linux/net/netfilter/ipvs/ip_vs_wrr.c
<<
>>
Prefs
   1/*
   2 * IPVS:        Weighted Round-Robin Scheduling module
   3 *
   4 * Authors:     Wensong Zhang <wensong@linuxvirtualserver.org>
   5 *
   6 *              This program is free software; you can redistribute it and/or
   7 *              modify it under the terms of the GNU General Public License
   8 *              as published by the Free Software Foundation; either version
   9 *              2 of the License, or (at your option) any later version.
  10 *
  11 * Changes:
  12 *     Wensong Zhang            :     changed the ip_vs_wrr_schedule to return dest
  13 *     Wensong Zhang            :     changed some comestics things for debugging
  14 *     Wensong Zhang            :     changed for the d-linked destination list
  15 *     Wensong Zhang            :     added the ip_vs_wrr_update_svc
  16 *     Julian Anastasov         :     fixed the bug of returning destination
  17 *                                    with weight 0 when all weights are zero
  18 *
  19 */
  20
  21#define KMSG_COMPONENT "IPVS"
  22#define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  23
  24#include <linux/module.h>
  25#include <linux/kernel.h>
  26#include <linux/slab.h>
  27#include <linux/net.h>
  28#include <linux/gcd.h>
  29
  30#include <net/ip_vs.h>
  31
  32/* The WRR algorithm depends on some caclulations:
  33 * - mw: maximum weight
  34 * - di: weight step, greatest common divisor from all weights
  35 * - cw: current required weight
  36 * As result, all weights are in the [di..mw] range with a step=di.
  37 *
  38 * First, we start with cw = mw and select dests with weight >= cw.
  39 * Then cw is reduced with di and all dests are checked again.
  40 * Last pass should be with cw = di. We have mw/di passes in total:
  41 *
  42 * pass 1: cw = max weight
  43 * pass 2: cw = max weight - di
  44 * pass 3: cw = max weight - 2 * di
  45 * ...
  46 * last pass: cw = di
  47 *
  48 * Weights are supposed to be >= di but we run in parallel with
  49 * weight changes, it is possible some dest weight to be reduced
  50 * below di, bad if it is the only available dest.
  51 *
  52 * So, we modify how mw is calculated, now it is reduced with (di - 1),
  53 * so that last cw is 1 to catch such dests with weight below di:
  54 * pass 1: cw = max weight - (di - 1)
  55 * pass 2: cw = max weight - di - (di - 1)
  56 * pass 3: cw = max weight - 2 * di - (di - 1)
  57 * ...
  58 * last pass: cw = 1
  59 *
  60 */
  61
  62/*
  63 * current destination pointer for weighted round-robin scheduling
  64 */
  65struct ip_vs_wrr_mark {
  66        struct ip_vs_dest *cl;  /* current dest or head */
  67        int cw;                 /* current weight */
  68        int mw;                 /* maximum weight */
  69        int di;                 /* decreasing interval */
  70        struct rcu_head         rcu_head;
  71};
  72
  73
  74static int ip_vs_wrr_gcd_weight(struct ip_vs_service *svc)
  75{
  76        struct ip_vs_dest *dest;
  77        int weight;
  78        int g = 0;
  79
  80        list_for_each_entry(dest, &svc->destinations, n_list) {
  81                weight = atomic_read(&dest->weight);
  82                if (weight > 0) {
  83                        if (g > 0)
  84                                g = gcd(weight, g);
  85                        else
  86                                g = weight;
  87                }
  88        }
  89        return g ? g : 1;
  90}
  91
  92
  93/*
  94 *    Get the maximum weight of the service destinations.
  95 */
  96static int ip_vs_wrr_max_weight(struct ip_vs_service *svc)
  97{
  98        struct ip_vs_dest *dest;
  99        int new_weight, weight = 0;
 100
 101        list_for_each_entry(dest, &svc->destinations, n_list) {
 102                new_weight = atomic_read(&dest->weight);
 103                if (new_weight > weight)
 104                        weight = new_weight;
 105        }
 106
 107        return weight;
 108}
 109
 110
 111static int ip_vs_wrr_init_svc(struct ip_vs_service *svc)
 112{
 113        struct ip_vs_wrr_mark *mark;
 114
 115        /*
 116         *    Allocate the mark variable for WRR scheduling
 117         */
 118        mark = kmalloc(sizeof(struct ip_vs_wrr_mark), GFP_KERNEL);
 119        if (mark == NULL)
 120                return -ENOMEM;
 121
 122        mark->cl = list_entry(&svc->destinations, struct ip_vs_dest, n_list);
 123        mark->di = ip_vs_wrr_gcd_weight(svc);
 124        mark->mw = ip_vs_wrr_max_weight(svc) - (mark->di - 1);
 125        mark->cw = mark->mw;
 126        svc->sched_data = mark;
 127
 128        return 0;
 129}
 130
 131
 132static void ip_vs_wrr_done_svc(struct ip_vs_service *svc)
 133{
 134        struct ip_vs_wrr_mark *mark = svc->sched_data;
 135
 136        /*
 137         *    Release the mark variable
 138         */
 139        kfree_rcu(mark, rcu_head);
 140}
 141
 142
 143static int ip_vs_wrr_dest_changed(struct ip_vs_service *svc,
 144                                  struct ip_vs_dest *dest)
 145{
 146        struct ip_vs_wrr_mark *mark = svc->sched_data;
 147
 148        spin_lock_bh(&svc->sched_lock);
 149        mark->cl = list_entry(&svc->destinations, struct ip_vs_dest, n_list);
 150        mark->di = ip_vs_wrr_gcd_weight(svc);
 151        mark->mw = ip_vs_wrr_max_weight(svc) - (mark->di - 1);
 152        if (mark->cw > mark->mw || !mark->cw) {
 153                gmb();
 154                mark->cw = mark->mw;
 155        } else if (mark->di > 1) {
 156                gmb();
 157                mark->cw = (mark->cw / mark->di) * mark->di + 1;
 158        }
 159        spin_unlock_bh(&svc->sched_lock);
 160        return 0;
 161}
 162
 163
 164/*
 165 *    Weighted Round-Robin Scheduling
 166 */
 167static struct ip_vs_dest *
 168ip_vs_wrr_schedule(struct ip_vs_service *svc, const struct sk_buff *skb,
 169                   struct ip_vs_iphdr *iph)
 170{
 171        struct ip_vs_dest *dest, *last, *stop = NULL;
 172        struct ip_vs_wrr_mark *mark = svc->sched_data;
 173        bool last_pass = false, restarted = false;
 174
 175        IP_VS_DBG(6, "%s(): Scheduling...\n", __func__);
 176
 177        spin_lock_bh(&svc->sched_lock);
 178        dest = mark->cl;
 179        /* No available dests? */
 180        if (mark->mw == 0)
 181                goto err_noavail;
 182        last = dest;
 183        /* Stop only after all dests were checked for weight >= 1 (last pass) */
 184        while (1) {
 185                list_for_each_entry_continue_rcu(dest,
 186                                                 &svc->destinations,
 187                                                 n_list) {
 188                        if (!(dest->flags & IP_VS_DEST_F_OVERLOAD) &&
 189                            atomic_read(&dest->weight) >= mark->cw)
 190                                goto found;
 191                        if (dest == stop)
 192                                goto err_over;
 193                }
 194                mark->cw -= mark->di;
 195                if (mark->cw <= 0) {
 196                        mark->cw = mark->mw;
 197                        /* Stop if we tried last pass from first dest:
 198                         * 1. last_pass: we started checks when cw > di but
 199                         *      then all dests were checked for w >= 1
 200                         * 2. last was head: the first and only traversal
 201                         *      was for weight >= 1, for all dests.
 202                         */
 203                        if (last_pass ||
 204                            &last->n_list == &svc->destinations)
 205                                goto err_over;
 206                        restarted = true;
 207                }
 208                last_pass = mark->cw <= mark->di;
 209                if (last_pass && restarted &&
 210                    &last->n_list != &svc->destinations) {
 211                        /* First traversal was for w >= 1 but only
 212                         * for dests after 'last', now do the same
 213                         * for all dests up to 'last'.
 214                         */
 215                        stop = last;
 216                }
 217        }
 218
 219found:
 220        IP_VS_DBG_BUF(6, "WRR: server %s:%u "
 221                      "activeconns %d refcnt %d weight %d\n",
 222                      IP_VS_DBG_ADDR(svc->af, &dest->addr), ntohs(dest->port),
 223                      atomic_read(&dest->activeconns),
 224                      atomic_read(&dest->refcnt),
 225                      atomic_read(&dest->weight));
 226        mark->cl = dest;
 227
 228  out:
 229        spin_unlock_bh(&svc->sched_lock);
 230        return dest;
 231
 232err_noavail:
 233        mark->cl = dest;
 234        dest = NULL;
 235        ip_vs_scheduler_err(svc, "no destination available");
 236        goto out;
 237
 238err_over:
 239        mark->cl = dest;
 240        dest = NULL;
 241        ip_vs_scheduler_err(svc, "no destination available: "
 242                            "all destinations are overloaded");
 243        goto out;
 244}
 245
 246
 247static struct ip_vs_scheduler ip_vs_wrr_scheduler = {
 248        .name =                 "wrr",
 249        .refcnt =               ATOMIC_INIT(0),
 250        .module =               THIS_MODULE,
 251        .n_list =               LIST_HEAD_INIT(ip_vs_wrr_scheduler.n_list),
 252        .init_service =         ip_vs_wrr_init_svc,
 253        .done_service =         ip_vs_wrr_done_svc,
 254        .add_dest =             ip_vs_wrr_dest_changed,
 255        .del_dest =             ip_vs_wrr_dest_changed,
 256        .upd_dest =             ip_vs_wrr_dest_changed,
 257        .schedule =             ip_vs_wrr_schedule,
 258};
 259
 260static int __init ip_vs_wrr_init(void)
 261{
 262        return register_ip_vs_scheduler(&ip_vs_wrr_scheduler) ;
 263}
 264
 265static void __exit ip_vs_wrr_cleanup(void)
 266{
 267        unregister_ip_vs_scheduler(&ip_vs_wrr_scheduler);
 268        synchronize_rcu();
 269}
 270
 271module_init(ip_vs_wrr_init);
 272module_exit(ip_vs_wrr_cleanup);
 273MODULE_LICENSE("GPL");
 274