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