Merge with /pub/scm/linux/kernel/git/torvalds/linux-2.6.git
[linux-2.6] / 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/module.h>
35 #include <linux/mroute.h>
36 #include <linux/init.h>
37 #include <net/ip.h>
38 #include <net/protocol.h>
39 #include <linux/skbuff.h>
40 #include <net/sock.h>
41 #include <net/icmp.h>
42 #include <net/udp.h>
43 #include <net/raw.h>
44 #include <linux/notifier.h>
45 #include <linux/if_arp.h>
46 #include <linux/netfilter_ipv4.h>
47 #include <net/ipip.h>
48 #include <net/checksum.h>
49 #include <net/ip_mp_alg.h>
50
51 struct multipath_device {
52         int             ifi; /* interface index of device */
53         atomic_t        usecount;
54         int             allocated;
55 };
56
57 #define MULTIPATH_MAX_DEVICECANDIDATES 10
58
59 static struct multipath_device state[MULTIPATH_MAX_DEVICECANDIDATES];
60 static DEFINE_SPINLOCK(state_lock);
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 static struct notifier_block drr_dev_notifier = {
111         .notifier_call  = drr_dev_event,
112 };
113
114
115 static void drr_safe_inc(atomic_t *usecount)
116 {
117         int n;
118
119         atomic_inc(usecount);
120
121         n = atomic_read(usecount);
122         if (n <= 0) {
123                 int i;
124
125                 spin_lock_bh(&state_lock);
126
127                 for (i = 0; i < MULTIPATH_MAX_DEVICECANDIDATES; i++)
128                         atomic_set(&state[i].usecount, 0);
129
130                 spin_unlock_bh(&state_lock);
131         }
132 }
133
134 static void drr_select_route(const struct flowi *flp,
135                              struct rtable *first, struct rtable **rp)
136 {
137         struct rtable *nh, *result, *cur_min;
138         int min_usecount = -1; 
139         int devidx = -1;
140         int cur_min_devidx = -1;
141
142         /* 1. make sure all alt. nexthops have the same GC related data */
143         /* 2. determine the new candidate to be returned */
144         result = NULL;
145         cur_min = NULL;
146         for (nh = rcu_dereference(first); nh;
147              nh = rcu_dereference(nh->u.rt_next)) {
148                 if ((nh->u.dst.flags & DST_BALANCED) != 0 &&
149                     multipath_comparekeys(&nh->fl, flp)) {
150                         int nh_ifidx = nh->u.dst.dev->ifindex;
151
152                         nh->u.dst.lastuse = jiffies;
153                         nh->u.dst.__use++;
154                         if (result != NULL)
155                                 continue;
156
157                         /* search for the output interface */
158
159                         /* this is not SMP safe, only add/remove are
160                          * SMP safe as wrong usecount updates have no big
161                          * impact
162                          */
163                         devidx = __multipath_finddev(nh_ifidx);
164                         if (devidx == -1) {
165                                 /* add the interface to the array 
166                                  * SMP safe
167                                  */
168                                 spin_lock_bh(&state_lock);
169
170                                 /* due to SMP: search again */
171                                 devidx = __multipath_finddev(nh_ifidx);
172                                 if (devidx == -1) {
173                                         /* add entry for device */
174                                         devidx = __multipath_findslot();
175                                         if (devidx == -1) {
176                                                 /* unlikely but possible */
177                                                 continue;
178                                         }
179
180                                         state[devidx].allocated = 1;
181                                         state[devidx].ifi = nh_ifidx;
182                                         atomic_set(&state[devidx].usecount, 0);
183                                         min_usecount = 0;
184                                 }
185
186                                 spin_unlock_bh(&state_lock);
187                         }
188
189                         if (min_usecount == 0) {
190                                 /* if the device has not been used it is
191                                  * the primary target
192                                  */
193                                 drr_safe_inc(&state[devidx].usecount);
194                                 result = nh;
195                         } else {
196                                 int count =
197                                         atomic_read(&state[devidx].usecount);
198
199                                 if (min_usecount == -1 ||
200                                     count < min_usecount) {
201                                         cur_min = nh;
202                                         cur_min_devidx = devidx;
203                                         min_usecount = count;
204                                 }
205                         }
206                 }
207         }
208
209         if (!result) {
210                 if (cur_min) {
211                         drr_safe_inc(&state[cur_min_devidx].usecount);
212                         result = cur_min;
213                 } else {
214                         result = first;
215                 }
216         }
217
218         *rp = result;
219 }
220
221 static struct ip_mp_alg_ops drr_ops = {
222         .mp_alg_select_route    =       drr_select_route,
223 };
224
225 static int __init drr_init(void)
226 {
227         int err = register_netdevice_notifier(&drr_dev_notifier);
228
229         if (err)
230                 return err;
231
232         err = multipath_alg_register(&drr_ops, IP_MP_ALG_DRR);
233         if (err)
234                 goto fail;
235
236         return 0;
237
238 fail:
239         unregister_netdevice_notifier(&drr_dev_notifier);
240         return err;
241 }
242
243 static void __exit drr_exit(void)
244 {
245         unregister_netdevice_notifier(&drr_dev_notifier);
246         multipath_alg_unregister(&drr_ops, IP_MP_ALG_DRR);
247 }
248
249 module_init(drr_init);
250 module_exit(drr_exit);
251 MODULE_LICENSE("GPL");