1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * IPVS:        Shortest Expected Delay scheduling module
4  *
5  * Authors:     Wensong Zhang <wensong@linuxvirtualserver.org>
6  *
7  * Changes:
8  */
9 
10 /*
11  * The SED algorithm attempts to minimize each job's expected delay until
12  * completion. The expected delay that the job will experience is
13  * (Ci + 1) / Ui if sent to the ith server, in which Ci is the number of
14  * jobs on the ith server and Ui is the fixed service rate (weight) of
15  * the ith server. The SED algorithm adopts a greedy policy that each does
16  * what is in its own best interest, i.e. to join the queue which would
17  * minimize its expected delay of completion.
18  *
19  * See the following paper for more information:
20  * A. Weinrib and S. Shenker, Greed is not enough: Adaptive load sharing
21  * in large heterogeneous systems. In Proceedings IEEE INFOCOM'88,
22  * pages 986-994, 1988.
23  *
24  * Thanks must go to Marko Buuri <marko@buuri.name> for talking SED to me.
25  *
26  * The difference between SED and WLC is that SED includes the incoming
27  * job in the cost function (the increment of 1). SED may outperform
28  * WLC, while scheduling big jobs under larger heterogeneous systems
29  * (the server weight varies a lot).
30  *
31  */
32 
33 #define KMSG_COMPONENT "IPVS"
34 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
35 
36 #include <linux/module.h>
37 #include <linux/kernel.h>
38 
39 #include <net/ip_vs.h>
40 
41 
42 static inline int
ip_vs_sed_dest_overhead(struct ip_vs_dest * dest)43 ip_vs_sed_dest_overhead(struct ip_vs_dest *dest)
44 {
45 	/*
46 	 * We only use the active connection number in the cost
47 	 * calculation here.
48 	 */
49 	return atomic_read(&dest->activeconns) + 1;
50 }
51 
52 
53 /*
54  *	Weighted Least Connection scheduling
55  */
56 static struct ip_vs_dest *
ip_vs_sed_schedule(struct ip_vs_service * svc,const struct sk_buff * skb,struct ip_vs_iphdr * iph)57 ip_vs_sed_schedule(struct ip_vs_service *svc, const struct sk_buff *skb,
58 		   struct ip_vs_iphdr *iph)
59 {
60 	struct ip_vs_dest *dest, *least;
61 	int loh, doh;
62 
63 	IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
64 
65 	/*
66 	 * We calculate the load of each dest server as follows:
67 	 *	(server expected overhead) / dest->weight
68 	 *
69 	 * Remember -- no floats in kernel mode!!!
70 	 * The comparison of h1*w2 > h2*w1 is equivalent to that of
71 	 *		  h1/w1 > h2/w2
72 	 * if every weight is larger than zero.
73 	 *
74 	 * The server with weight=0 is quiesced and will not receive any
75 	 * new connections.
76 	 */
77 
78 	list_for_each_entry_rcu(dest, &svc->destinations, n_list) {
79 		if (!(dest->flags & IP_VS_DEST_F_OVERLOAD) &&
80 		    atomic_read(&dest->weight) > 0) {
81 			least = dest;
82 			loh = ip_vs_sed_dest_overhead(least);
83 			goto nextstage;
84 		}
85 	}
86 	ip_vs_scheduler_err(svc, "no destination available");
87 	return NULL;
88 
89 	/*
90 	 *    Find the destination with the least load.
91 	 */
92   nextstage:
93 	list_for_each_entry_continue_rcu(dest, &svc->destinations, n_list) {
94 		if (dest->flags & IP_VS_DEST_F_OVERLOAD)
95 			continue;
96 		doh = ip_vs_sed_dest_overhead(dest);
97 		if ((__s64)loh * atomic_read(&dest->weight) >
98 		    (__s64)doh * atomic_read(&least->weight)) {
99 			least = dest;
100 			loh = doh;
101 		}
102 	}
103 
104 	IP_VS_DBG_BUF(6, "SED: server %s:%u "
105 		      "activeconns %d refcnt %d weight %d overhead %d\n",
106 		      IP_VS_DBG_ADDR(least->af, &least->addr),
107 		      ntohs(least->port),
108 		      atomic_read(&least->activeconns),
109 		      refcount_read(&least->refcnt),
110 		      atomic_read(&least->weight), loh);
111 
112 	return least;
113 }
114 
115 
116 static struct ip_vs_scheduler ip_vs_sed_scheduler =
117 {
118 	.name =			"sed",
119 	.refcnt =		ATOMIC_INIT(0),
120 	.module =		THIS_MODULE,
121 	.n_list =		LIST_HEAD_INIT(ip_vs_sed_scheduler.n_list),
122 	.schedule =		ip_vs_sed_schedule,
123 };
124 
125 
ip_vs_sed_init(void)126 static int __init ip_vs_sed_init(void)
127 {
128 	return register_ip_vs_scheduler(&ip_vs_sed_scheduler);
129 }
130 
ip_vs_sed_cleanup(void)131 static void __exit ip_vs_sed_cleanup(void)
132 {
133 	unregister_ip_vs_scheduler(&ip_vs_sed_scheduler);
134 	synchronize_rcu();
135 }
136 
137 module_init(ip_vs_sed_init);
138 module_exit(ip_vs_sed_cleanup);
139 MODULE_LICENSE("GPL");
140