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 */
44static int trace_state = TRACE_OFF;
45static DEFINE_MUTEX(trace_state_mutex);
46
47struct 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
54struct 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
62static struct genl_family net_drop_monitor_family;
63
64static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
65
66static int dm_hit_limit = 64;
67static int dm_delay = 1;
68static unsigned long dm_hw_check_delta = 2*HZ;
69static LIST_HEAD(hw_stats_list);
70
71static struct sk_buff *reset_per_cpu_data(struct per_cpu_dm_data *data)
72{
73 size_t al;
74 struct net_dm_alert_msg *msg;
75 struct nlattr *nla;
76 struct sk_buff *skb;
77 unsigned long flags;
78 void *msg_header;
79
80 al = sizeof(struct net_dm_alert_msg);
81 al += dm_hit_limit * sizeof(struct net_dm_drop_point);
82 al += sizeof(struct nlattr);
83
84 skb = genlmsg_new(al, GFP_KERNEL);
85
86 if (!skb)
87 goto err;
88
89 msg_header = genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
90 0, NET_DM_CMD_ALERT);
91 if (!msg_header) {
92 nlmsg_free(skb);
93 skb = NULL;
94 goto err;
95 }
96 nla = nla_reserve(skb, NLA_UNSPEC,
97 sizeof(struct net_dm_alert_msg));
98 if (!nla) {
99 nlmsg_free(skb);
100 skb = NULL;
101 goto err;
102 }
103 msg = nla_data(nla);
104 memset(msg, 0, al);
105 goto out;
106
107err:
108 mod_timer(&data->send_timer, jiffies + HZ / 10);
109out:
110 spin_lock_irqsave(&data->lock, flags);
111 swap(data->skb, skb);
112 spin_unlock_irqrestore(&data->lock, flags);
113
114 if (skb) {
115 struct nlmsghdr *nlh = (struct nlmsghdr *)skb->data;
116 struct genlmsghdr *gnlh = (struct genlmsghdr *)nlmsg_data(nlh);
117
118 genlmsg_end(skb, genlmsg_data(gnlh));
119 }
120
121 return skb;
122}
123
124static const struct genl_multicast_group dropmon_mcgrps[] = {
125 { .name = "events", },
126};
127
128static void send_dm_alert(struct work_struct *work)
129{
130 struct sk_buff *skb;
131 struct per_cpu_dm_data *data;
132
133 data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
134
135 skb = reset_per_cpu_data(data);
136
137 if (skb)
138 genlmsg_multicast(&net_drop_monitor_family, skb, 0,
139 0, GFP_KERNEL);
140}
141
142/*
143 * This is the timer function to delay the sending of an alert
144 * in the event that more drops will arrive during the
145 * hysteresis period.
146 */
147static void sched_send_work(struct timer_list *t)
148{
149 struct per_cpu_dm_data *data = from_timer(data, t, send_timer);
150
151 schedule_work(&data->dm_alert_work);
152}
153
154static void trace_drop_common(struct sk_buff *skb, void *location)
155{
156 struct net_dm_alert_msg *msg;
157 struct nlmsghdr *nlh;
158 struct nlattr *nla;
159 int i;
160 struct sk_buff *dskb;
161 struct per_cpu_dm_data *data;
162 unsigned long flags;
163
164 local_irq_save(flags);
165 data = this_cpu_ptr(&dm_cpu_data);
166 spin_lock(&data->lock);
167 dskb = data->skb;
168
169 if (!dskb)
170 goto out;
171
172 nlh = (struct nlmsghdr *)dskb->data;
173 nla = genlmsg_data(nlmsg_data(nlh));
174 msg = nla_data(nla);
175 for (i = 0; i < msg->entries; i++) {
176 if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
177 msg->points[i].count++;
178 goto out;
179 }
180 }
181 if (msg->entries == dm_hit_limit)
182 goto out;
183 /*
184 * We need to create a new entry
185 */
186 __nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
187 nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
188 memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
189 msg->points[msg->entries].count = 1;
190 msg->entries++;
191
192 if (!timer_pending(&data->send_timer)) {
193 data->send_timer.expires = jiffies + dm_delay * HZ;
194 add_timer(&data->send_timer);
195 }
196
197out:
198 spin_unlock_irqrestore(&data->lock, flags);
199}
200
201static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
202{
203 trace_drop_common(skb, location);
204}
205
206static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi,
207 int work, int budget)
208{
209 struct dm_hw_stat_delta *new_stat;
210
211 /*
212 * Don't check napi structures with no associated device
213 */
214 if (!napi->dev)
215 return;
216
217 rcu_read_lock();
218 list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
219 /*
220 * only add a note to our monitor buffer if:
221 * 1) this is the dev we received on
222 * 2) its after the last_rx delta
223 * 3) our rx_dropped count has gone up
224 */
225 if ((new_stat->dev == napi->dev) &&
226 (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
227 (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
228 trace_drop_common(NULL, NULL);
229 new_stat->last_drop_val = napi->dev->stats.rx_dropped;
230 new_stat->last_rx = jiffies;
231 break;
232 }
233 }
234 rcu_read_unlock();
235}
236
237static int set_all_monitor_traces(int state)
238{
239 int rc = 0;
240 struct dm_hw_stat_delta *new_stat = NULL;
241 struct dm_hw_stat_delta *temp;
242
243 mutex_lock(&trace_state_mutex);
244
245 if (state == trace_state) {
246 rc = -EAGAIN;
247 goto out_unlock;
248 }
249
250 switch (state) {
251 case TRACE_ON:
252 if (!try_module_get(THIS_MODULE)) {
253 rc = -ENODEV;
254 break;
255 }
256
257 rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
258 rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
259 break;
260
261 case TRACE_OFF:
262 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
263 rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
264
265 tracepoint_synchronize_unregister();
266
267 /*
268 * Clean the device list
269 */
270 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
271 if (new_stat->dev == NULL) {
272 list_del_rcu(&new_stat->list);
273 kfree_rcu(new_stat, rcu);
274 }
275 }
276
277 module_put(THIS_MODULE);
278
279 break;
280 default:
281 rc = 1;
282 break;
283 }
284
285 if (!rc)
286 trace_state = state;
287 else
288 rc = -EINPROGRESS;
289
290out_unlock:
291 mutex_unlock(&trace_state_mutex);
292
293 return rc;
294}
295
296
297static int net_dm_cmd_config(struct sk_buff *skb,
298 struct genl_info *info)
299{
300 return -ENOTSUPP;
301}
302
303static int net_dm_cmd_trace(struct sk_buff *skb,
304 struct genl_info *info)
305{
306 switch (info->genlhdr->cmd) {
307 case NET_DM_CMD_START:
308 return set_all_monitor_traces(TRACE_ON);
309 case NET_DM_CMD_STOP:
310 return set_all_monitor_traces(TRACE_OFF);
311 }
312
313 return -ENOTSUPP;
314}
315
316static int dropmon_net_event(struct notifier_block *ev_block,
317 unsigned long event, void *ptr)
318{
319 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
320 struct dm_hw_stat_delta *new_stat = NULL;
321 struct dm_hw_stat_delta *tmp;
322
323 switch (event) {
324 case NETDEV_REGISTER:
325 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
326
327 if (!new_stat)
328 goto out;
329
330 new_stat->dev = dev;
331 new_stat->last_rx = jiffies;
332 mutex_lock(&trace_state_mutex);
333 list_add_rcu(&new_stat->list, &hw_stats_list);
334 mutex_unlock(&trace_state_mutex);
335 break;
336 case NETDEV_UNREGISTER:
337 mutex_lock(&trace_state_mutex);
338 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
339 if (new_stat->dev == dev) {
340 new_stat->dev = NULL;
341 if (trace_state == TRACE_OFF) {
342 list_del_rcu(&new_stat->list);
343 kfree_rcu(new_stat, rcu);
344 break;
345 }
346 }
347 }
348 mutex_unlock(&trace_state_mutex);
349 break;
350 }
351out:
352 return NOTIFY_DONE;
353}
354
355static const struct genl_ops dropmon_ops[] = {
356 {
357 .cmd = NET_DM_CMD_CONFIG,
358 .doit = net_dm_cmd_config,
359 },
360 {
361 .cmd = NET_DM_CMD_START,
362 .doit = net_dm_cmd_trace,
363 },
364 {
365 .cmd = NET_DM_CMD_STOP,
366 .doit = net_dm_cmd_trace,
367 },
368};
369
370static struct genl_family net_drop_monitor_family __ro_after_init = {
371 .hdrsize = 0,
372 .name = "NET_DM",
373 .version = 2,
374 .module = THIS_MODULE,
375 .ops = dropmon_ops,
376 .n_ops = ARRAY_SIZE(dropmon_ops),
377 .mcgrps = dropmon_mcgrps,
378 .n_mcgrps = ARRAY_SIZE(dropmon_mcgrps),
379};
380
381static struct notifier_block dropmon_net_notifier = {
382 .notifier_call = dropmon_net_event
383};
384
385static int __init init_net_drop_monitor(void)
386{
387 struct per_cpu_dm_data *data;
388 int cpu, rc;
389
390 pr_info("Initializing network drop monitor service\n");
391
392 if (sizeof(void *) > 8) {
393 pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
394 return -ENOSPC;
395 }
396
397 rc = genl_register_family(&net_drop_monitor_family);
398 if (rc) {
399 pr_err("Could not create drop monitor netlink family\n");
400 return rc;
401 }
402 WARN_ON(net_drop_monitor_family.mcgrp_offset != NET_DM_GRP_ALERT);
403
404 rc = register_netdevice_notifier(&dropmon_net_notifier);
405 if (rc < 0) {
406 pr_crit("Failed to register netdevice notifier\n");
407 goto out_unreg;
408 }
409
410 rc = 0;
411
412 for_each_possible_cpu(cpu) {
413 data = &per_cpu(dm_cpu_data, cpu);
414 INIT_WORK(&data->dm_alert_work, send_dm_alert);
415 timer_setup(&data->send_timer, sched_send_work, 0);
416 spin_lock_init(&data->lock);
417 reset_per_cpu_data(data);
418 }
419
420
421 goto out;
422
423out_unreg:
424 genl_unregister_family(&net_drop_monitor_family);
425out:
426 return rc;
427}
428
429static void exit_net_drop_monitor(void)
430{
431 struct per_cpu_dm_data *data;
432 int cpu;
433
434 BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
435
436 /*
437 * Because of the module_get/put we do in the trace state change path
438 * we are guarnateed not to have any current users when we get here
439 * all we need to do is make sure that we don't have any running timers
440 * or pending schedule calls
441 */
442
443 for_each_possible_cpu(cpu) {
444 data = &per_cpu(dm_cpu_data, cpu);
445 del_timer_sync(&data->send_timer);
446 cancel_work_sync(&data->dm_alert_work);
447 /*
448 * At this point, we should have exclusive access
449 * to this struct and can free the skb inside it
450 */
451 kfree_skb(data->skb);
452 }
453
454 BUG_ON(genl_unregister_family(&net_drop_monitor_family));
455}
456
457module_init(init_net_drop_monitor);
458module_exit(exit_net_drop_monitor);
459
460MODULE_LICENSE("GPL v2");
461MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
462MODULE_ALIAS_GENL_FAMILY("NET_DM");
463