xref: /linux/net/core/drop_monitor.c (revision b6ebbac51bedf9e98e837688bc838f400196da5e)
1 /*
2  * Monitoring code for network dropped packet alerts
3  *
4  * Copyright (C) 2009 Neil Horman <nhorman@tuxdriver.com>
5  */
6 
7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
8 
9 #include <linux/netdevice.h>
10 #include <linux/etherdevice.h>
11 #include <linux/string.h>
12 #include <linux/if_arp.h>
13 #include <linux/inetdevice.h>
14 #include <linux/inet.h>
15 #include <linux/interrupt.h>
16 #include <linux/netpoll.h>
17 #include <linux/sched.h>
18 #include <linux/delay.h>
19 #include <linux/types.h>
20 #include <linux/workqueue.h>
21 #include <linux/netlink.h>
22 #include <linux/net_dropmon.h>
23 #include <linux/percpu.h>
24 #include <linux/timer.h>
25 #include <linux/bitops.h>
26 #include <linux/slab.h>
27 #include <linux/module.h>
28 #include <net/genetlink.h>
29 #include <net/netevent.h>
30 
31 #include <trace/events/skb.h>
32 #include <trace/events/napi.h>
33 
34 #include <asm/unaligned.h>
35 
36 #define TRACE_ON 1
37 #define TRACE_OFF 0
38 
39 /*
40  * Globals, our netlink socket pointer
41  * and the work handle that will send up
42  * netlink alerts
43  */
44 static int trace_state = TRACE_OFF;
45 static DEFINE_MUTEX(trace_state_mutex);
46 
47 struct per_cpu_dm_data {
48 	spinlock_t		lock;
49 	struct sk_buff		*skb;
50 	struct work_struct	dm_alert_work;
51 	struct timer_list	send_timer;
52 };
53 
54 struct dm_hw_stat_delta {
55 	struct net_device *dev;
56 	unsigned long last_rx;
57 	struct list_head list;
58 	struct rcu_head rcu;
59 	unsigned long last_drop_val;
60 };
61 
62 static struct genl_family net_drop_monitor_family = {
63 	.id             = GENL_ID_GENERATE,
64 	.hdrsize        = 0,
65 	.name           = "NET_DM",
66 	.version        = 2,
67 };
68 
69 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
70 
71 static int dm_hit_limit = 64;
72 static int dm_delay = 1;
73 static unsigned long dm_hw_check_delta = 2*HZ;
74 static LIST_HEAD(hw_stats_list);
75 
76 static struct sk_buff *reset_per_cpu_data(struct per_cpu_dm_data *data)
77 {
78 	size_t al;
79 	struct net_dm_alert_msg *msg;
80 	struct nlattr *nla;
81 	struct sk_buff *skb;
82 	unsigned long flags;
83 
84 	al = sizeof(struct net_dm_alert_msg);
85 	al += dm_hit_limit * sizeof(struct net_dm_drop_point);
86 	al += sizeof(struct nlattr);
87 
88 	skb = genlmsg_new(al, GFP_KERNEL);
89 
90 	if (skb) {
91 		genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
92 				0, NET_DM_CMD_ALERT);
93 		nla = nla_reserve(skb, NLA_UNSPEC,
94 				  sizeof(struct net_dm_alert_msg));
95 		msg = nla_data(nla);
96 		memset(msg, 0, al);
97 	} else {
98 		mod_timer(&data->send_timer, jiffies + HZ / 10);
99 	}
100 
101 	spin_lock_irqsave(&data->lock, flags);
102 	swap(data->skb, skb);
103 	spin_unlock_irqrestore(&data->lock, flags);
104 
105 	return skb;
106 }
107 
108 static struct genl_multicast_group dropmon_mcgrps[] = {
109 	{ .name = "events", },
110 };
111 
112 static void send_dm_alert(struct work_struct *work)
113 {
114 	struct sk_buff *skb;
115 	struct per_cpu_dm_data *data;
116 
117 	data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
118 
119 	skb = reset_per_cpu_data(data);
120 
121 	if (skb)
122 		genlmsg_multicast(&net_drop_monitor_family, skb, 0,
123 				  0, GFP_KERNEL);
124 }
125 
126 /*
127  * This is the timer function to delay the sending of an alert
128  * in the event that more drops will arrive during the
129  * hysteresis period.
130  */
131 static void sched_send_work(unsigned long _data)
132 {
133 	struct per_cpu_dm_data *data = (struct per_cpu_dm_data *)_data;
134 
135 	schedule_work(&data->dm_alert_work);
136 }
137 
138 static void trace_drop_common(struct sk_buff *skb, void *location)
139 {
140 	struct net_dm_alert_msg *msg;
141 	struct nlmsghdr *nlh;
142 	struct nlattr *nla;
143 	int i;
144 	struct sk_buff *dskb;
145 	struct per_cpu_dm_data *data;
146 	unsigned long flags;
147 
148 	local_irq_save(flags);
149 	data = this_cpu_ptr(&dm_cpu_data);
150 	spin_lock(&data->lock);
151 	dskb = data->skb;
152 
153 	if (!dskb)
154 		goto out;
155 
156 	nlh = (struct nlmsghdr *)dskb->data;
157 	nla = genlmsg_data(nlmsg_data(nlh));
158 	msg = nla_data(nla);
159 	for (i = 0; i < msg->entries; i++) {
160 		if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
161 			msg->points[i].count++;
162 			goto out;
163 		}
164 	}
165 	if (msg->entries == dm_hit_limit)
166 		goto out;
167 	/*
168 	 * We need to create a new entry
169 	 */
170 	__nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
171 	nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
172 	memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
173 	msg->points[msg->entries].count = 1;
174 	msg->entries++;
175 
176 	if (!timer_pending(&data->send_timer)) {
177 		data->send_timer.expires = jiffies + dm_delay * HZ;
178 		add_timer(&data->send_timer);
179 	}
180 
181 out:
182 	spin_unlock_irqrestore(&data->lock, flags);
183 }
184 
185 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
186 {
187 	trace_drop_common(skb, location);
188 }
189 
190 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi,
191 				int work, int budget)
192 {
193 	struct dm_hw_stat_delta *new_stat;
194 
195 	/*
196 	 * Don't check napi structures with no associated device
197 	 */
198 	if (!napi->dev)
199 		return;
200 
201 	rcu_read_lock();
202 	list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
203 		/*
204 		 * only add a note to our monitor buffer if:
205 		 * 1) this is the dev we received on
206 		 * 2) its after the last_rx delta
207 		 * 3) our rx_dropped count has gone up
208 		 */
209 		if ((new_stat->dev == napi->dev)  &&
210 		    (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
211 		    (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
212 			trace_drop_common(NULL, NULL);
213 			new_stat->last_drop_val = napi->dev->stats.rx_dropped;
214 			new_stat->last_rx = jiffies;
215 			break;
216 		}
217 	}
218 	rcu_read_unlock();
219 }
220 
221 static int set_all_monitor_traces(int state)
222 {
223 	int rc = 0;
224 	struct dm_hw_stat_delta *new_stat = NULL;
225 	struct dm_hw_stat_delta *temp;
226 
227 	mutex_lock(&trace_state_mutex);
228 
229 	if (state == trace_state) {
230 		rc = -EAGAIN;
231 		goto out_unlock;
232 	}
233 
234 	switch (state) {
235 	case TRACE_ON:
236 		if (!try_module_get(THIS_MODULE)) {
237 			rc = -ENODEV;
238 			break;
239 		}
240 
241 		rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
242 		rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
243 		break;
244 
245 	case TRACE_OFF:
246 		rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
247 		rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
248 
249 		tracepoint_synchronize_unregister();
250 
251 		/*
252 		 * Clean the device list
253 		 */
254 		list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
255 			if (new_stat->dev == NULL) {
256 				list_del_rcu(&new_stat->list);
257 				kfree_rcu(new_stat, rcu);
258 			}
259 		}
260 
261 		module_put(THIS_MODULE);
262 
263 		break;
264 	default:
265 		rc = 1;
266 		break;
267 	}
268 
269 	if (!rc)
270 		trace_state = state;
271 	else
272 		rc = -EINPROGRESS;
273 
274 out_unlock:
275 	mutex_unlock(&trace_state_mutex);
276 
277 	return rc;
278 }
279 
280 
281 static int net_dm_cmd_config(struct sk_buff *skb,
282 			struct genl_info *info)
283 {
284 	return -ENOTSUPP;
285 }
286 
287 static int net_dm_cmd_trace(struct sk_buff *skb,
288 			struct genl_info *info)
289 {
290 	switch (info->genlhdr->cmd) {
291 	case NET_DM_CMD_START:
292 		return set_all_monitor_traces(TRACE_ON);
293 	case NET_DM_CMD_STOP:
294 		return set_all_monitor_traces(TRACE_OFF);
295 	}
296 
297 	return -ENOTSUPP;
298 }
299 
300 static int dropmon_net_event(struct notifier_block *ev_block,
301 			     unsigned long event, void *ptr)
302 {
303 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
304 	struct dm_hw_stat_delta *new_stat = NULL;
305 	struct dm_hw_stat_delta *tmp;
306 
307 	switch (event) {
308 	case NETDEV_REGISTER:
309 		new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
310 
311 		if (!new_stat)
312 			goto out;
313 
314 		new_stat->dev = dev;
315 		new_stat->last_rx = jiffies;
316 		mutex_lock(&trace_state_mutex);
317 		list_add_rcu(&new_stat->list, &hw_stats_list);
318 		mutex_unlock(&trace_state_mutex);
319 		break;
320 	case NETDEV_UNREGISTER:
321 		mutex_lock(&trace_state_mutex);
322 		list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
323 			if (new_stat->dev == dev) {
324 				new_stat->dev = NULL;
325 				if (trace_state == TRACE_OFF) {
326 					list_del_rcu(&new_stat->list);
327 					kfree_rcu(new_stat, rcu);
328 					break;
329 				}
330 			}
331 		}
332 		mutex_unlock(&trace_state_mutex);
333 		break;
334 	}
335 out:
336 	return NOTIFY_DONE;
337 }
338 
339 static const struct genl_ops dropmon_ops[] = {
340 	{
341 		.cmd = NET_DM_CMD_CONFIG,
342 		.doit = net_dm_cmd_config,
343 	},
344 	{
345 		.cmd = NET_DM_CMD_START,
346 		.doit = net_dm_cmd_trace,
347 	},
348 	{
349 		.cmd = NET_DM_CMD_STOP,
350 		.doit = net_dm_cmd_trace,
351 	},
352 };
353 
354 static struct notifier_block dropmon_net_notifier = {
355 	.notifier_call = dropmon_net_event
356 };
357 
358 static int __init init_net_drop_monitor(void)
359 {
360 	struct per_cpu_dm_data *data;
361 	int cpu, rc;
362 
363 	pr_info("Initializing network drop monitor service\n");
364 
365 	if (sizeof(void *) > 8) {
366 		pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
367 		return -ENOSPC;
368 	}
369 
370 	rc = genl_register_family_with_ops_groups(&net_drop_monitor_family,
371 						  dropmon_ops, dropmon_mcgrps);
372 	if (rc) {
373 		pr_err("Could not create drop monitor netlink family\n");
374 		return rc;
375 	}
376 	WARN_ON(net_drop_monitor_family.mcgrp_offset != NET_DM_GRP_ALERT);
377 
378 	rc = register_netdevice_notifier(&dropmon_net_notifier);
379 	if (rc < 0) {
380 		pr_crit("Failed to register netdevice notifier\n");
381 		goto out_unreg;
382 	}
383 
384 	rc = 0;
385 
386 	for_each_possible_cpu(cpu) {
387 		data = &per_cpu(dm_cpu_data, cpu);
388 		INIT_WORK(&data->dm_alert_work, send_dm_alert);
389 		init_timer(&data->send_timer);
390 		data->send_timer.data = (unsigned long)data;
391 		data->send_timer.function = sched_send_work;
392 		spin_lock_init(&data->lock);
393 		reset_per_cpu_data(data);
394 	}
395 
396 
397 	goto out;
398 
399 out_unreg:
400 	genl_unregister_family(&net_drop_monitor_family);
401 out:
402 	return rc;
403 }
404 
405 static void exit_net_drop_monitor(void)
406 {
407 	struct per_cpu_dm_data *data;
408 	int cpu;
409 
410 	BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
411 
412 	/*
413 	 * Because of the module_get/put we do in the trace state change path
414 	 * we are guarnateed not to have any current users when we get here
415 	 * all we need to do is make sure that we don't have any running timers
416 	 * or pending schedule calls
417 	 */
418 
419 	for_each_possible_cpu(cpu) {
420 		data = &per_cpu(dm_cpu_data, cpu);
421 		del_timer_sync(&data->send_timer);
422 		cancel_work_sync(&data->dm_alert_work);
423 		/*
424 		 * At this point, we should have exclusive access
425 		 * to this struct and can free the skb inside it
426 		 */
427 		kfree_skb(data->skb);
428 	}
429 
430 	BUG_ON(genl_unregister_family(&net_drop_monitor_family));
431 }
432 
433 module_init(init_net_drop_monitor);
434 module_exit(exit_net_drop_monitor);
435 
436 MODULE_LICENSE("GPL v2");
437 MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
438 MODULE_ALIAS_GENL_FAMILY("NET_DM");
439