Merge branch 'master' of git://dev.medozas.de/linux
[linux-2.6] / net / core / drop_monitor.c
1 /*
2  * Monitoring code for network dropped packet alerts
3  *
4  * Copyright (C) 2009 Neil Horman <nhorman@tuxdriver.com>
5  */
6
7 #include <linux/netdevice.h>
8 #include <linux/etherdevice.h>
9 #include <linux/string.h>
10 #include <linux/if_arp.h>
11 #include <linux/inetdevice.h>
12 #include <linux/inet.h>
13 #include <linux/interrupt.h>
14 #include <linux/netpoll.h>
15 #include <linux/sched.h>
16 #include <linux/delay.h>
17 #include <linux/types.h>
18 #include <linux/workqueue.h>
19 #include <linux/netlink.h>
20 #include <linux/net_dropmon.h>
21 #include <linux/percpu.h>
22 #include <linux/timer.h>
23 #include <linux/bitops.h>
24 #include <net/genetlink.h>
25 #include <net/netevent.h>
26
27 #include <trace/skb.h>
28 #include <trace/napi.h>
29
30 #include <asm/unaligned.h>
31
32 #define TRACE_ON 1
33 #define TRACE_OFF 0
34
35 static void send_dm_alert(struct work_struct *unused);
36
37
38 /*
39  * Globals, our netlink socket pointer
40  * and the work handle that will send up
41  * netlink alerts
42  */
43 static int trace_state = TRACE_OFF;
44 static spinlock_t trace_state_lock = SPIN_LOCK_UNLOCKED;
45
46 struct per_cpu_dm_data {
47         struct work_struct dm_alert_work;
48         struct sk_buff *skb;
49         atomic_t dm_hit_count;
50         struct timer_list send_timer;
51 };
52
53 struct dm_hw_stat_delta {
54         struct net_device *dev;
55         struct list_head list;
56         struct rcu_head rcu;
57         unsigned long last_drop_val;
58 };
59
60 static struct genl_family net_drop_monitor_family = {
61         .id             = GENL_ID_GENERATE,
62         .hdrsize        = 0,
63         .name           = "NET_DM",
64         .version        = 2,
65         .maxattr        = NET_DM_CMD_MAX,
66 };
67
68 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
69
70 static int dm_hit_limit = 64;
71 static int dm_delay = 1;
72 static unsigned long dm_hw_check_delta = 2*HZ;
73 static LIST_HEAD(hw_stats_list);
74
75 static void reset_per_cpu_data(struct per_cpu_dm_data *data)
76 {
77         size_t al;
78         struct net_dm_alert_msg *msg;
79         struct nlattr *nla;
80
81         al = sizeof(struct net_dm_alert_msg);
82         al += dm_hit_limit * sizeof(struct net_dm_drop_point);
83         al += sizeof(struct nlattr);
84
85         data->skb = genlmsg_new(al, GFP_KERNEL);
86         genlmsg_put(data->skb, 0, 0, &net_drop_monitor_family,
87                         0, NET_DM_CMD_ALERT);
88         nla = nla_reserve(data->skb, NLA_UNSPEC, sizeof(struct net_dm_alert_msg));
89         msg = nla_data(nla);
90         memset(msg, 0, al);
91         atomic_set(&data->dm_hit_count, dm_hit_limit);
92 }
93
94 static void send_dm_alert(struct work_struct *unused)
95 {
96         struct sk_buff *skb;
97         struct per_cpu_dm_data *data = &__get_cpu_var(dm_cpu_data);
98
99         /*
100          * Grab the skb we're about to send
101          */
102         skb = data->skb;
103
104         /*
105          * Replace it with a new one
106          */
107         reset_per_cpu_data(data);
108
109         /*
110          * Ship it!
111          */
112         genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL);
113
114 }
115
116 /*
117  * This is the timer function to delay the sending of an alert
118  * in the event that more drops will arrive during the
119  * hysteresis period.  Note that it operates under the timer interrupt
120  * so we don't need to disable preemption here
121  */
122 static void sched_send_work(unsigned long unused)
123 {
124         struct per_cpu_dm_data *data =  &__get_cpu_var(dm_cpu_data);
125
126         schedule_work(&data->dm_alert_work);
127 }
128
129 static void trace_drop_common(struct sk_buff *skb, void *location)
130 {
131         struct net_dm_alert_msg *msg;
132         struct nlmsghdr *nlh;
133         struct nlattr *nla;
134         int i;
135         struct per_cpu_dm_data *data = &__get_cpu_var(dm_cpu_data);
136
137
138         if (!atomic_add_unless(&data->dm_hit_count, -1, 0)) {
139                 /*
140                  * we're already at zero, discard this hit
141                  */
142                 goto out;
143         }
144
145         nlh = (struct nlmsghdr *)data->skb->data;
146         nla = genlmsg_data(nlmsg_data(nlh));
147         msg = nla_data(nla);
148         for (i = 0; i < msg->entries; i++) {
149                 if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
150                         msg->points[i].count++;
151                         goto out;
152                 }
153         }
154
155         /*
156          * We need to create a new entry
157          */
158         __nla_reserve_nohdr(data->skb, sizeof(struct net_dm_drop_point));
159         nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
160         memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
161         msg->points[msg->entries].count = 1;
162         msg->entries++;
163
164         if (!timer_pending(&data->send_timer)) {
165                 data->send_timer.expires = jiffies + dm_delay * HZ;
166                 add_timer_on(&data->send_timer, smp_processor_id());
167         }
168
169 out:
170         return;
171 }
172
173 static void trace_kfree_skb_hit(struct sk_buff *skb, void *location)
174 {
175         trace_drop_common(skb, location);
176 }
177
178 static void trace_napi_poll_hit(struct napi_struct *napi)
179 {
180         struct dm_hw_stat_delta *new_stat;
181
182         /*
183          * Ratelimit our check time to dm_hw_check_delta jiffies
184          */
185         if (!time_after(jiffies, napi->dev->last_rx + dm_hw_check_delta))
186                 return;
187
188         rcu_read_lock();
189         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
190                 if ((new_stat->dev == napi->dev)  &&
191                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
192                         trace_drop_common(NULL, NULL);
193                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
194                         break;
195                 }
196         }
197         rcu_read_unlock();
198 }
199
200
201 static void free_dm_hw_stat(struct rcu_head *head)
202 {
203         struct dm_hw_stat_delta *n;
204         n = container_of(head, struct dm_hw_stat_delta, rcu);
205         kfree(n);
206 }
207
208 static int set_all_monitor_traces(int state)
209 {
210         int rc = 0;
211         struct dm_hw_stat_delta *new_stat = NULL;
212         struct dm_hw_stat_delta *temp;
213
214         spin_lock(&trace_state_lock);
215
216         switch (state) {
217         case TRACE_ON:
218                 rc |= register_trace_kfree_skb(trace_kfree_skb_hit);
219                 rc |= register_trace_napi_poll(trace_napi_poll_hit);
220                 break;
221         case TRACE_OFF:
222                 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit);
223                 rc |= unregister_trace_napi_poll(trace_napi_poll_hit);
224
225                 tracepoint_synchronize_unregister();
226
227                 /*
228                  * Clean the device list
229                  */
230                 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
231                         if (new_stat->dev == NULL) {
232                                 list_del_rcu(&new_stat->list);
233                                 call_rcu(&new_stat->rcu, free_dm_hw_stat);
234                         }
235                 }
236                 break;
237         default:
238                 rc = 1;
239                 break;
240         }
241
242         if (!rc)
243                 trace_state = state;
244
245         spin_unlock(&trace_state_lock);
246
247         if (rc)
248                 return -EINPROGRESS;
249         return rc;
250 }
251
252
253 static int net_dm_cmd_config(struct sk_buff *skb,
254                         struct genl_info *info)
255 {
256         return -ENOTSUPP;
257 }
258
259 static int net_dm_cmd_trace(struct sk_buff *skb,
260                         struct genl_info *info)
261 {
262         switch (info->genlhdr->cmd) {
263         case NET_DM_CMD_START:
264                 return set_all_monitor_traces(TRACE_ON);
265                 break;
266         case NET_DM_CMD_STOP:
267                 return set_all_monitor_traces(TRACE_OFF);
268                 break;
269         }
270
271         return -ENOTSUPP;
272 }
273
274 static int dropmon_net_event(struct notifier_block *ev_block,
275                         unsigned long event, void *ptr)
276 {
277         struct net_device *dev = ptr;
278         struct dm_hw_stat_delta *new_stat = NULL;
279         struct dm_hw_stat_delta *tmp;
280
281         switch (event) {
282         case NETDEV_REGISTER:
283                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
284
285                 if (!new_stat)
286                         goto out;
287
288                 new_stat->dev = dev;
289                 INIT_RCU_HEAD(&new_stat->rcu);
290                 spin_lock(&trace_state_lock);
291                 list_add_rcu(&new_stat->list, &hw_stats_list);
292                 spin_unlock(&trace_state_lock);
293                 break;
294         case NETDEV_UNREGISTER:
295                 spin_lock(&trace_state_lock);
296                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
297                         if (new_stat->dev == dev) {
298                                 new_stat->dev = NULL;
299                                 if (trace_state == TRACE_OFF) {
300                                         list_del_rcu(&new_stat->list);
301                                         call_rcu(&new_stat->rcu, free_dm_hw_stat);
302                                         break;
303                                 }
304                         }
305                 }
306                 spin_unlock(&trace_state_lock);
307                 break;
308         }
309 out:
310         return NOTIFY_DONE;
311 }
312
313 static struct genl_ops dropmon_ops[] = {
314         {
315                 .cmd = NET_DM_CMD_CONFIG,
316                 .doit = net_dm_cmd_config,
317         },
318         {
319                 .cmd = NET_DM_CMD_START,
320                 .doit = net_dm_cmd_trace,
321         },
322         {
323                 .cmd = NET_DM_CMD_STOP,
324                 .doit = net_dm_cmd_trace,
325         },
326 };
327
328 static struct notifier_block dropmon_net_notifier = {
329         .notifier_call = dropmon_net_event
330 };
331
332 static int __init init_net_drop_monitor(void)
333 {
334         int cpu;
335         int rc, i, ret;
336         struct per_cpu_dm_data *data;
337         printk(KERN_INFO "Initalizing network drop monitor service\n");
338
339         if (sizeof(void *) > 8) {
340                 printk(KERN_ERR "Unable to store program counters on this arch, Drop monitor failed\n");
341                 return -ENOSPC;
342         }
343
344         if (genl_register_family(&net_drop_monitor_family) < 0) {
345                 printk(KERN_ERR "Could not create drop monitor netlink family\n");
346                 return -EFAULT;
347         }
348
349         rc = -EFAULT;
350
351         for (i = 0; i < ARRAY_SIZE(dropmon_ops); i++) {
352                 ret = genl_register_ops(&net_drop_monitor_family,
353                                         &dropmon_ops[i]);
354                 if (ret) {
355                         printk(KERN_CRIT "Failed to register operation %d\n",
356                                 dropmon_ops[i].cmd);
357                         goto out_unreg;
358                 }
359         }
360
361         rc = register_netdevice_notifier(&dropmon_net_notifier);
362         if (rc < 0) {
363                 printk(KERN_CRIT "Failed to register netdevice notifier\n");
364                 goto out_unreg;
365         }
366
367         rc = 0;
368
369         for_each_present_cpu(cpu) {
370                 data = &per_cpu(dm_cpu_data, cpu);
371                 reset_per_cpu_data(data);
372                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
373                 init_timer(&data->send_timer);
374                 data->send_timer.data = cpu;
375                 data->send_timer.function = sched_send_work;
376         }
377
378         goto out;
379
380 out_unreg:
381         genl_unregister_family(&net_drop_monitor_family);
382 out:
383         return rc;
384 }
385
386 late_initcall(init_net_drop_monitor);