net/netfilter/ipvs/ip_vs_sed.c

   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");