_vs_lblc_table
{
102 struct rcu_head rcu_head
;
103 struct hlist_head bucket
[IP_VS_LBLC_TAB_SIZE
]; /* hash bucket */
104 struct timer_list periodic_timer
; /* collect stale entries */
105 struct ip_vs_service
*svc
; /* pointer back to service */
106 atomic_t entries
; /* number of entries */
107 int max_size
; /* maximum size of entries */
108 int rover
; /* rover for expire check */
109 int counter
; /* counter for no expire */
115 * IPVS LBLC sysctl table
118 static struct ctl_table vs_vars_table
[] = {
120 .procname
= "lblc_expiration",
122 .maxlen
= sizeof(int),
124 .proc_handler
= proc_dointvec_jiffies
,
130 static void ip_vs_lblc_rcu_free(struct rcu_head
*head
)
132 struct ip_vs_lblc_entry
*en
= container_of(head
,
133 struct ip_vs_lblc_entry
,
136 ip_vs_dest_put_and_free(en
->dest
);
140 static inline void ip_vs_lblc_del(struct ip_vs_lblc_entry
*en
)
142 hlist_del_rcu(&en
->list
);
143 call_rcu(&en
->rcu_head
, ip_vs_lblc_rcu_free
);
147 * Returns hash value for IPVS LBLC entry
149 static inline unsigned int
150 ip_vs_lblc_hashkey(int af
, const union nf_inet_addr
*addr
)
152 __be32 addr_fold
= addr
->ip
;
154 #ifdef CONFIG_IP_VS_IPV6
156 addr_fold
= addr
->ip6
[0]^addr
->ip6
[1]^
157 addr
->ip6
[2]^addr
->ip6
[3];
159 return hash_32(ntohl(addr_fold
), IP_VS_LBLC_TAB_BITS
);
164 * Hash an entry in the ip_vs_lblc_table.
165 * returns bool success.
168 ip_vs_lblc_hash(struct ip_vs_lblc_table
*tbl
, struct ip_vs_lblc_entry
*en
)
170 unsigned int hash
= ip_vs_lblc_hashkey(en
->af
, &en
->addr
);
172 hlist_add_head_rcu(&en
->list
, &tbl
->bucket
[hash
]);
173 atomic_inc(&tbl
->entries
);
177 /* Get ip_vs_lblc_entry associated with supplied parameters. */
178 static inline struct ip_vs_lblc_entry
*
179 ip_vs_lblc_get(int af
, struct ip_vs_lblc_table
*tbl
,
180 const union nf_inet_addr
*addr
)
182 unsigned int hash
= ip_vs_lblc_hashkey(af
, addr
);
183 struct ip_vs_lblc_entry
*en
;
185 hlist_for_each_entry_rcu(en
, &tbl
->bucket
[hash
], list
)
186 if (ip_vs_addr_equal(af
, &en
->addr
, addr
))
194 * Create or update an ip_vs_lblc_entry, which is a mapping of a destination IP
195 * address to a server. Called under spin lock.
197 static inline struct ip_vs_lblc_entry
*
198 ip_vs_lblc_new(struct ip_vs_lblc_table
*tbl
, const union nf_inet_addr
*daddr
,
199 u16 af
, struct ip_vs_dest
*dest
)
201 struct ip_vs_lblc_entry
*en
;
203 en
= ip_vs_lblc_get(af
, tbl
, daddr
);
205 if (en
->dest
== dest
)
209 en
= kmalloc(sizeof(*en
), GFP_ATOMIC
);
214 ip_vs_addr_copy(af
, &en
->addr
, daddr
);
215 en
->lastuse
= jiffies
;
217 ip_vs_dest_hold(dest
);
220 ip_vs_lblc_hash(tbl
, en
);
227 * Flush all the entries of the specified table.
229 static void ip_vs_lblc_flush(struct ip_vs_service
*svc
)
231 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
232 struct ip_vs_lblc_entry
*en
;
233 struct hlist_node
*next
;
236 spin_lock_bh(&svc
->sched_lock
);
238 for (i
= 0; i
< IP_VS_LBLC_TAB_SIZE
; i
++) {
239 hlist_for_each_entry_safe(en
, next
, &tbl
->bucket
[i
], list
) {
241 atomic_dec(&tbl
->entries
);
244 spin_unlock_bh(&svc
->sched_lock
);
247 static int sysctl_lblc_expiration(struct ip_vs_service
*svc
)
250 return svc
->ipvs
->sysctl_lblc_expiration
;
252 return DEFAULT_EXPIRATION
;
256 static inline void ip_vs_lblc_full_check(struct ip_vs_service
*svc
)
258 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
259 struct ip_vs_lblc_entry
*en
;
260 struct hlist_node
*next
;
261 unsigned long now
= jiffies
;
264 for (i
= 0, j
= tbl
->rover
; i
< IP_VS_LBLC_TAB_SIZE
; i
++) {
265 j
= (j
+ 1) & IP_VS_LBLC_TAB_MASK
;
267 spin_lock(&svc
->sched_lock
);
268 hlist_for_each_entry_safe(en
, next
, &tbl
->bucket
[j
], list
) {
271 sysctl_lblc_expiration(svc
)))
275 atomic_dec(&tbl
->entries
);
277 spin_unlock(&svc
->sched_lock
);
284 * Periodical timer handler for IPVS lblc table
285 * It is used to collect stale entries when the number of entries
286 * exceeds the maximum size of the table.
288 * Fixme: we probably need more complicated algorithm to collect
289 * entries that have not been used for a long time even
290 * if the number of entries doesn't exceed the maximum size
292 * The full expiration check is for this purpose now.
294 static void ip_vs_lblc_check_expire(struct timer_list
*t
)
296 struct ip_vs_lblc_table
*tbl
= from_timer(tbl
, t
, periodic_timer
);
297 struct ip_vs_service
*svc
= tbl
->svc
;
298 unsigned long now
= jiffies
;
301 struct ip_vs_lblc_entry
*en
;
302 struct hlist_node
*next
;
304 if ((tbl
->counter
% COUNT_FOR_FULL_EXPIRATION
) == 0) {
305 /* do full expiration check */
306 ip_vs_lblc_full_check(svc
);
311 if (atomic_read(&tbl
->entries
) <= tbl
->max_size
) {
316 goal
= (atomic_read(&tbl
->entries
) - tbl
->max_size
)*4/3;
317 if (goal
> tbl
->max_size
/2)
318 goal
= tbl
->max_size
/2;
320 for (i
= 0, j
= tbl
->rover
; i
< IP_VS_LBLC_TAB_SIZE
; i
++) {
321 j
= (j
+ 1) & IP_VS_LBLC_TAB_MASK
;
323 spin_lock(&svc
->sched_lock
);
324 hlist_for_each_entry_safe(en
, next
, &tbl
->bucket
[j
], list
) {
325 if (time_before(now
, en
->lastuse
+ ENTRY_TIMEOUT
))
329 atomic_dec(&tbl
->entries
);
332 spin_unlock(&svc
->sched_lock
);
339 mod_timer(&tbl
->periodic_timer
, jiffies
+ CHECK_EXPIRE_INTERVAL
);
343 static int ip_vs_lblc_init_svc(struct ip_vs_service
*svc
)
346 struct ip_vs_lblc_table
*tbl
;
349 * Allocate the ip_vs_lblc_table for this service
351 tbl
= kmalloc(sizeof(*tbl
), GFP_KERNEL
);
355 svc
->sched_data
= tbl
;
356 IP_VS_DBG(6, "LBLC hash table (memory=%zdbytes) allocated for "
357 "current service\n", sizeof(*tbl
));
360 * Initialize the hash buckets
362 for (i
= 0; i
< IP_VS_LBLC_TAB_SIZE
; i
++) {
363 INIT_HLIST_HEAD(&tbl
->bucket
[i
]);
365 tbl
->max_size
= IP_VS_LBLC_TAB_SIZE
*16;
370 atomic_set(&tbl
->entries
, 0);
373 * Hook periodic timer for garbage collection
375 timer_setup(&tbl
->periodic_timer
, ip_vs_lblc_check_expire
, 0);
376 mod_timer(&tbl
->periodic_timer
, jiffies
+ CHECK_EXPIRE_INTERVAL
);
382 static void ip_vs_lblc_done_svc(struct ip_vs_service
*svc
)
384 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
386 /* remove periodic timer */
387 del_timer_sync(&tbl
->periodic_timer
);
389 /* got to clean up table entries here */
390 ip_vs_lblc_flush(svc
);
392 /* release the table itself */
393 kfree_rcu(tbl
, rcu_head
);
394 IP_VS_DBG(6, "LBLC hash table (memory=%zdbytes) released\n",
399 static inline struct ip_vs_dest
*
400 __ip_vs_lblc_schedule(struct ip_vs_service
*svc
)
402 struct ip_vs_dest
*dest
, *least
;
406 * We use the following formula to estimate the load:
407 * (dest overhead) / dest->weight
409 * Remember -- no floats in kernel mode!!!
410 * The comparison of h1*w2 > h2*w1 is equivalent to that of
412 * if every weight is larger than zero.
414 * The server with weight=0 is quiesced and will not receive any
417 list_for_each_entry_rcu(dest
, &svc
->destinations
, n_list
) {
418 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
420 if (atomic_read(&dest
->weight
) > 0) {
422 loh
= ip_vs_dest_conn_overhead(least
);
429 * Find the destination with the least load.
432 list_for_each_entry_continue_rcu(dest
, &svc
->destinations
, n_list
) {
433 if (dest
->flags
& IP_VS_DEST_F_OVERLOAD
)
436 doh
= ip_vs_dest_conn_overhead(dest
);
437 if ((__s64
)loh
* atomic_read(&dest
->weight
) >
438 (__s64
)doh
* atomic_read(&least
->weight
)) {
444 IP_VS_DBG_BUF(6, "LBLC: server %s:%d "
445 "activeconns %d refcnt %d weight %d overhead %d\n",
446 IP_VS_DBG_ADDR(least
->af
, &least
->addr
),
448 atomic_read(&least
->activeconns
),
449 refcount_read(&least
->refcnt
),
450 atomic_read(&least
->weight
), loh
);
457 * If this destination server is overloaded and there is a less loaded
458 * server, then return true.
461 is_overloaded(struct ip_vs_dest
*dest
, struct ip_vs_service
*svc
)
463 if (atomic_read(&dest
->activeconns
) > atomic_read(&dest
->weight
)) {
464 struct ip_vs_dest
*d
;
466 list_for_each_entry_rcu(d
, &svc
->destinations
, n_list
) {
467 if (atomic_read(&d
->activeconns
)*2
468 < atomic_read(&d
->weight
)) {
478 * Locality-Based (weighted) Least-Connection scheduling
480 static struct ip_vs_dest
*
481 ip_vs_lblc_schedule(struct ip_vs_service
*svc
, const struct sk_buff
*skb
,
482 struct ip_vs_iphdr
*iph
)
484 struct ip_vs_lblc_table
*tbl
= svc
->sched_data
;
485 struct ip_vs_dest
*dest
= NULL
;
486 struct ip_vs_lblc_entry
*en
;
488 IP_VS_DBG(6, "%s(): Scheduling...\n", __func__
);
490 /* First look in our cache */
491 en
= ip_vs_lblc_get(svc
->af
, tbl
, &iph
->daddr
);
493 /* We only hold a read lock, but this is atomic */
494 en
->lastuse
= jiffies
;
497 * If the destination is not available, i.e. it's in the trash,
498 * we must ignore it, as it may be removed from under our feet,
499 * if someone drops our reference count. Our caller only makes
500 * sure that destinations, that are not in the trash, are not
501 * moved to the trash, while we are scheduling. But anyone can
502 * free up entries from the trash at any time.
506 if ((dest
->flags
& IP_VS_DEST_F_AVAILABLE
) &&
507 atomic_read(&dest
->weight
) > 0 && !is_overloaded(dest
, svc
))
511 /* No cache entry or it is invalid, time to schedule */
512 dest
= __ip_vs_lblc_schedule(svc
);
514 ip_vs_scheduler_err(svc
, "no destination available");
518 /* If we fail to create a cache entry, we'll just use the valid dest */
519 spin_lock_bh(&svc
->sched_lock
);
521 ip_vs_lblc_new(tbl
, &iph
->daddr
, svc
->af
, dest
);
522 spin_unlock_bh(&svc
->sched_lock
);
525 IP_VS_DBG_BUF(6, "LBLC: destination IP address %s --> server %s:%d\n",
526 IP_VS_DBG_ADDR(svc
->af
, &iph
->daddr
),
527 IP_VS_DBG_ADDR(dest
->af
, &dest
->addr
), ntohs(dest
->port
));
534 * IPVS LBLC Scheduler structure
536 static struct ip_vs_scheduler ip_vs_lblc_scheduler
= {
538 .refcnt
= ATOMIC_INIT(0),
539 .module
= THIS_MODULE
,
540 .n_list
= LIST_HEAD_INIT(ip_vs_lblc_scheduler
.n_list
),
541 .init_service
= ip_vs_lblc_init_svc
,
542 .done_service
= ip_vs_lblc_done_svc
,
543 .schedule
= ip_vs_lblc_schedule
,
550 static int __net_init
__ip_vs_lblc_init(struct net
*net
)
552 struct netns_ipvs
*ipvs
= net_ipvs(net
);
557 if (!net_eq(net
, &init_net
)) {
558 ipvs
->lblc_ctl_table
= kmemdup(vs_vars_table
,
559 sizeof(vs_vars_table
),
561 if (ipvs
->lblc_ctl_table
== NULL
)
564 /* Don't export sysctls to unprivileged users */
565 if (net
->user_ns
!= &init_user_ns
)
566 ipvs
->lblc_ctl_table
[0].procname
= NULL
;
569 ipvs
->lblc_ctl_table
= vs_vars_table
;
570 ipvs
->sysctl_lblc_expiration
= DEFAULT_EXPIRATION
;
571 ipvs
->lblc_ctl_table
[0].data
= &ipvs
->sysctl_lblc_expiration
;
573 ipvs
->lblc_ctl_header
=
574 register_net_sysctl(net
, "net/ipv4/vs", ipvs
->lblc_ctl_table
);
575 if (!ipvs
->lblc_ctl_header
) {
576 if (!net_eq(net
, &init_net
))
577 kfree(ipvs
->lblc_ctl_table
);
584 static void __net_exit
__ip_vs_lblc_exit(struct net
*net
)
586 struct netns_ipvs
*ipvs
= net_ipvs(net
);
588 unregister_net_sysctl_table(ipvs
->lblc_ctl_header
);
590 if (!net_eq(net
, &init_net
))
591 kfree(ipvs
->lblc_ctl_table
);
596 static int __net_init
__ip_vs_lblc_init(struct net
*net
) { return 0; }
597 static void __net_exit
__ip_vs_lblc_exit(struct net
*net
) { }
601 static struct pernet_operations ip_vs_lblc_ops
= {
602 .init
= __ip_vs_lblc_init
,
603 .exit
= __ip_vs_lblc_exit
,
606 static int __init
ip_vs_lblc_init(void)
610 ret
= register_pernet_subsys(&ip_vs_lblc_ops
);
614 ret
= register_ip_vs_scheduler(&ip_vs_lblc_scheduler
);
616 unregister_pernet_subsys(&ip_vs_lblc_ops
);
620 static void __exit
ip_vs_lblc_cleanup(void)
622 unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler
);
623 unregister_pernet_subsys(&ip_vs_lblc_ops
);
628 module_init(ip_vs_lblc_init
);
629 module_exit(ip_vs_lblc_cleanup
);
630 MODULE_LICENSE("GPL");