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 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 * 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 }; 124 125 126 static int __init ip_vs_sed_init(void) 127 { 128 return register_ip_vs_scheduler(&ip_vs_sed_scheduler); 129 } 130 131 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