1/*
2 * IPVS:        Weighted Round-Robin Scheduling module
3 *
4 * Authors:     Wensong Zhang <wensong@linuxvirtualserver.org>
5 *
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.
10 *
11 * Changes:
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
18 *
19 */
20
21#define KMSG_COMPONENT "IPVS"
22#define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
23
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>
29
30#include <net/ip_vs.h>
31
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.
37 *
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:
41 *
42 * pass 1: cw = max weight
43 * pass 2: cw = max weight - di
44 * pass 3: cw = max weight - 2 * di
45 * ...
46 * last pass: cw = di
47 *
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.
51 *
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)
57 * ...
58 * last pass: cw = 1
59 *
60 */
61
62/*
63 * current destination pointer for weighted round-robin scheduling
64 */
65struct 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;
71};
72
73
74static int ip_vs_wrr_gcd_weight(struct ip_vs_service *svc)
75{
76	struct ip_vs_dest *dest;
77	int weight;
78	int g = 0;
79
80	list_for_each_entry(dest, &svc->destinations, n_list) {
81		weight = atomic_read(&dest->weight);
82		if (weight > 0) {
83			if (g > 0)
84				g = gcd(weight, g);
85			else
86				g = weight;
87		}
88	}
89	return g ? g : 1;
90}
91
92
93/*
94 *    Get the maximum weight of the service destinations.
95 */
96static int ip_vs_wrr_max_weight(struct ip_vs_service *svc)
97{
98	struct ip_vs_dest *dest;
99	int new_weight, weight = 0;
100
101	list_for_each_entry(dest, &svc->destinations, n_list) {
102		new_weight = atomic_read(&dest->weight);
103		if (new_weight > weight)
104			weight = new_weight;
105	}
106
107	return weight;
108}
109
110
111static int ip_vs_wrr_init_svc(struct ip_vs_service *svc)
112{
113	struct ip_vs_wrr_mark *mark;
114
115	/*
116	 *    Allocate the mark variable for WRR scheduling
117	 */
118	mark = kmalloc(sizeof(struct ip_vs_wrr_mark), GFP_KERNEL);
119	if (mark == NULL)
120		return -ENOMEM;
121
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);
125	mark->cw = mark->mw;
126	svc->sched_data = mark;
127
128	return 0;
129}
130
131
132static void ip_vs_wrr_done_svc(struct ip_vs_service *svc)
133{
134	struct ip_vs_wrr_mark *mark = svc->sched_data;
135
136	/*
137	 *    Release the mark variable
138	 */
139	kfree_rcu(mark, rcu_head);
140}
141
142
143static int ip_vs_wrr_dest_changed(struct ip_vs_service *svc,
144				  struct ip_vs_dest *dest)
145{
146	struct ip_vs_wrr_mark *mark = svc->sched_data;
147
148	spin_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)
153		mark->cw = mark->mw;
154	else if (mark->di > 1)
155		mark->cw = (mark->cw / mark->di) * mark->di + 1;
156	spin_unlock_bh(&svc->sched_lock);
157	return 0;
158}
159
160
161/*
162 *    Weighted Round-Robin Scheduling
163 */
164static struct ip_vs_dest *
165ip_vs_wrr_schedule(struct ip_vs_service *svc, const struct sk_buff *skb,
166		   struct ip_vs_iphdr *iph)
167{
168	struct ip_vs_dest *dest, *last, *stop = NULL;
169	struct ip_vs_wrr_mark *mark = svc->sched_data;
170	bool last_pass = false, restarted = false;
171
172	IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
173
174	spin_lock_bh(&svc->sched_lock);
175	dest = mark->cl;
176	/* No available dests? */
177	if (mark->mw == 0)
178		goto err_noavail;
179	last = dest;
180	/* Stop only after all dests were checked for weight >= 1 (last pass) */
181	while (1) {
182		list_for_each_entry_continue_rcu(dest,
183						 &svc->destinations,
184						 n_list) {
185			if (!(dest->flags & IP_VS_DEST_F_OVERLOAD) &&
186			    atomic_read(&dest->weight) >= mark->cw)
187				goto found;
188			if (dest == stop)
189				goto err_over;
190		}
191		mark->cw -= mark->di;
192		if (mark->cw <= 0) {
193			mark->cw = mark->mw;
194			/* Stop if we tried last pass from first dest:
195			 * 1. last_pass: we started checks when cw > di but
196			 *	then all dests were checked for w >= 1
197			 * 2. last was head: the first and only traversal
198			 *	was for weight >= 1, for all dests.
199			 */
200			if (last_pass ||
201			    &last->n_list == &svc->destinations)
202				goto err_over;
203			restarted = true;
204		}
205		last_pass = mark->cw <= mark->di;
206		if (last_pass && restarted &&
207		    &last->n_list != &svc->destinations) {
208			/* First traversal was for w >= 1 but only
209			 * for dests after 'last', now do the same
210			 * for all dests up to 'last'.
211			 */
212			stop = last;
213		}
214	}
215
216found:
217	IP_VS_DBG_BUF(6, "WRR: server %s:%u "
218		      "activeconns %d refcnt %d weight %d\n",
219		      IP_VS_DBG_ADDR(dest->af, &dest->addr), ntohs(dest->port),
220		      atomic_read(&dest->activeconns),
221		      atomic_read(&dest->refcnt),
222		      atomic_read(&dest->weight));
223	mark->cl = dest;
224
225  out:
226	spin_unlock_bh(&svc->sched_lock);
227	return dest;
228
229err_noavail:
230	mark->cl = dest;
231	dest = NULL;
232	ip_vs_scheduler_err(svc, "no destination available");
233	goto out;
234
235err_over:
236	mark->cl = dest;
237	dest = NULL;
238	ip_vs_scheduler_err(svc, "no destination available: "
239			    "all destinations are overloaded");
240	goto out;
241}
242
243
244static struct ip_vs_scheduler ip_vs_wrr_scheduler = {
245	.name =			"wrr",
246	.refcnt =		ATOMIC_INIT(0),
247	.module =		THIS_MODULE,
248	.n_list =		LIST_HEAD_INIT(ip_vs_wrr_scheduler.n_list),
249	.init_service =		ip_vs_wrr_init_svc,
250	.done_service =		ip_vs_wrr_done_svc,
251	.add_dest =		ip_vs_wrr_dest_changed,
252	.del_dest =		ip_vs_wrr_dest_changed,
253	.upd_dest =		ip_vs_wrr_dest_changed,
254	.schedule =		ip_vs_wrr_schedule,
255};
256
257static int __init ip_vs_wrr_init(void)
258{
259	return register_ip_vs_scheduler(&ip_vs_wrr_scheduler) ;
260}
261
262static void __exit ip_vs_wrr_cleanup(void)
263{
264	unregister_ip_vs_scheduler(&ip_vs_wrr_scheduler);
265	synchronize_rcu();
266}
267
268module_init(ip_vs_wrr_init);
269module_exit(ip_vs_wrr_cleanup);
270MODULE_LICENSE("GPL");
271