swh:1:snp:173f8deb0c56c557784b4fd217e7608ac6197844
Tip revision: 219d54332a09e8d8741c1e1982f5eae56099de85 authored by Linus Torvalds on 25 November 2019, 00:32:01 UTC
Linux 5.4
Linux 5.4
Tip revision: 219d543
ip_vs_nq.c
// SPDX-License-Identifier: GPL-2.0-or-later
/*
* IPVS: Never Queue scheduling module
*
* Authors: Wensong Zhang <wensong@linuxvirtualserver.org>
*
* Changes:
*/
/*
* The NQ algorithm adopts a two-speed model. When there is an idle server
* available, the job will be sent to the idle server, instead of waiting
* for a fast one. When there is no idle server available, the job will be
* sent to the server that minimize its expected delay (The Shortest
* Expected Delay scheduling algorithm).
*
* See the following paper for more information:
* A. Weinrib and S. Shenker, Greed is not enough: Adaptive load sharing
* in large heterogeneous systems. In Proceedings IEEE INFOCOM'88,
* pages 986-994, 1988.
*
* Thanks must go to Marko Buuri <marko@buuri.name> for talking NQ to me.
*
* The difference between NQ and SED is that NQ can improve overall
* system utilization.
*
*/
#define KMSG_COMPONENT "IPVS"
#define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
#include <linux/module.h>
#include <linux/kernel.h>
#include <net/ip_vs.h>
static inline int
ip_vs_nq_dest_overhead(struct ip_vs_dest *dest)
{
/*
* We only use the active connection number in the cost
* calculation here.
*/
return atomic_read(&dest->activeconns) + 1;
}
/*
* Weighted Least Connection scheduling
*/
static struct ip_vs_dest *
ip_vs_nq_schedule(struct ip_vs_service *svc, const struct sk_buff *skb,
struct ip_vs_iphdr *iph)
{
struct ip_vs_dest *dest, *least = NULL;
int loh = 0, doh;
IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
/*
* We calculate the load of each dest server as follows:
* (server expected overhead) / dest->weight
*
* Remember -- no floats in kernel mode!!!
* The comparison of h1*w2 > h2*w1 is equivalent to that of
* h1/w1 > h2/w2
* if every weight is larger than zero.
*
* The server with weight=0 is quiesced and will not receive any
* new connections.
*/
list_for_each_entry_rcu(dest, &svc->destinations, n_list) {
if (dest->flags & IP_VS_DEST_F_OVERLOAD ||
!atomic_read(&dest->weight))
continue;
doh = ip_vs_nq_dest_overhead(dest);
/* return the server directly if it is idle */
if (atomic_read(&dest->activeconns) == 0) {
least = dest;
loh = doh;
goto out;
}
if (!least ||
((__s64)loh * atomic_read(&dest->weight) >
(__s64)doh * atomic_read(&least->weight))) {
least = dest;
loh = doh;
}
}
if (!least) {
ip_vs_scheduler_err(svc, "no destination available");
return NULL;
}
out:
IP_VS_DBG_BUF(6, "NQ: server %s:%u "
"activeconns %d refcnt %d weight %d overhead %d\n",
IP_VS_DBG_ADDR(least->af, &least->addr),
ntohs(least->port),
atomic_read(&least->activeconns),
refcount_read(&least->refcnt),
atomic_read(&least->weight), loh);
return least;
}
static struct ip_vs_scheduler ip_vs_nq_scheduler =
{
.name = "nq",
.refcnt = ATOMIC_INIT(0),
.module = THIS_MODULE,
.n_list = LIST_HEAD_INIT(ip_vs_nq_scheduler.n_list),
.schedule = ip_vs_nq_schedule,
};
static int __init ip_vs_nq_init(void)
{
return register_ip_vs_scheduler(&ip_vs_nq_scheduler);
}
static void __exit ip_vs_nq_cleanup(void)
{
unregister_ip_vs_scheduler(&ip_vs_nq_scheduler);
synchronize_rcu();
}
module_init(ip_vs_nq_init);
module_exit(ip_vs_nq_cleanup);
MODULE_LICENSE("GPL");