xref: /openbmc/linux/net/core/drop_monitor.c (revision b9ccfda2)
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 void send_dm_alert(struct work_struct *work)
110 {
111 	struct sk_buff *skb;
112 	struct per_cpu_dm_data *data;
113 
114 	data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
115 
116 	skb = reset_per_cpu_data(data);
117 
118 	if (skb)
119 		genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL);
120 }
121 
122 /*
123  * This is the timer function to delay the sending of an alert
124  * in the event that more drops will arrive during the
125  * hysteresis period.
126  */
127 static void sched_send_work(unsigned long _data)
128 {
129 	struct per_cpu_dm_data *data = (struct per_cpu_dm_data *)_data;
130 
131 	schedule_work(&data->dm_alert_work);
132 }
133 
134 static void trace_drop_common(struct sk_buff *skb, void *location)
135 {
136 	struct net_dm_alert_msg *msg;
137 	struct nlmsghdr *nlh;
138 	struct nlattr *nla;
139 	int i;
140 	struct sk_buff *dskb;
141 	struct per_cpu_dm_data *data;
142 	unsigned long flags;
143 
144 	local_irq_save(flags);
145 	data = &__get_cpu_var(dm_cpu_data);
146 	spin_lock(&data->lock);
147 	dskb = data->skb;
148 
149 	if (!dskb)
150 		goto out;
151 
152 	nlh = (struct nlmsghdr *)dskb->data;
153 	nla = genlmsg_data(nlmsg_data(nlh));
154 	msg = nla_data(nla);
155 	for (i = 0; i < msg->entries; i++) {
156 		if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
157 			msg->points[i].count++;
158 			goto out;
159 		}
160 	}
161 	if (msg->entries == dm_hit_limit)
162 		goto out;
163 	/*
164 	 * We need to create a new entry
165 	 */
166 	__nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
167 	nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
168 	memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
169 	msg->points[msg->entries].count = 1;
170 	msg->entries++;
171 
172 	if (!timer_pending(&data->send_timer)) {
173 		data->send_timer.expires = jiffies + dm_delay * HZ;
174 		add_timer(&data->send_timer);
175 	}
176 
177 out:
178 	spin_unlock_irqrestore(&data->lock, flags);
179 }
180 
181 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
182 {
183 	trace_drop_common(skb, location);
184 }
185 
186 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi)
187 {
188 	struct dm_hw_stat_delta *new_stat;
189 
190 	/*
191 	 * Don't check napi structures with no associated device
192 	 */
193 	if (!napi->dev)
194 		return;
195 
196 	rcu_read_lock();
197 	list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
198 		/*
199 		 * only add a note to our monitor buffer if:
200 		 * 1) this is the dev we received on
201 		 * 2) its after the last_rx delta
202 		 * 3) our rx_dropped count has gone up
203 		 */
204 		if ((new_stat->dev == napi->dev)  &&
205 		    (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
206 		    (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
207 			trace_drop_common(NULL, NULL);
208 			new_stat->last_drop_val = napi->dev->stats.rx_dropped;
209 			new_stat->last_rx = jiffies;
210 			break;
211 		}
212 	}
213 	rcu_read_unlock();
214 }
215 
216 static int set_all_monitor_traces(int state)
217 {
218 	int rc = 0;
219 	struct dm_hw_stat_delta *new_stat = NULL;
220 	struct dm_hw_stat_delta *temp;
221 
222 	mutex_lock(&trace_state_mutex);
223 
224 	if (state == trace_state) {
225 		rc = -EAGAIN;
226 		goto out_unlock;
227 	}
228 
229 	switch (state) {
230 	case TRACE_ON:
231 		if (!try_module_get(THIS_MODULE)) {
232 			rc = -ENODEV;
233 			break;
234 		}
235 
236 		rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
237 		rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
238 		break;
239 
240 	case TRACE_OFF:
241 		rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
242 		rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
243 
244 		tracepoint_synchronize_unregister();
245 
246 		/*
247 		 * Clean the device list
248 		 */
249 		list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
250 			if (new_stat->dev == NULL) {
251 				list_del_rcu(&new_stat->list);
252 				kfree_rcu(new_stat, rcu);
253 			}
254 		}
255 
256 		module_put(THIS_MODULE);
257 
258 		break;
259 	default:
260 		rc = 1;
261 		break;
262 	}
263 
264 	if (!rc)
265 		trace_state = state;
266 	else
267 		rc = -EINPROGRESS;
268 
269 out_unlock:
270 	mutex_unlock(&trace_state_mutex);
271 
272 	return rc;
273 }
274 
275 
276 static int net_dm_cmd_config(struct sk_buff *skb,
277 			struct genl_info *info)
278 {
279 	return -ENOTSUPP;
280 }
281 
282 static int net_dm_cmd_trace(struct sk_buff *skb,
283 			struct genl_info *info)
284 {
285 	switch (info->genlhdr->cmd) {
286 	case NET_DM_CMD_START:
287 		return set_all_monitor_traces(TRACE_ON);
288 		break;
289 	case NET_DM_CMD_STOP:
290 		return set_all_monitor_traces(TRACE_OFF);
291 		break;
292 	}
293 
294 	return -ENOTSUPP;
295 }
296 
297 static int dropmon_net_event(struct notifier_block *ev_block,
298 			unsigned long event, void *ptr)
299 {
300 	struct net_device *dev = ptr;
301 	struct dm_hw_stat_delta *new_stat = NULL;
302 	struct dm_hw_stat_delta *tmp;
303 
304 	switch (event) {
305 	case NETDEV_REGISTER:
306 		new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
307 
308 		if (!new_stat)
309 			goto out;
310 
311 		new_stat->dev = dev;
312 		new_stat->last_rx = jiffies;
313 		mutex_lock(&trace_state_mutex);
314 		list_add_rcu(&new_stat->list, &hw_stats_list);
315 		mutex_unlock(&trace_state_mutex);
316 		break;
317 	case NETDEV_UNREGISTER:
318 		mutex_lock(&trace_state_mutex);
319 		list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
320 			if (new_stat->dev == dev) {
321 				new_stat->dev = NULL;
322 				if (trace_state == TRACE_OFF) {
323 					list_del_rcu(&new_stat->list);
324 					kfree_rcu(new_stat, rcu);
325 					break;
326 				}
327 			}
328 		}
329 		mutex_unlock(&trace_state_mutex);
330 		break;
331 	}
332 out:
333 	return NOTIFY_DONE;
334 }
335 
336 static struct genl_ops dropmon_ops[] = {
337 	{
338 		.cmd = NET_DM_CMD_CONFIG,
339 		.doit = net_dm_cmd_config,
340 	},
341 	{
342 		.cmd = NET_DM_CMD_START,
343 		.doit = net_dm_cmd_trace,
344 	},
345 	{
346 		.cmd = NET_DM_CMD_STOP,
347 		.doit = net_dm_cmd_trace,
348 	},
349 };
350 
351 static struct notifier_block dropmon_net_notifier = {
352 	.notifier_call = dropmon_net_event
353 };
354 
355 static int __init init_net_drop_monitor(void)
356 {
357 	struct per_cpu_dm_data *data;
358 	int cpu, rc;
359 
360 	pr_info("Initializing network drop monitor service\n");
361 
362 	if (sizeof(void *) > 8) {
363 		pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
364 		return -ENOSPC;
365 	}
366 
367 	rc = genl_register_family_with_ops(&net_drop_monitor_family,
368 					   dropmon_ops,
369 					   ARRAY_SIZE(dropmon_ops));
370 	if (rc) {
371 		pr_err("Could not create drop monitor netlink family\n");
372 		return rc;
373 	}
374 
375 	rc = register_netdevice_notifier(&dropmon_net_notifier);
376 	if (rc < 0) {
377 		pr_crit("Failed to register netdevice notifier\n");
378 		goto out_unreg;
379 	}
380 
381 	rc = 0;
382 
383 	for_each_possible_cpu(cpu) {
384 		data = &per_cpu(dm_cpu_data, cpu);
385 		INIT_WORK(&data->dm_alert_work, send_dm_alert);
386 		init_timer(&data->send_timer);
387 		data->send_timer.data = (unsigned long)data;
388 		data->send_timer.function = sched_send_work;
389 		spin_lock_init(&data->lock);
390 		reset_per_cpu_data(data);
391 	}
392 
393 
394 	goto out;
395 
396 out_unreg:
397 	genl_unregister_family(&net_drop_monitor_family);
398 out:
399 	return rc;
400 }
401 
402 static void exit_net_drop_monitor(void)
403 {
404 	struct per_cpu_dm_data *data;
405 	int cpu;
406 
407 	BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
408 
409 	/*
410 	 * Because of the module_get/put we do in the trace state change path
411 	 * we are guarnateed not to have any current users when we get here
412 	 * all we need to do is make sure that we don't have any running timers
413 	 * or pending schedule calls
414 	 */
415 
416 	for_each_possible_cpu(cpu) {
417 		data = &per_cpu(dm_cpu_data, cpu);
418 		del_timer_sync(&data->send_timer);
419 		cancel_work_sync(&data->dm_alert_work);
420 		/*
421 		 * At this point, we should have exclusive access
422 		 * to this struct and can free the skb inside it
423 		 */
424 		kfree_skb(data->skb);
425 	}
426 
427 	BUG_ON(genl_unregister_family(&net_drop_monitor_family));
428 }
429 
430 module_init(init_net_drop_monitor);
431 module_exit(exit_net_drop_monitor);
432 
433 MODULE_LICENSE("GPL v2");
434 MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
435 MODULE_ALIAS_GENL_FAMILY("NET_DM");
436