net/ipv4/multipath_drr.c

   1 /*
   2  *              Device round robin policy for multipath.
   3  *
   4  *
   5  * Version:     $Id: multipath_drr.c,v 1.1.2.1 2004/09/16 07:42:34 elueck Exp $
   6  *
   7  * Authors:     Einar Lueck <elueck@de.ibm.com><lkml@einar-lueck.de>
   8  *
   9  *              This program is free software; you can redistribute it and/or
  10  *              modify it under the terms of the GNU General Public License
  11  *              as published by the Free Software Foundation; either version
  12  *              2 of the License, or (at your option) any later version.
  13  */
  14
  15 #include <asm/system.h>
  16 #include <asm/uaccess.h>
  17 #include <linux/types.h>
  18 #include <linux/sched.h>
  19 #include <linux/errno.h>
  20 #include <linux/timer.h>
  21 #include <linux/mm.h>
  22 #include <linux/kernel.h>
  23 #include <linux/fcntl.h>
  24 #include <linux/stat.h>
  25 #include <linux/socket.h>
  26 #include <linux/in.h>
  27 #include <linux/inet.h>
  28 #include <linux/netdevice.h>
  29 #include <linux/inetdevice.h>
  30 #include <linux/igmp.h>
  31 #include <linux/proc_fs.h>
  32 #include <linux/seq_file.h>
  33 #include <linux/module.h>
  34 #include <linux/mroute.h>
  35 #include <linux/init.h>
  36 #include <net/ip.h>
  37 #include <net/protocol.h>
  38 #include <linux/skbuff.h>
  39 #include <net/sock.h>
  40 #include <net/icmp.h>
  41 #include <net/udp.h>
  42 #include <net/raw.h>
  43 #include <linux/notifier.h>
  44 #include <linux/if_arp.h>
  45 #include <linux/netfilter_ipv4.h>
  46 #include <net/ipip.h>
  47 #include <net/checksum.h>
  48 #include <net/ip_mp_alg.h>
  49
  50 struct multipath_device {
  51         int             ifi; /* interface index of device */
  52         atomic_t        usecount;
  53         int             allocated;
  54 };
  55
  56 #define MULTIPATH_MAX_DEVICECANDIDATES 10
  57
  58 static struct multipath_device state[MULTIPATH_MAX_DEVICECANDIDATES];
  59 static DEFINE_SPINLOCK(state_lock);
  60
  61 static int inline __multipath_findslot(void)
  62 {
  63         int i;
  64
  65         for (i = 0; i < MULTIPATH_MAX_DEVICECANDIDATES; i++) {
  66                 if (state[i].allocated == 0)
  67                         return i;
  68         }
  69         return -1;
  70 }
  71
  72 static int inline __multipath_finddev(int ifindex)
  73 {
  74         int i;
  75
  76         for (i = 0; i < MULTIPATH_MAX_DEVICECANDIDATES; i++) {
  77                 if (state[i].allocated != 0 &&
  78                     state[i].ifi == ifindex)
  79                         return i;
  80         }
  81         return -1;
  82 }
  83
  84 static int drr_dev_event(struct notifier_block *this,
  85                          unsigned long event, void *ptr)
  86 {
  87         struct net_device *dev = ptr;
  88         int devidx;
  89
  90         switch (event) {
  91         case NETDEV_UNREGISTER:
  92         case NETDEV_DOWN:
  93                 spin_lock_bh(&state_lock);
  94
  95                 devidx = __multipath_finddev(dev->ifindex);
  96                 if (devidx != -1) {
  97                         state[devidx].allocated = 0;
  98                         state[devidx].ifi = 0;
  99                         atomic_set(&state[devidx].usecount, 0);
 100                 }
 101
 102                 spin_unlock_bh(&state_lock);
 103                 break;
 104         };
 105
 106         return NOTIFY_DONE;
 107 }
 108
 109 static struct notifier_block drr_dev_notifier = {
 110         .notifier_call  = drr_dev_event,
 111 };
 112
 113
 114 static void drr_safe_inc(atomic_t *usecount)
 115 {
 116         int n;
 117
 118         atomic_inc(usecount);
 119
 120         n = atomic_read(usecount);
 121         if (n <= 0) {
 122                 int i;
 123
 124                 spin_lock_bh(&state_lock);
 125
 126                 for (i = 0; i < MULTIPATH_MAX_DEVICECANDIDATES; i++)
 127                         atomic_set(&state[i].usecount, 0);
 128
 129                 spin_unlock_bh(&state_lock);
 130         }
 131 }
 132
 133 static void drr_select_route(const struct flowi *flp,
 134                              struct rtable *first, struct rtable **rp)
 135 {
 136         struct rtable *nh, *result, *cur_min;
 137         int min_usecount = -1;
 138         int devidx = -1;
 139         int cur_min_devidx = -1;
 140
 141         /* 1. make sure all alt. nexthops have the same GC related data */
 142         /* 2. determine the new candidate to be returned */
 143         result = NULL;
 144         cur_min = NULL;
 145         for (nh = rcu_dereference(first); nh;
 146              nh = rcu_dereference(nh->u.dst.rt_next)) {
 147                 if ((nh->u.dst.flags & DST_BALANCED) != 0 &&
 148                     multipath_comparekeys(&nh->fl, flp)) {
 149                         int nh_ifidx = nh->u.dst.dev->ifindex;
 150
 151                         nh->u.dst.lastuse = jiffies;
 152                         nh->u.dst.__use++;
 153                         if (result != NULL)
 154                                 continue;
 155
 156                         /* search for the output interface */
 157
 158                         /* this is not SMP safe, only add/remove are
 159                          * SMP safe as wrong usecount updates have no big
 160                          * impact
 161                          */
 162                         devidx = __multipath_finddev(nh_ifidx);
 163                         if (devidx == -1) {
 164                                 /* add the interface to the array
 165                                  * SMP safe
 166                                  */
 167                                 spin_lock_bh(&state_lock);
 168
 169                                 /* due to SMP: search again */
 170                                 devidx = __multipath_finddev(nh_ifidx);
 171                                 if (devidx == -1) {
 172                                         /* add entry for device */
 173                                         devidx = __multipath_findslot();
 174                                         if (devidx == -1) {
 175                                                 /* unlikely but possible */
 176                                                 continue;
 177                                         }
 178
 179                                         state[devidx].allocated = 1;
 180                                         state[devidx].ifi = nh_ifidx;
 181                                         atomic_set(&state[devidx].usecount, 0);
 182                                         min_usecount = 0;
 183                                 }
 184
 185                                 spin_unlock_bh(&state_lock);
 186                         }
 187
 188                         if (min_usecount == 0) {
 189                                 /* if the device has not been used it is
 190                                  * the primary target
 191                                  */
 192                                 drr_safe_inc(&state[devidx].usecount);
 193                                 result = nh;
 194                         } else {
 195                                 int count =
 196                                         atomic_read(&state[devidx].usecount);
 197
 198                                 if (min_usecount == -1 ||
 199                                     count < min_usecount) {
 200                                         cur_min = nh;
 201                                         cur_min_devidx = devidx;
 202                                         min_usecount = count;
 203                                 }
 204                         }
 205                 }
 206         }
 207
 208         if (!result) {
 209                 if (cur_min) {
 210                         drr_safe_inc(&state[cur_min_devidx].usecount);
 211                         result = cur_min;
 212                 } else {
 213                         result = first;
 214                 }
 215         }
 216
 217         *rp = result;
 218 }
 219
 220 static struct ip_mp_alg_ops drr_ops = {
 221         .mp_alg_select_route    =       drr_select_route,
 222 };
 223
 224 static int __init drr_init(void)
 225 {
 226         int err = register_netdevice_notifier(&drr_dev_notifier);
 227
 228         if (err)
 229                 return err;
 230
 231         err = multipath_alg_register(&drr_ops, IP_MP_ALG_DRR);
 232         if (err)
 233                 goto fail;
 234
 235         return 0;
 236
 237 fail:
 238         unregister_netdevice_notifier(&drr_dev_notifier);
 239         return err;
 240 }
 241
 242 static void __exit drr_exit(void)
 243 {
 244         unregister_netdevice_notifier(&drr_dev_notifier);
 245         multipath_alg_unregister(&drr_ops, IP_MP_ALG_DRR);
 246 }
 247
 248 module_init(drr_init);
 249 module_exit(drr_exit);
 250 MODULE_LICENSE("GPL");