xref: /linux/net/netfilter/ipvs/ip_vs_wrr.c (revision 0be3ff0c)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * IPVS:        Weighted Round-Robin Scheduling module
4  *
5  * Authors:     Wensong Zhang <wensong@linuxvirtualserver.org>
6  *
7  * Changes:
8  *     Wensong Zhang            :     changed the ip_vs_wrr_schedule to return dest
9  *     Wensong Zhang            :     changed some comestics things for debugging
10  *     Wensong Zhang            :     changed for the d-linked destination list
11  *     Wensong Zhang            :     added the ip_vs_wrr_update_svc
12  *     Julian Anastasov         :     fixed the bug of returning destination
13  *                                    with weight 0 when all weights are zero
14  */
15 
16 #define KMSG_COMPONENT "IPVS"
17 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
18 
19 #include <linux/module.h>
20 #include <linux/kernel.h>
21 #include <linux/slab.h>
22 #include <linux/net.h>
23 #include <linux/gcd.h>
24 
25 #include <net/ip_vs.h>
26 
27 /* The WRR algorithm depends on some caclulations:
28  * - mw: maximum weight
29  * - di: weight step, greatest common divisor from all weights
30  * - cw: current required weight
31  * As result, all weights are in the [di..mw] range with a step=di.
32  *
33  * First, we start with cw = mw and select dests with weight >= cw.
34  * Then cw is reduced with di and all dests are checked again.
35  * Last pass should be with cw = di. We have mw/di passes in total:
36  *
37  * pass 1: cw = max weight
38  * pass 2: cw = max weight - di
39  * pass 3: cw = max weight - 2 * di
40  * ...
41  * last pass: cw = di
42  *
43  * Weights are supposed to be >= di but we run in parallel with
44  * weight changes, it is possible some dest weight to be reduced
45  * below di, bad if it is the only available dest.
46  *
47  * So, we modify how mw is calculated, now it is reduced with (di - 1),
48  * so that last cw is 1 to catch such dests with weight below di:
49  * pass 1: cw = max weight - (di - 1)
50  * pass 2: cw = max weight - di - (di - 1)
51  * pass 3: cw = max weight - 2 * di - (di - 1)
52  * ...
53  * last pass: cw = 1
54  *
55  */
56 
57 /*
58  * current destination pointer for weighted round-robin scheduling
59  */
60 struct ip_vs_wrr_mark {
61 	struct ip_vs_dest *cl;	/* current dest or head */
62 	int cw;			/* current weight */
63 	int mw;			/* maximum weight */
64 	int di;			/* decreasing interval */
65 	struct rcu_head		rcu_head;
66 };
67 
68 
69 static int ip_vs_wrr_gcd_weight(struct ip_vs_service *svc)
70 {
71 	struct ip_vs_dest *dest;
72 	int weight;
73 	int g = 0;
74 
75 	list_for_each_entry(dest, &svc->destinations, n_list) {
76 		weight = atomic_read(&dest->weight);
77 		if (weight > 0) {
78 			if (g > 0)
79 				g = gcd(weight, g);
80 			else
81 				g = weight;
82 		}
83 	}
84 	return g ? g : 1;
85 }
86 
87 
88 /*
89  *    Get the maximum weight of the service destinations.
90  */
91 static int ip_vs_wrr_max_weight(struct ip_vs_service *svc)
92 {
93 	struct ip_vs_dest *dest;
94 	int new_weight, weight = 0;
95 
96 	list_for_each_entry(dest, &svc->destinations, n_list) {
97 		new_weight = atomic_read(&dest->weight);
98 		if (new_weight > weight)
99 			weight = new_weight;
100 	}
101 
102 	return weight;
103 }
104 
105 
106 static int ip_vs_wrr_init_svc(struct ip_vs_service *svc)
107 {
108 	struct ip_vs_wrr_mark *mark;
109 
110 	/*
111 	 *    Allocate the mark variable for WRR scheduling
112 	 */
113 	mark = kmalloc(sizeof(struct ip_vs_wrr_mark), GFP_KERNEL);
114 	if (mark == NULL)
115 		return -ENOMEM;
116 
117 	mark->cl = list_entry(&svc->destinations, struct ip_vs_dest, n_list);
118 	mark->di = ip_vs_wrr_gcd_weight(svc);
119 	mark->mw = ip_vs_wrr_max_weight(svc) - (mark->di - 1);
120 	mark->cw = mark->mw;
121 	svc->sched_data = mark;
122 
123 	return 0;
124 }
125 
126 
127 static void ip_vs_wrr_done_svc(struct ip_vs_service *svc)
128 {
129 	struct ip_vs_wrr_mark *mark = svc->sched_data;
130 
131 	/*
132 	 *    Release the mark variable
133 	 */
134 	kfree_rcu(mark, rcu_head);
135 }
136 
137 
138 static int ip_vs_wrr_dest_changed(struct ip_vs_service *svc,
139 				  struct ip_vs_dest *dest)
140 {
141 	struct ip_vs_wrr_mark *mark = svc->sched_data;
142 
143 	spin_lock_bh(&svc->sched_lock);
144 	mark->cl = list_entry(&svc->destinations, struct ip_vs_dest, n_list);
145 	mark->di = ip_vs_wrr_gcd_weight(svc);
146 	mark->mw = ip_vs_wrr_max_weight(svc) - (mark->di - 1);
147 	if (mark->cw > mark->mw || !mark->cw)
148 		mark->cw = mark->mw;
149 	else if (mark->di > 1)
150 		mark->cw = (mark->cw / mark->di) * mark->di + 1;
151 	spin_unlock_bh(&svc->sched_lock);
152 	return 0;
153 }
154 
155 
156 /*
157  *    Weighted Round-Robin Scheduling
158  */
159 static struct ip_vs_dest *
160 ip_vs_wrr_schedule(struct ip_vs_service *svc, const struct sk_buff *skb,
161 		   struct ip_vs_iphdr *iph)
162 {
163 	struct ip_vs_dest *dest, *last, *stop = NULL;
164 	struct ip_vs_wrr_mark *mark = svc->sched_data;
165 	bool last_pass = false, restarted = false;
166 
167 	IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
168 
169 	spin_lock_bh(&svc->sched_lock);
170 	dest = mark->cl;
171 	/* No available dests? */
172 	if (mark->mw == 0)
173 		goto err_noavail;
174 	last = dest;
175 	/* Stop only after all dests were checked for weight >= 1 (last pass) */
176 	while (1) {
177 		list_for_each_entry_continue_rcu(dest,
178 						 &svc->destinations,
179 						 n_list) {
180 			if (!(dest->flags & IP_VS_DEST_F_OVERLOAD) &&
181 			    atomic_read(&dest->weight) >= mark->cw)
182 				goto found;
183 			if (dest == stop)
184 				goto err_over;
185 		}
186 		mark->cw -= mark->di;
187 		if (mark->cw <= 0) {
188 			mark->cw = mark->mw;
189 			/* Stop if we tried last pass from first dest:
190 			 * 1. last_pass: we started checks when cw > di but
191 			 *	then all dests were checked for w >= 1
192 			 * 2. last was head: the first and only traversal
193 			 *	was for weight >= 1, for all dests.
194 			 */
195 			if (last_pass ||
196 			    &last->n_list == &svc->destinations)
197 				goto err_over;
198 			restarted = true;
199 		}
200 		last_pass = mark->cw <= mark->di;
201 		if (last_pass && restarted &&
202 		    &last->n_list != &svc->destinations) {
203 			/* First traversal was for w >= 1 but only
204 			 * for dests after 'last', now do the same
205 			 * for all dests up to 'last'.
206 			 */
207 			stop = last;
208 		}
209 	}
210 
211 found:
212 	IP_VS_DBG_BUF(6, "WRR: server %s:%u "
213 		      "activeconns %d refcnt %d weight %d\n",
214 		      IP_VS_DBG_ADDR(dest->af, &dest->addr), ntohs(dest->port),
215 		      atomic_read(&dest->activeconns),
216 		      refcount_read(&dest->refcnt),
217 		      atomic_read(&dest->weight));
218 	mark->cl = dest;
219 
220   out:
221 	spin_unlock_bh(&svc->sched_lock);
222 	return dest;
223 
224 err_noavail:
225 	mark->cl = dest;
226 	dest = NULL;
227 	ip_vs_scheduler_err(svc, "no destination available");
228 	goto out;
229 
230 err_over:
231 	mark->cl = dest;
232 	dest = NULL;
233 	ip_vs_scheduler_err(svc, "no destination available: "
234 			    "all destinations are overloaded");
235 	goto out;
236 }
237 
238 
239 static struct ip_vs_scheduler ip_vs_wrr_scheduler = {
240 	.name =			"wrr",
241 	.refcnt =		ATOMIC_INIT(0),
242 	.module =		THIS_MODULE,
243 	.n_list =		LIST_HEAD_INIT(ip_vs_wrr_scheduler.n_list),
244 	.init_service =		ip_vs_wrr_init_svc,
245 	.done_service =		ip_vs_wrr_done_svc,
246 	.add_dest =		ip_vs_wrr_dest_changed,
247 	.del_dest =		ip_vs_wrr_dest_changed,
248 	.upd_dest =		ip_vs_wrr_dest_changed,
249 	.schedule =		ip_vs_wrr_schedule,
250 };
251 
252 static int __init ip_vs_wrr_init(void)
253 {
254 	return register_ip_vs_scheduler(&ip_vs_wrr_scheduler) ;
255 }
256 
257 static void __exit ip_vs_wrr_cleanup(void)
258 {
259 	unregister_ip_vs_scheduler(&ip_vs_wrr_scheduler);
260 	synchronize_rcu();
261 }
262 
263 module_init(ip_vs_wrr_init);
264 module_exit(ip_vs_wrr_cleanup);
265 MODULE_LICENSE("GPL");
266