2 * IPVS: Weighted Round-Robin Scheduling module
4 * Authors: Wensong Zhang <wensong@linuxvirtualserver.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 * Wensong Zhang : changed the ip_vs_wrr_schedule to return dest
13 * Wensong Zhang : changed some comestics things for debugging
14 * Wensong Zhang : changed for the d-linked destination list
15 * Wensong Zhang : added the ip_vs_wrr_update_svc
16 * Julian Anastasov : fixed the bug of returning destination
17 * with weight 0 when all weights are zero
21 #define KMSG_COMPONENT "IPVS"
22 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
24 #include <linux/module.h>
25 #include <linux/kernel.h>
26 #include <linux/slab.h>
27 #include <linux/net.h>
28 #include <linux/gcd.h>
30 #include <net/ip_vs.h>
32 /* The WRR algorithm depends on some caclulations:
33 * - mw: maximum weight
34 * - di: weight step, greatest common divisor from all weights
35 * - cw: current required weight
36 * As result, all weights are in the [di..mw] range with a step=di.
38 * First, we start with cw = mw and select dests with weight >= cw.
39 * Then cw is reduced with di and all dests are checked again.
40 * Last pass should be with cw = di. We have mw/di passes in total:
42 * pass 1: cw = max weight
43 * pass 2: cw = max weight - di
44 * pass 3: cw = max weight - 2 * di
48 * Weights are supposed to be >= di but we run in parallel with
49 * weight changes, it is possible some dest weight to be reduced
50 * below di, bad if it is the only available dest.
52 * So, we modify how mw is calculated, now it is reduced with (di - 1),
53 * so that last cw is 1 to catch such dests with weight below di:
54 * pass 1: cw = max weight - (di - 1)
55 * pass 2: cw = max weight - di - (di - 1)
56 * pass 3: cw = max weight - 2 * di - (di - 1)
63 * current destination pointer for weighted round-robin scheduling
65 struct ip_vs_wrr_mark
{
66 struct ip_vs_dest
*cl
; /* current dest or head */
67 int cw
; /* current weight */
68 int mw
; /* maximum weight */
69 int di
; /* decreasing interval */
70 struct rcu_head rcu_head
;
74 static int ip_vs_wrr_gcd_weight(struct ip_vs_service
*svc
)
76 struct ip_vs_dest
*dest
;
80 list_for_each_entry(dest
, &svc
->destinations
, n_list
) {
81 weight
= atomic_read(&dest
->weight
);
94 * Get the maximum weight of the service destinations.
96 static int ip_vs_wrr_max_weight(struct ip_vs_service
*svc
)
98 struct ip_vs_dest
*dest
;
99 int new_weight
, weight
= 0;
101 list_for_each_entry(dest
, &svc
->destinations
, n_list
) {
102 new_weight
= atomic_read(&dest
->weight
);
103 if (new_weight
> weight
)
111 static int ip_vs_wrr_init_svc(struct ip_vs_service
*svc
)
113 struct ip_vs_wrr_mark
*mark
;
116 * Allocate the mark variable for WRR scheduling
118 mark
= kmalloc(sizeof(struct ip_vs_wrr_mark
), GFP_KERNEL
);
122 mark
->cl
= list_entry(&svc
->destinations
, struct ip_vs_dest
, n_list
);
123 mark
->di
= ip_vs_wrr_gcd_weight(svc
);
124 mark
->mw
= ip_vs_wrr_max_weight(svc
) - (mark
->di
- 1);
126 svc
->sched_data
= mark
;
132 static void ip_vs_wrr_done_svc(struct ip_vs_service
*svc
)
134 struct ip_vs_wrr_mark
*mark
= svc
->sched_data
;
137 * Release the mark variable
139 kfree_rcu(mark
, rcu_head
);
143 static int ip_vs_wrr_dest_changed(struct ip_vs_service
*svc
,
144 struct ip_vs_dest
*dest
)
146 struct ip_vs_wrr_mark
*mark
= svc
->sched_data
;
148 write_lock_bh(&svc
->sched_lock
);
149 mark
->cl
= list_entry(&svc
->destinations
, struct ip_vs_dest
, n_list
);
150 mark
->di
= ip_vs_wrr_gcd_weight(svc
);
151 mark
->mw
= ip_vs_wrr_max_weight(svc
) - (mark
->di
- 1);
152 if (mark
->cw
> mark
->mw
|| !mark
->cw
)
154 else if (mark
->di
> 1)
155 mark
->cw
= (mark
->cw
/ mark
->di
) * mark
->di
+ 1;
156 write_unlock_bh(&svc
->sched_lock
);
162 * Weighted Round-Robin Scheduling
164 static struct ip_vs_dest
*
165 ip_vs_wrr_schedule(struct ip_vs_service
*svc
, const struct sk_buff
*skb
)
167 struct ip_vs_dest
*dest
, *last
, *stop
= NULL
;
168 struct ip_vs_wrr_mark
*mark
= svc
->sched_data
;
169 bool last_pass
= false, restarted
= false;
171 IP_VS_DBG(6, "%s(): Scheduling...\n", __func__
);
173 write_lock(&svc
->sched_lock
);
175 /* No available dests? */
179 /* Stop only after all dests were checked for weight >= 1 (last pass) */
181 list_for_each_entry_continue_rcu(dest
,
184 if (!(dest
->flags
& IP_VS_DEST_F_OVERLOAD
) &&
185 atomic_read(&dest
->weight
) >= mark
->cw
)
190 mark
->cw
-= mark
->di
;
193 /* Stop if we tried last pass from first dest:
194 * 1. last_pass: we started checks when cw > di but
195 * then all dests were checked for w >= 1
196 * 2. last was head: the first and only traversal
197 * was for weight >= 1, for all dests.
200 &last
->n_list
== &svc
->destinations
)
204 last_pass
= mark
->cw
<= mark
->di
;
205 if (last_pass
&& restarted
&&
206 &last
->n_list
!= &svc
->destinations
) {
207 /* First traversal was for w >= 1 but only
208 * for dests after 'last', now do the same
209 * for all dests up to 'last'.
216 IP_VS_DBG_BUF(6, "WRR: server %s:%u "
217 "activeconns %d refcnt %d weight %d\n",
218 IP_VS_DBG_ADDR(svc
->af
, &dest
->addr
), ntohs(dest
->port
),
219 atomic_read(&dest
->activeconns
),
220 atomic_read(&dest
->refcnt
),
221 atomic_read(&dest
->weight
));
225 write_unlock(&svc
->sched_lock
);
231 ip_vs_scheduler_err(svc
, "no destination available");
237 ip_vs_scheduler_err(svc
, "no destination available: "
238 "all destinations are overloaded");
243 static struct ip_vs_scheduler ip_vs_wrr_scheduler
= {
245 .refcnt
= ATOMIC_INIT(0),
246 .module
= THIS_MODULE
,
247 .n_list
= LIST_HEAD_INIT(ip_vs_wrr_scheduler
.n_list
),
248 .init_service
= ip_vs_wrr_init_svc
,
249 .done_service
= ip_vs_wrr_done_svc
,
250 .add_dest
= ip_vs_wrr_dest_changed
,
251 .del_dest
= ip_vs_wrr_dest_changed
,
252 .upd_dest
= ip_vs_wrr_dest_changed
,
253 .schedule
= ip_vs_wrr_schedule
,
256 static int __init
ip_vs_wrr_init(void)
258 return register_ip_vs_scheduler(&ip_vs_wrr_scheduler
) ;
261 static void __exit
ip_vs_wrr_cleanup(void)
263 unregister_ip_vs_scheduler(&ip_vs_wrr_scheduler
);
266 module_init(ip_vs_wrr_init
);
267 module_exit(ip_vs_wrr_cleanup
);
268 MODULE_LICENSE("GPL");