xref: /openbmc/linux/net/core/drop_monitor.c (revision 79f08d9e)
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 	.maxattr        = NET_DM_CMD_MAX,
68 };
69 
70 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
71 
72 static int dm_hit_limit = 64;
73 static int dm_delay = 1;
74 static unsigned long dm_hw_check_delta = 2*HZ;
75 static LIST_HEAD(hw_stats_list);
76 
77 static struct sk_buff *reset_per_cpu_data(struct per_cpu_dm_data *data)
78 {
79 	size_t al;
80 	struct net_dm_alert_msg *msg;
81 	struct nlattr *nla;
82 	struct sk_buff *skb;
83 	unsigned long flags;
84 
85 	al = sizeof(struct net_dm_alert_msg);
86 	al += dm_hit_limit * sizeof(struct net_dm_drop_point);
87 	al += sizeof(struct nlattr);
88 
89 	skb = genlmsg_new(al, GFP_KERNEL);
90 
91 	if (skb) {
92 		genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
93 				0, NET_DM_CMD_ALERT);
94 		nla = nla_reserve(skb, NLA_UNSPEC,
95 				  sizeof(struct net_dm_alert_msg));
96 		msg = nla_data(nla);
97 		memset(msg, 0, al);
98 	} else {
99 		mod_timer(&data->send_timer, jiffies + HZ / 10);
100 	}
101 
102 	spin_lock_irqsave(&data->lock, flags);
103 	swap(data->skb, skb);
104 	spin_unlock_irqrestore(&data->lock, flags);
105 
106 	return skb;
107 }
108 
109 static struct genl_multicast_group dropmon_mcgrps[] = {
110 	{ .name = "events", },
111 };
112 
113 static void send_dm_alert(struct work_struct *work)
114 {
115 	struct sk_buff *skb;
116 	struct per_cpu_dm_data *data;
117 
118 	data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
119 
120 	skb = reset_per_cpu_data(data);
121 
122 	if (skb)
123 		genlmsg_multicast(&net_drop_monitor_family, skb, 0,
124 				  0, GFP_KERNEL);
125 }
126 
127 /*
128  * This is the timer function to delay the sending of an alert
129  * in the event that more drops will arrive during the
130  * hysteresis period.
131  */
132 static void sched_send_work(unsigned long _data)
133 {
134 	struct per_cpu_dm_data *data = (struct per_cpu_dm_data *)_data;
135 
136 	schedule_work(&data->dm_alert_work);
137 }
138 
139 static void trace_drop_common(struct sk_buff *skb, void *location)
140 {
141 	struct net_dm_alert_msg *msg;
142 	struct nlmsghdr *nlh;
143 	struct nlattr *nla;
144 	int i;
145 	struct sk_buff *dskb;
146 	struct per_cpu_dm_data *data;
147 	unsigned long flags;
148 
149 	local_irq_save(flags);
150 	data = &__get_cpu_var(dm_cpu_data);
151 	spin_lock(&data->lock);
152 	dskb = data->skb;
153 
154 	if (!dskb)
155 		goto out;
156 
157 	nlh = (struct nlmsghdr *)dskb->data;
158 	nla = genlmsg_data(nlmsg_data(nlh));
159 	msg = nla_data(nla);
160 	for (i = 0; i < msg->entries; i++) {
161 		if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
162 			msg->points[i].count++;
163 			goto out;
164 		}
165 	}
166 	if (msg->entries == dm_hit_limit)
167 		goto out;
168 	/*
169 	 * We need to create a new entry
170 	 */
171 	__nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
172 	nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
173 	memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
174 	msg->points[msg->entries].count = 1;
175 	msg->entries++;
176 
177 	if (!timer_pending(&data->send_timer)) {
178 		data->send_timer.expires = jiffies + dm_delay * HZ;
179 		add_timer(&data->send_timer);
180 	}
181 
182 out:
183 	spin_unlock_irqrestore(&data->lock, flags);
184 }
185 
186 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
187 {
188 	trace_drop_common(skb, location);
189 }
190 
191 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi)
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 		break;
294 	case NET_DM_CMD_STOP:
295 		return set_all_monitor_traces(TRACE_OFF);
296 		break;
297 	}
298 
299 	return -ENOTSUPP;
300 }
301 
302 static int dropmon_net_event(struct notifier_block *ev_block,
303 			     unsigned long event, void *ptr)
304 {
305 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
306 	struct dm_hw_stat_delta *new_stat = NULL;
307 	struct dm_hw_stat_delta *tmp;
308 
309 	switch (event) {
310 	case NETDEV_REGISTER:
311 		new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
312 
313 		if (!new_stat)
314 			goto out;
315 
316 		new_stat->dev = dev;
317 		new_stat->last_rx = jiffies;
318 		mutex_lock(&trace_state_mutex);
319 		list_add_rcu(&new_stat->list, &hw_stats_list);
320 		mutex_unlock(&trace_state_mutex);
321 		break;
322 	case NETDEV_UNREGISTER:
323 		mutex_lock(&trace_state_mutex);
324 		list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
325 			if (new_stat->dev == dev) {
326 				new_stat->dev = NULL;
327 				if (trace_state == TRACE_OFF) {
328 					list_del_rcu(&new_stat->list);
329 					kfree_rcu(new_stat, rcu);
330 					break;
331 				}
332 			}
333 		}
334 		mutex_unlock(&trace_state_mutex);
335 		break;
336 	}
337 out:
338 	return NOTIFY_DONE;
339 }
340 
341 static const struct genl_ops dropmon_ops[] = {
342 	{
343 		.cmd = NET_DM_CMD_CONFIG,
344 		.doit = net_dm_cmd_config,
345 	},
346 	{
347 		.cmd = NET_DM_CMD_START,
348 		.doit = net_dm_cmd_trace,
349 	},
350 	{
351 		.cmd = NET_DM_CMD_STOP,
352 		.doit = net_dm_cmd_trace,
353 	},
354 };
355 
356 static struct notifier_block dropmon_net_notifier = {
357 	.notifier_call = dropmon_net_event
358 };
359 
360 static int __init init_net_drop_monitor(void)
361 {
362 	struct per_cpu_dm_data *data;
363 	int cpu, rc;
364 
365 	pr_info("Initializing network drop monitor service\n");
366 
367 	if (sizeof(void *) > 8) {
368 		pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
369 		return -ENOSPC;
370 	}
371 
372 	rc = genl_register_family_with_ops_groups(&net_drop_monitor_family,
373 						  dropmon_ops, dropmon_mcgrps);
374 	if (rc) {
375 		pr_err("Could not create drop monitor netlink family\n");
376 		return rc;
377 	}
378 	WARN_ON(net_drop_monitor_family.mcgrp_offset != NET_DM_GRP_ALERT);
379 
380 	rc = register_netdevice_notifier(&dropmon_net_notifier);
381 	if (rc < 0) {
382 		pr_crit("Failed to register netdevice notifier\n");
383 		goto out_unreg;
384 	}
385 
386 	rc = 0;
387 
388 	for_each_possible_cpu(cpu) {
389 		data = &per_cpu(dm_cpu_data, cpu);
390 		INIT_WORK(&data->dm_alert_work, send_dm_alert);
391 		init_timer(&data->send_timer);
392 		data->send_timer.data = (unsigned long)data;
393 		data->send_timer.function = sched_send_work;
394 		spin_lock_init(&data->lock);
395 		reset_per_cpu_data(data);
396 	}
397 
398 
399 	goto out;
400 
401 out_unreg:
402 	genl_unregister_family(&net_drop_monitor_family);
403 out:
404 	return rc;
405 }
406 
407 static void exit_net_drop_monitor(void)
408 {
409 	struct per_cpu_dm_data *data;
410 	int cpu;
411 
412 	BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
413 
414 	/*
415 	 * Because of the module_get/put we do in the trace state change path
416 	 * we are guarnateed not to have any current users when we get here
417 	 * all we need to do is make sure that we don't have any running timers
418 	 * or pending schedule calls
419 	 */
420 
421 	for_each_possible_cpu(cpu) {
422 		data = &per_cpu(dm_cpu_data, cpu);
423 		del_timer_sync(&data->send_timer);
424 		cancel_work_sync(&data->dm_alert_work);
425 		/*
426 		 * At this point, we should have exclusive access
427 		 * to this struct and can free the skb inside it
428 		 */
429 		kfree_skb(data->skb);
430 	}
431 
432 	BUG_ON(genl_unregister_family(&net_drop_monitor_family));
433 }
434 
435 module_init(init_net_drop_monitor);
436 module_exit(exit_net_drop_monitor);
437 
438 MODULE_LICENSE("GPL v2");
439 MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
440 MODULE_ALIAS_GENL_FAMILY("NET_DM");
441