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 <linux/config.h>
  16 #include <asm/system.h>
  17 #include <asm/uaccess.h>
  18 #include <linux/types.h>
  19 #include <linux/sched.h>
  20 #include <linux/errno.h>
  21 #include <linux/timer.h>
  22 #include <linux/mm.h>
  23 #include <linux/kernel.h>
  24 #include <linux/fcntl.h>
  25 #include <linux/stat.h>
  26 #include <linux/socket.h>
  27 #include <linux/in.h>
  28 #include <linux/inet.h>
  29 #include <linux/netdevice.h>
  30 #include <linux/inetdevice.h>
  31 #include <linux/igmp.h>
  32 #include <linux/proc_fs.h>
  33 #include <linux/seq_file.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 static struct rtable *last_selection = NULL;
  61
  62 static int inline __multipath_findslot(void)
  63 {
  64         int i;
  65
  66         for (i = 0; i < MULTIPATH_MAX_DEVICECANDIDATES; i++) {
  67                 if (state[i].allocated == 0)
  68                         return i;
  69         }
  70         return -1;
  71 }
  72
  73 static int inline __multipath_finddev(int ifindex)
  74 {
  75         int i;
  76
  77         for (i = 0; i < MULTIPATH_MAX_DEVICECANDIDATES; i++) {
  78                 if (state[i].allocated != 0 &&
  79                     state[i].ifi == ifindex)
  80                         return i;
  81         }
  82         return -1;
  83 }
  84
  85 static int drr_dev_event(struct notifier_block *this,
  86                          unsigned long event, void *ptr)
  87 {
  88         struct net_device *dev = ptr;
  89         int devidx;
  90
  91         switch (event) {
  92         case NETDEV_UNREGISTER:
  93         case NETDEV_DOWN:
  94                 spin_lock_bh(&state_lock);
  95
  96                 devidx = __multipath_finddev(dev->ifindex);
  97                 if (devidx != -1) {
  98                         state[devidx].allocated = 0;
  99                         state[devidx].ifi = 0;
 100                         atomic_set(&state[devidx].usecount, 0);
 101                 }
 102
 103                 spin_unlock_bh(&state_lock);
 104                 break;
 105         };
 106
 107         return NOTIFY_DONE;
 108 }
 109
 110 struct notifier_block drr_dev_notifier = {
 111         .notifier_call  = drr_dev_event,
 112 };
 113
 114 static void drr_remove(struct rtable *rt)
 115 {
 116         if (last_selection == rt)
 117                 last_selection = NULL;
 118 }
 119
 120 static void drr_safe_inc(atomic_t *usecount)
 121 {
 122         int n;
 123
 124         atomic_inc(usecount);
 125
 126         n = atomic_read(usecount);
 127         if (n <= 0) {
 128                 int i;
 129
 130                 spin_lock_bh(&state_lock);
 131
 132                 for (i = 0; i < MULTIPATH_MAX_DEVICECANDIDATES; i++)
 133                         atomic_set(&state[i].usecount, 0);
 134
 135                 spin_unlock_bh(&state_lock);
 136         }
 137 }
 138
 139 static void drr_select_route(const struct flowi *flp,
 140                              struct rtable *first, struct rtable **rp)
 141 {
 142         struct rtable *nh, *result, *cur_min;
 143         int min_usecount = -1;
 144         int devidx = -1;
 145         int cur_min_devidx = -1;
 146
 147         /* if necessary and possible utilize the old alternative */
 148         if ((flp->flags & FLOWI_FLAG_MULTIPATHOLDROUTE) != 0 &&
 149             last_selection != NULL) {
 150                 result = last_selection;
 151                 *rp = result;
 152                 return;
 153         }
 154
 155         /* 1. make sure all alt. nexthops have the same GC related data */
 156         /* 2. determine the new candidate to be returned */
 157         result = NULL;
 158         cur_min = NULL;
 159         for (nh = rcu_dereference(first); nh;
 160              nh = rcu_dereference(nh->u.rt_next)) {
 161                 if ((nh->u.dst.flags & DST_BALANCED) != 0 &&
 162                     multipath_comparekeys(&nh->fl, flp)) {
 163                         int nh_ifidx = nh->u.dst.dev->ifindex;
 164
 165                         nh->u.dst.lastuse = jiffies;
 166                         nh->u.dst.__use++;
 167                         if (result != NULL)
 168                                 continue;
 169
 170                         /* search for the output interface */
 171
 172                         /* this is not SMP safe, only add/remove are
 173                          * SMP safe as wrong usecount updates have no big
 174                          * impact
 175                          */
 176                         devidx = __multipath_finddev(nh_ifidx);
 177                         if (devidx == -1) {
 178                                 /* add the interface to the array
 179                                  * SMP safe
 180                                  */
 181                                 spin_lock_bh(&state_lock);
 182
 183                                 /* due to SMP: search again */
 184                                 devidx = __multipath_finddev(nh_ifidx);
 185                                 if (devidx == -1) {
 186                                         /* add entry for device */
 187                                         devidx = __multipath_findslot();
 188                                         if (devidx == -1) {
 189                                                 /* unlikely but possible */
 190                                                 continue;
 191                                         }
 192
 193                                         state[devidx].allocated = 1;
 194                                         state[devidx].ifi = nh_ifidx;
 195                                         atomic_set(&state[devidx].usecount, 0);
 196                                         min_usecount = 0;
 197                                 }
 198
 199                                 spin_unlock_bh(&state_lock);
 200                         }
 201
 202                         if (min_usecount == 0) {
 203                                 /* if the device has not been used it is
 204                                  * the primary target
 205                                  */
 206                                 drr_safe_inc(&state[devidx].usecount);
 207                                 result = nh;
 208                         } else {
 209                                 int count =
 210                                         atomic_read(&state[devidx].usecount);
 211
 212                                 if (min_usecount == -1 ||
 213                                     count < min_usecount) {
 214                                         cur_min = nh;
 215                                         cur_min_devidx = devidx;
 216                                         min_usecount = count;
 217                                 }
 218                         }
 219                 }
 220         }
 221
 222         if (!result) {
 223                 if (cur_min) {
 224                         drr_safe_inc(&state[cur_min_devidx].usecount);
 225                         result = cur_min;
 226                 } else {
 227                         result = first;
 228                 }
 229         }
 230
 231         *rp = result;
 232         last_selection = result;
 233 }
 234
 235 static struct ip_mp_alg_ops drr_ops = {
 236         .mp_alg_select_route    =       drr_select_route,
 237         .mp_alg_remove          =       drr_remove,
 238 };
 239
 240 static int __init drr_init(void)
 241 {
 242         int err = register_netdevice_notifier(&drr_dev_notifier);
 243
 244         if (err)
 245                 return err;
 246
 247         err = multipath_alg_register(&drr_ops, IP_MP_ALG_RR);
 248         if (err)
 249                 goto fail;
 250
 251         return 0;
 252
 253 fail:
 254         unregister_netdevice_notifier(&drr_dev_notifier);
 255         return err;
 256 }
 257
 258 static void __exit drr_exit(void)
 259 {
 260         unregister_netdevice_notifier(&drr_dev_notifier);
 261         multipath_alg_unregister(&drr_ops, IP_MP_ALG_DRR);
 262 }
 263
 264 module_init(drr_init);
 265 module_exit(drr_exit);