| 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
| 2 | /* |
| 3 | * IPVS: Never Queue scheduling module |
| 4 | * |
| 5 | * Authors: Wensong Zhang <wensong@linuxvirtualserver.org> |
| 6 | * |
| 7 | * Changes: |
| 8 | */ |
| 9 | |
| 10 | /* |
| 11 | * The NQ algorithm adopts a two-speed model. When there is an idle server |
| 12 | * available, the job will be sent to the idle server, instead of waiting |
| 13 | * for a fast one. When there is no idle server available, the job will be |
| 14 | * sent to the server that minimize its expected delay (The Shortest |
| 15 | * Expected Delay scheduling algorithm). |
| 16 | * |
| 17 | * See the following paper for more information: |
| 18 | * A. Weinrib and S. Shenker, Greed is not enough: Adaptive load sharing |
| 19 | * in large heterogeneous systems. In Proceedings IEEE INFOCOM'88, |
| 20 | * pages 986-994, 1988. |
| 21 | * |
| 22 | * Thanks must go to Marko Buuri <marko@buuri.name> for talking NQ to me. |
| 23 | * |
| 24 | * The difference between NQ and SED is that NQ can improve overall |
| 25 | * system utilization. |
| 26 | * |
| 27 | */ |
| 28 | |
| 29 | #define pr_fmt(fmt) "IPVS: " fmt |
| 30 | |
| 31 | #include <linux/module.h> |
| 32 | #include <linux/kernel.h> |
| 33 | |
| 34 | #include <net/ip_vs.h> |
| 35 | |
| 36 | |
| 37 | static inline int |
| 38 | ip_vs_nq_dest_overhead(struct ip_vs_dest *dest) |
| 39 | { |
| 40 | /* |
| 41 | * We only use the active connection number in the cost |
| 42 | * calculation here. |
| 43 | */ |
| 44 | return atomic_read(v: &dest->activeconns) + 1; |
| 45 | } |
| 46 | |
| 47 | |
| 48 | /* |
| 49 | * Weighted Least Connection scheduling |
| 50 | */ |
| 51 | static struct ip_vs_dest * |
| 52 | ip_vs_nq_schedule(struct ip_vs_service *svc, const struct sk_buff *skb, |
| 53 | struct ip_vs_iphdr *iph) |
| 54 | { |
| 55 | struct ip_vs_dest *dest, *least = NULL; |
| 56 | int loh = 0, doh; |
| 57 | |
| 58 | IP_VS_DBG(6, "%s(): Scheduling...\n" , __func__); |
| 59 | |
| 60 | /* |
| 61 | * We calculate the load of each dest server as follows: |
| 62 | * (server expected overhead) / dest->weight |
| 63 | * |
| 64 | * Remember -- no floats in kernel mode!!! |
| 65 | * The comparison of h1*w2 > h2*w1 is equivalent to that of |
| 66 | * h1/w1 > h2/w2 |
| 67 | * if every weight is larger than zero. |
| 68 | * |
| 69 | * The server with weight=0 is quiesced and will not receive any |
| 70 | * new connections. |
| 71 | */ |
| 72 | |
| 73 | list_for_each_entry_rcu(dest, &svc->destinations, n_list) { |
| 74 | |
| 75 | if (dest->flags & IP_VS_DEST_F_OVERLOAD || |
| 76 | !atomic_read(v: &dest->weight)) |
| 77 | continue; |
| 78 | |
| 79 | doh = ip_vs_nq_dest_overhead(dest); |
| 80 | |
| 81 | /* return the server directly if it is idle */ |
| 82 | if (atomic_read(v: &dest->activeconns) == 0) { |
| 83 | least = dest; |
| 84 | loh = doh; |
| 85 | goto out; |
| 86 | } |
| 87 | |
| 88 | if (!least || |
| 89 | ((__s64)loh * atomic_read(v: &dest->weight) > |
| 90 | (__s64)doh * atomic_read(v: &least->weight))) { |
| 91 | least = dest; |
| 92 | loh = doh; |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | if (!least) { |
| 97 | ip_vs_scheduler_err(svc, msg: "no destination available" ); |
| 98 | return NULL; |
| 99 | } |
| 100 | |
| 101 | out: |
| 102 | IP_VS_DBG_BUF(6, "NQ: server %s:%u " |
| 103 | "activeconns %d refcnt %d weight %d overhead %d\n" , |
| 104 | IP_VS_DBG_ADDR(least->af, &least->addr), |
| 105 | ntohs(least->port), |
| 106 | atomic_read(&least->activeconns), |
| 107 | refcount_read(&least->refcnt), |
| 108 | atomic_read(&least->weight), loh); |
| 109 | |
| 110 | return least; |
| 111 | } |
| 112 | |
| 113 | |
| 114 | static struct ip_vs_scheduler ip_vs_nq_scheduler = |
| 115 | { |
| 116 | .name = "nq" , |
| 117 | .refcnt = ATOMIC_INIT(0), |
| 118 | .module = THIS_MODULE, |
| 119 | .n_list = LIST_HEAD_INIT(ip_vs_nq_scheduler.n_list), |
| 120 | .schedule = ip_vs_nq_schedule, |
| 121 | }; |
| 122 | |
| 123 | |
| 124 | static int __init ip_vs_nq_init(void) |
| 125 | { |
| 126 | return register_ip_vs_scheduler(scheduler: &ip_vs_nq_scheduler); |
| 127 | } |
| 128 | |
| 129 | static void __exit ip_vs_nq_cleanup(void) |
| 130 | { |
| 131 | unregister_ip_vs_scheduler(scheduler: &ip_vs_nq_scheduler); |
| 132 | synchronize_rcu(); |
| 133 | } |
| 134 | |
| 135 | module_init(ip_vs_nq_init); |
| 136 | module_exit(ip_vs_nq_cleanup); |
| 137 | MODULE_LICENSE("GPL" ); |
| 138 | MODULE_DESCRIPTION("ipvs never queue scheduler" ); |
| 139 | |