xref: /openbmc/linux/net/core/drop_monitor.c (revision 78560d41)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Monitoring code for network dropped packet alerts
4  *
5  * Copyright (C) 2009 Neil Horman <nhorman@tuxdriver.com>
6  */
7 
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 
10 #include <linux/netdevice.h>
11 #include <linux/etherdevice.h>
12 #include <linux/string.h>
13 #include <linux/if_arp.h>
14 #include <linux/inetdevice.h>
15 #include <linux/inet.h>
16 #include <linux/interrupt.h>
17 #include <linux/netpoll.h>
18 #include <linux/sched.h>
19 #include <linux/delay.h>
20 #include <linux/types.h>
21 #include <linux/workqueue.h>
22 #include <linux/netlink.h>
23 #include <linux/net_dropmon.h>
24 #include <linux/percpu.h>
25 #include <linux/timer.h>
26 #include <linux/bitops.h>
27 #include <linux/slab.h>
28 #include <linux/module.h>
29 #include <net/drop_monitor.h>
30 #include <net/genetlink.h>
31 #include <net/netevent.h>
32 
33 #include <trace/events/skb.h>
34 #include <trace/events/napi.h>
35 
36 #include <asm/unaligned.h>
37 
38 #define TRACE_ON 1
39 #define TRACE_OFF 0
40 
41 /*
42  * Globals, our netlink socket pointer
43  * and the work handle that will send up
44  * netlink alerts
45  */
46 static int trace_state = TRACE_OFF;
47 static bool monitor_hw;
48 
49 /* net_dm_mutex
50  *
51  * An overall lock guarding every operation coming from userspace.
52  * It also guards the global 'hw_stats_list' list.
53  */
54 static DEFINE_MUTEX(net_dm_mutex);
55 
56 struct net_dm_stats {
57 	u64 dropped;
58 	struct u64_stats_sync syncp;
59 };
60 
61 #define NET_DM_MAX_HW_TRAP_NAME_LEN 40
62 
63 struct net_dm_hw_entry {
64 	char trap_name[NET_DM_MAX_HW_TRAP_NAME_LEN];
65 	u32 count;
66 };
67 
68 struct net_dm_hw_entries {
69 	u32 num_entries;
70 	struct net_dm_hw_entry entries[0];
71 };
72 
73 struct per_cpu_dm_data {
74 	spinlock_t		lock;	/* Protects 'skb', 'hw_entries' and
75 					 * 'send_timer'
76 					 */
77 	union {
78 		struct sk_buff			*skb;
79 		struct net_dm_hw_entries	*hw_entries;
80 	};
81 	struct sk_buff_head	drop_queue;
82 	struct work_struct	dm_alert_work;
83 	struct timer_list	send_timer;
84 	struct net_dm_stats	stats;
85 };
86 
87 struct dm_hw_stat_delta {
88 	struct net_device *dev;
89 	unsigned long last_rx;
90 	struct list_head list;
91 	struct rcu_head rcu;
92 	unsigned long last_drop_val;
93 };
94 
95 static struct genl_family net_drop_monitor_family;
96 
97 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
98 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_hw_cpu_data);
99 
100 static int dm_hit_limit = 64;
101 static int dm_delay = 1;
102 static unsigned long dm_hw_check_delta = 2*HZ;
103 static LIST_HEAD(hw_stats_list);
104 
105 static enum net_dm_alert_mode net_dm_alert_mode = NET_DM_ALERT_MODE_SUMMARY;
106 static u32 net_dm_trunc_len;
107 static u32 net_dm_queue_len = 1000;
108 
109 struct net_dm_alert_ops {
110 	void (*kfree_skb_probe)(void *ignore, struct sk_buff *skb,
111 				void *location);
112 	void (*napi_poll_probe)(void *ignore, struct napi_struct *napi,
113 				int work, int budget);
114 	void (*work_item_func)(struct work_struct *work);
115 	void (*hw_work_item_func)(struct work_struct *work);
116 	void (*hw_probe)(struct sk_buff *skb,
117 			 const struct net_dm_hw_metadata *hw_metadata);
118 };
119 
120 struct net_dm_skb_cb {
121 	union {
122 		struct net_dm_hw_metadata *hw_metadata;
123 		void *pc;
124 	};
125 };
126 
127 #define NET_DM_SKB_CB(__skb) ((struct net_dm_skb_cb *)&((__skb)->cb[0]))
128 
129 static struct sk_buff *reset_per_cpu_data(struct per_cpu_dm_data *data)
130 {
131 	size_t al;
132 	struct net_dm_alert_msg *msg;
133 	struct nlattr *nla;
134 	struct sk_buff *skb;
135 	unsigned long flags;
136 	void *msg_header;
137 
138 	al = sizeof(struct net_dm_alert_msg);
139 	al += dm_hit_limit * sizeof(struct net_dm_drop_point);
140 	al += sizeof(struct nlattr);
141 
142 	skb = genlmsg_new(al, GFP_KERNEL);
143 
144 	if (!skb)
145 		goto err;
146 
147 	msg_header = genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
148 				 0, NET_DM_CMD_ALERT);
149 	if (!msg_header) {
150 		nlmsg_free(skb);
151 		skb = NULL;
152 		goto err;
153 	}
154 	nla = nla_reserve(skb, NLA_UNSPEC,
155 			  sizeof(struct net_dm_alert_msg));
156 	if (!nla) {
157 		nlmsg_free(skb);
158 		skb = NULL;
159 		goto err;
160 	}
161 	msg = nla_data(nla);
162 	memset(msg, 0, al);
163 	goto out;
164 
165 err:
166 	mod_timer(&data->send_timer, jiffies + HZ / 10);
167 out:
168 	spin_lock_irqsave(&data->lock, flags);
169 	swap(data->skb, skb);
170 	spin_unlock_irqrestore(&data->lock, flags);
171 
172 	if (skb) {
173 		struct nlmsghdr *nlh = (struct nlmsghdr *)skb->data;
174 		struct genlmsghdr *gnlh = (struct genlmsghdr *)nlmsg_data(nlh);
175 
176 		genlmsg_end(skb, genlmsg_data(gnlh));
177 	}
178 
179 	return skb;
180 }
181 
182 static const struct genl_multicast_group dropmon_mcgrps[] = {
183 	{ .name = "events", },
184 };
185 
186 static void send_dm_alert(struct work_struct *work)
187 {
188 	struct sk_buff *skb;
189 	struct per_cpu_dm_data *data;
190 
191 	data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
192 
193 	skb = reset_per_cpu_data(data);
194 
195 	if (skb)
196 		genlmsg_multicast(&net_drop_monitor_family, skb, 0,
197 				  0, GFP_KERNEL);
198 }
199 
200 /*
201  * This is the timer function to delay the sending of an alert
202  * in the event that more drops will arrive during the
203  * hysteresis period.
204  */
205 static void sched_send_work(struct timer_list *t)
206 {
207 	struct per_cpu_dm_data *data = from_timer(data, t, send_timer);
208 
209 	schedule_work(&data->dm_alert_work);
210 }
211 
212 static void trace_drop_common(struct sk_buff *skb, void *location)
213 {
214 	struct net_dm_alert_msg *msg;
215 	struct nlmsghdr *nlh;
216 	struct nlattr *nla;
217 	int i;
218 	struct sk_buff *dskb;
219 	struct per_cpu_dm_data *data;
220 	unsigned long flags;
221 
222 	local_irq_save(flags);
223 	data = this_cpu_ptr(&dm_cpu_data);
224 	spin_lock(&data->lock);
225 	dskb = data->skb;
226 
227 	if (!dskb)
228 		goto out;
229 
230 	nlh = (struct nlmsghdr *)dskb->data;
231 	nla = genlmsg_data(nlmsg_data(nlh));
232 	msg = nla_data(nla);
233 	for (i = 0; i < msg->entries; i++) {
234 		if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
235 			msg->points[i].count++;
236 			goto out;
237 		}
238 	}
239 	if (msg->entries == dm_hit_limit)
240 		goto out;
241 	/*
242 	 * We need to create a new entry
243 	 */
244 	__nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
245 	nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
246 	memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
247 	msg->points[msg->entries].count = 1;
248 	msg->entries++;
249 
250 	if (!timer_pending(&data->send_timer)) {
251 		data->send_timer.expires = jiffies + dm_delay * HZ;
252 		add_timer(&data->send_timer);
253 	}
254 
255 out:
256 	spin_unlock_irqrestore(&data->lock, flags);
257 }
258 
259 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
260 {
261 	trace_drop_common(skb, location);
262 }
263 
264 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi,
265 				int work, int budget)
266 {
267 	struct dm_hw_stat_delta *new_stat;
268 
269 	/*
270 	 * Don't check napi structures with no associated device
271 	 */
272 	if (!napi->dev)
273 		return;
274 
275 	rcu_read_lock();
276 	list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
277 		/*
278 		 * only add a note to our monitor buffer if:
279 		 * 1) this is the dev we received on
280 		 * 2) its after the last_rx delta
281 		 * 3) our rx_dropped count has gone up
282 		 */
283 		if ((new_stat->dev == napi->dev)  &&
284 		    (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
285 		    (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
286 			trace_drop_common(NULL, NULL);
287 			new_stat->last_drop_val = napi->dev->stats.rx_dropped;
288 			new_stat->last_rx = jiffies;
289 			break;
290 		}
291 	}
292 	rcu_read_unlock();
293 }
294 
295 static struct net_dm_hw_entries *
296 net_dm_hw_reset_per_cpu_data(struct per_cpu_dm_data *hw_data)
297 {
298 	struct net_dm_hw_entries *hw_entries;
299 	unsigned long flags;
300 
301 	hw_entries = kzalloc(struct_size(hw_entries, entries, dm_hit_limit),
302 			     GFP_KERNEL);
303 	if (!hw_entries) {
304 		/* If the memory allocation failed, we try to perform another
305 		 * allocation in 1/10 second. Otherwise, the probe function
306 		 * will constantly bail out.
307 		 */
308 		mod_timer(&hw_data->send_timer, jiffies + HZ / 10);
309 	}
310 
311 	spin_lock_irqsave(&hw_data->lock, flags);
312 	swap(hw_data->hw_entries, hw_entries);
313 	spin_unlock_irqrestore(&hw_data->lock, flags);
314 
315 	return hw_entries;
316 }
317 
318 static int net_dm_hw_entry_put(struct sk_buff *msg,
319 			       const struct net_dm_hw_entry *hw_entry)
320 {
321 	struct nlattr *attr;
322 
323 	attr = nla_nest_start(msg, NET_DM_ATTR_HW_ENTRY);
324 	if (!attr)
325 		return -EMSGSIZE;
326 
327 	if (nla_put_string(msg, NET_DM_ATTR_HW_TRAP_NAME, hw_entry->trap_name))
328 		goto nla_put_failure;
329 
330 	if (nla_put_u32(msg, NET_DM_ATTR_HW_TRAP_COUNT, hw_entry->count))
331 		goto nla_put_failure;
332 
333 	nla_nest_end(msg, attr);
334 
335 	return 0;
336 
337 nla_put_failure:
338 	nla_nest_cancel(msg, attr);
339 	return -EMSGSIZE;
340 }
341 
342 static int net_dm_hw_entries_put(struct sk_buff *msg,
343 				 const struct net_dm_hw_entries *hw_entries)
344 {
345 	struct nlattr *attr;
346 	int i;
347 
348 	attr = nla_nest_start(msg, NET_DM_ATTR_HW_ENTRIES);
349 	if (!attr)
350 		return -EMSGSIZE;
351 
352 	for (i = 0; i < hw_entries->num_entries; i++) {
353 		int rc;
354 
355 		rc = net_dm_hw_entry_put(msg, &hw_entries->entries[i]);
356 		if (rc)
357 			goto nla_put_failure;
358 	}
359 
360 	nla_nest_end(msg, attr);
361 
362 	return 0;
363 
364 nla_put_failure:
365 	nla_nest_cancel(msg, attr);
366 	return -EMSGSIZE;
367 }
368 
369 static int
370 net_dm_hw_summary_report_fill(struct sk_buff *msg,
371 			      const struct net_dm_hw_entries *hw_entries)
372 {
373 	struct net_dm_alert_msg anc_hdr = { 0 };
374 	void *hdr;
375 	int rc;
376 
377 	hdr = genlmsg_put(msg, 0, 0, &net_drop_monitor_family, 0,
378 			  NET_DM_CMD_ALERT);
379 	if (!hdr)
380 		return -EMSGSIZE;
381 
382 	/* We need to put the ancillary header in order not to break user
383 	 * space.
384 	 */
385 	if (nla_put(msg, NLA_UNSPEC, sizeof(anc_hdr), &anc_hdr))
386 		goto nla_put_failure;
387 
388 	rc = net_dm_hw_entries_put(msg, hw_entries);
389 	if (rc)
390 		goto nla_put_failure;
391 
392 	genlmsg_end(msg, hdr);
393 
394 	return 0;
395 
396 nla_put_failure:
397 	genlmsg_cancel(msg, hdr);
398 	return -EMSGSIZE;
399 }
400 
401 static void net_dm_hw_summary_work(struct work_struct *work)
402 {
403 	struct net_dm_hw_entries *hw_entries;
404 	struct per_cpu_dm_data *hw_data;
405 	struct sk_buff *msg;
406 	int rc;
407 
408 	hw_data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
409 
410 	hw_entries = net_dm_hw_reset_per_cpu_data(hw_data);
411 	if (!hw_entries)
412 		return;
413 
414 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
415 	if (!msg)
416 		goto out;
417 
418 	rc = net_dm_hw_summary_report_fill(msg, hw_entries);
419 	if (rc) {
420 		nlmsg_free(msg);
421 		goto out;
422 	}
423 
424 	genlmsg_multicast(&net_drop_monitor_family, msg, 0, 0, GFP_KERNEL);
425 
426 out:
427 	kfree(hw_entries);
428 }
429 
430 static void
431 net_dm_hw_summary_probe(struct sk_buff *skb,
432 			const struct net_dm_hw_metadata *hw_metadata)
433 {
434 	struct net_dm_hw_entries *hw_entries;
435 	struct net_dm_hw_entry *hw_entry;
436 	struct per_cpu_dm_data *hw_data;
437 	unsigned long flags;
438 	int i;
439 
440 	hw_data = this_cpu_ptr(&dm_hw_cpu_data);
441 	spin_lock_irqsave(&hw_data->lock, flags);
442 	hw_entries = hw_data->hw_entries;
443 
444 	if (!hw_entries)
445 		goto out;
446 
447 	for (i = 0; i < hw_entries->num_entries; i++) {
448 		hw_entry = &hw_entries->entries[i];
449 		if (!strncmp(hw_entry->trap_name, hw_metadata->trap_name,
450 			     NET_DM_MAX_HW_TRAP_NAME_LEN - 1)) {
451 			hw_entry->count++;
452 			goto out;
453 		}
454 	}
455 	if (WARN_ON_ONCE(hw_entries->num_entries == dm_hit_limit))
456 		goto out;
457 
458 	hw_entry = &hw_entries->entries[hw_entries->num_entries];
459 	strlcpy(hw_entry->trap_name, hw_metadata->trap_name,
460 		NET_DM_MAX_HW_TRAP_NAME_LEN - 1);
461 	hw_entry->count = 1;
462 	hw_entries->num_entries++;
463 
464 	if (!timer_pending(&hw_data->send_timer)) {
465 		hw_data->send_timer.expires = jiffies + dm_delay * HZ;
466 		add_timer(&hw_data->send_timer);
467 	}
468 
469 out:
470 	spin_unlock_irqrestore(&hw_data->lock, flags);
471 }
472 
473 static const struct net_dm_alert_ops net_dm_alert_summary_ops = {
474 	.kfree_skb_probe	= trace_kfree_skb_hit,
475 	.napi_poll_probe	= trace_napi_poll_hit,
476 	.work_item_func		= send_dm_alert,
477 	.hw_work_item_func	= net_dm_hw_summary_work,
478 	.hw_probe		= net_dm_hw_summary_probe,
479 };
480 
481 static void net_dm_packet_trace_kfree_skb_hit(void *ignore,
482 					      struct sk_buff *skb,
483 					      void *location)
484 {
485 	ktime_t tstamp = ktime_get_real();
486 	struct per_cpu_dm_data *data;
487 	struct sk_buff *nskb;
488 	unsigned long flags;
489 
490 	if (!skb_mac_header_was_set(skb))
491 		return;
492 
493 	nskb = skb_clone(skb, GFP_ATOMIC);
494 	if (!nskb)
495 		return;
496 
497 	NET_DM_SKB_CB(nskb)->pc = location;
498 	/* Override the timestamp because we care about the time when the
499 	 * packet was dropped.
500 	 */
501 	nskb->tstamp = tstamp;
502 
503 	data = this_cpu_ptr(&dm_cpu_data);
504 
505 	spin_lock_irqsave(&data->drop_queue.lock, flags);
506 	if (skb_queue_len(&data->drop_queue) < net_dm_queue_len)
507 		__skb_queue_tail(&data->drop_queue, nskb);
508 	else
509 		goto unlock_free;
510 	spin_unlock_irqrestore(&data->drop_queue.lock, flags);
511 
512 	schedule_work(&data->dm_alert_work);
513 
514 	return;
515 
516 unlock_free:
517 	spin_unlock_irqrestore(&data->drop_queue.lock, flags);
518 	u64_stats_update_begin(&data->stats.syncp);
519 	data->stats.dropped++;
520 	u64_stats_update_end(&data->stats.syncp);
521 	consume_skb(nskb);
522 }
523 
524 static void net_dm_packet_trace_napi_poll_hit(void *ignore,
525 					      struct napi_struct *napi,
526 					      int work, int budget)
527 {
528 }
529 
530 static size_t net_dm_in_port_size(void)
531 {
532 	       /* NET_DM_ATTR_IN_PORT nest */
533 	return nla_total_size(0) +
534 	       /* NET_DM_ATTR_PORT_NETDEV_IFINDEX */
535 	       nla_total_size(sizeof(u32)) +
536 	       /* NET_DM_ATTR_PORT_NETDEV_NAME */
537 	       nla_total_size(IFNAMSIZ + 1);
538 }
539 
540 #define NET_DM_MAX_SYMBOL_LEN 40
541 
542 static size_t net_dm_packet_report_size(size_t payload_len)
543 {
544 	size_t size;
545 
546 	size = nlmsg_msg_size(GENL_HDRLEN + net_drop_monitor_family.hdrsize);
547 
548 	return NLMSG_ALIGN(size) +
549 	       /* NET_DM_ATTR_ORIGIN */
550 	       nla_total_size(sizeof(u16)) +
551 	       /* NET_DM_ATTR_PC */
552 	       nla_total_size(sizeof(u64)) +
553 	       /* NET_DM_ATTR_SYMBOL */
554 	       nla_total_size(NET_DM_MAX_SYMBOL_LEN + 1) +
555 	       /* NET_DM_ATTR_IN_PORT */
556 	       net_dm_in_port_size() +
557 	       /* NET_DM_ATTR_TIMESTAMP */
558 	       nla_total_size(sizeof(u64)) +
559 	       /* NET_DM_ATTR_ORIG_LEN */
560 	       nla_total_size(sizeof(u32)) +
561 	       /* NET_DM_ATTR_PROTO */
562 	       nla_total_size(sizeof(u16)) +
563 	       /* NET_DM_ATTR_PAYLOAD */
564 	       nla_total_size(payload_len);
565 }
566 
567 static int net_dm_packet_report_in_port_put(struct sk_buff *msg, int ifindex,
568 					    const char *name)
569 {
570 	struct nlattr *attr;
571 
572 	attr = nla_nest_start(msg, NET_DM_ATTR_IN_PORT);
573 	if (!attr)
574 		return -EMSGSIZE;
575 
576 	if (ifindex &&
577 	    nla_put_u32(msg, NET_DM_ATTR_PORT_NETDEV_IFINDEX, ifindex))
578 		goto nla_put_failure;
579 
580 	if (name && nla_put_string(msg, NET_DM_ATTR_PORT_NETDEV_NAME, name))
581 		goto nla_put_failure;
582 
583 	nla_nest_end(msg, attr);
584 
585 	return 0;
586 
587 nla_put_failure:
588 	nla_nest_cancel(msg, attr);
589 	return -EMSGSIZE;
590 }
591 
592 static int net_dm_packet_report_fill(struct sk_buff *msg, struct sk_buff *skb,
593 				     size_t payload_len)
594 {
595 	u64 pc = (u64)(uintptr_t) NET_DM_SKB_CB(skb)->pc;
596 	char buf[NET_DM_MAX_SYMBOL_LEN];
597 	struct nlattr *attr;
598 	void *hdr;
599 	int rc;
600 
601 	hdr = genlmsg_put(msg, 0, 0, &net_drop_monitor_family, 0,
602 			  NET_DM_CMD_PACKET_ALERT);
603 	if (!hdr)
604 		return -EMSGSIZE;
605 
606 	if (nla_put_u16(msg, NET_DM_ATTR_ORIGIN, NET_DM_ORIGIN_SW))
607 		goto nla_put_failure;
608 
609 	if (nla_put_u64_64bit(msg, NET_DM_ATTR_PC, pc, NET_DM_ATTR_PAD))
610 		goto nla_put_failure;
611 
612 	snprintf(buf, sizeof(buf), "%pS", NET_DM_SKB_CB(skb)->pc);
613 	if (nla_put_string(msg, NET_DM_ATTR_SYMBOL, buf))
614 		goto nla_put_failure;
615 
616 	rc = net_dm_packet_report_in_port_put(msg, skb->skb_iif, NULL);
617 	if (rc)
618 		goto nla_put_failure;
619 
620 	if (nla_put_u64_64bit(msg, NET_DM_ATTR_TIMESTAMP,
621 			      ktime_to_ns(skb->tstamp), NET_DM_ATTR_PAD))
622 		goto nla_put_failure;
623 
624 	if (nla_put_u32(msg, NET_DM_ATTR_ORIG_LEN, skb->len))
625 		goto nla_put_failure;
626 
627 	if (!payload_len)
628 		goto out;
629 
630 	if (nla_put_u16(msg, NET_DM_ATTR_PROTO, be16_to_cpu(skb->protocol)))
631 		goto nla_put_failure;
632 
633 	attr = skb_put(msg, nla_total_size(payload_len));
634 	attr->nla_type = NET_DM_ATTR_PAYLOAD;
635 	attr->nla_len = nla_attr_size(payload_len);
636 	if (skb_copy_bits(skb, 0, nla_data(attr), payload_len))
637 		goto nla_put_failure;
638 
639 out:
640 	genlmsg_end(msg, hdr);
641 
642 	return 0;
643 
644 nla_put_failure:
645 	genlmsg_cancel(msg, hdr);
646 	return -EMSGSIZE;
647 }
648 
649 #define NET_DM_MAX_PACKET_SIZE (0xffff - NLA_HDRLEN - NLA_ALIGNTO)
650 
651 static void net_dm_packet_report(struct sk_buff *skb)
652 {
653 	struct sk_buff *msg;
654 	size_t payload_len;
655 	int rc;
656 
657 	/* Make sure we start copying the packet from the MAC header */
658 	if (skb->data > skb_mac_header(skb))
659 		skb_push(skb, skb->data - skb_mac_header(skb));
660 	else
661 		skb_pull(skb, skb_mac_header(skb) - skb->data);
662 
663 	/* Ensure packet fits inside a single netlink attribute */
664 	payload_len = min_t(size_t, skb->len, NET_DM_MAX_PACKET_SIZE);
665 	if (net_dm_trunc_len)
666 		payload_len = min_t(size_t, net_dm_trunc_len, payload_len);
667 
668 	msg = nlmsg_new(net_dm_packet_report_size(payload_len), GFP_KERNEL);
669 	if (!msg)
670 		goto out;
671 
672 	rc = net_dm_packet_report_fill(msg, skb, payload_len);
673 	if (rc) {
674 		nlmsg_free(msg);
675 		goto out;
676 	}
677 
678 	genlmsg_multicast(&net_drop_monitor_family, msg, 0, 0, GFP_KERNEL);
679 
680 out:
681 	consume_skb(skb);
682 }
683 
684 static void net_dm_packet_work(struct work_struct *work)
685 {
686 	struct per_cpu_dm_data *data;
687 	struct sk_buff_head list;
688 	struct sk_buff *skb;
689 	unsigned long flags;
690 
691 	data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
692 
693 	__skb_queue_head_init(&list);
694 
695 	spin_lock_irqsave(&data->drop_queue.lock, flags);
696 	skb_queue_splice_tail_init(&data->drop_queue, &list);
697 	spin_unlock_irqrestore(&data->drop_queue.lock, flags);
698 
699 	while ((skb = __skb_dequeue(&list)))
700 		net_dm_packet_report(skb);
701 }
702 
703 static size_t
704 net_dm_hw_packet_report_size(size_t payload_len,
705 			     const struct net_dm_hw_metadata *hw_metadata)
706 {
707 	size_t size;
708 
709 	size = nlmsg_msg_size(GENL_HDRLEN + net_drop_monitor_family.hdrsize);
710 
711 	return NLMSG_ALIGN(size) +
712 	       /* NET_DM_ATTR_ORIGIN */
713 	       nla_total_size(sizeof(u16)) +
714 	       /* NET_DM_ATTR_HW_TRAP_GROUP_NAME */
715 	       nla_total_size(strlen(hw_metadata->trap_group_name) + 1) +
716 	       /* NET_DM_ATTR_HW_TRAP_NAME */
717 	       nla_total_size(strlen(hw_metadata->trap_name) + 1) +
718 	       /* NET_DM_ATTR_IN_PORT */
719 	       net_dm_in_port_size() +
720 	       /* NET_DM_ATTR_TIMESTAMP */
721 	       nla_total_size(sizeof(u64)) +
722 	       /* NET_DM_ATTR_ORIG_LEN */
723 	       nla_total_size(sizeof(u32)) +
724 	       /* NET_DM_ATTR_PROTO */
725 	       nla_total_size(sizeof(u16)) +
726 	       /* NET_DM_ATTR_PAYLOAD */
727 	       nla_total_size(payload_len);
728 }
729 
730 static int net_dm_hw_packet_report_fill(struct sk_buff *msg,
731 					struct sk_buff *skb, size_t payload_len)
732 {
733 	struct net_dm_hw_metadata *hw_metadata;
734 	struct nlattr *attr;
735 	void *hdr;
736 
737 	hw_metadata = NET_DM_SKB_CB(skb)->hw_metadata;
738 
739 	hdr = genlmsg_put(msg, 0, 0, &net_drop_monitor_family, 0,
740 			  NET_DM_CMD_PACKET_ALERT);
741 	if (!hdr)
742 		return -EMSGSIZE;
743 
744 	if (nla_put_u16(msg, NET_DM_ATTR_ORIGIN, NET_DM_ORIGIN_HW))
745 		goto nla_put_failure;
746 
747 	if (nla_put_string(msg, NET_DM_ATTR_HW_TRAP_GROUP_NAME,
748 			   hw_metadata->trap_group_name))
749 		goto nla_put_failure;
750 
751 	if (nla_put_string(msg, NET_DM_ATTR_HW_TRAP_NAME,
752 			   hw_metadata->trap_name))
753 		goto nla_put_failure;
754 
755 	if (hw_metadata->input_dev) {
756 		struct net_device *dev = hw_metadata->input_dev;
757 		int rc;
758 
759 		rc = net_dm_packet_report_in_port_put(msg, dev->ifindex,
760 						      dev->name);
761 		if (rc)
762 			goto nla_put_failure;
763 	}
764 
765 	if (nla_put_u64_64bit(msg, NET_DM_ATTR_TIMESTAMP,
766 			      ktime_to_ns(skb->tstamp), NET_DM_ATTR_PAD))
767 		goto nla_put_failure;
768 
769 	if (nla_put_u32(msg, NET_DM_ATTR_ORIG_LEN, skb->len))
770 		goto nla_put_failure;
771 
772 	if (!payload_len)
773 		goto out;
774 
775 	if (nla_put_u16(msg, NET_DM_ATTR_PROTO, be16_to_cpu(skb->protocol)))
776 		goto nla_put_failure;
777 
778 	attr = skb_put(msg, nla_total_size(payload_len));
779 	attr->nla_type = NET_DM_ATTR_PAYLOAD;
780 	attr->nla_len = nla_attr_size(payload_len);
781 	if (skb_copy_bits(skb, 0, nla_data(attr), payload_len))
782 		goto nla_put_failure;
783 
784 out:
785 	genlmsg_end(msg, hdr);
786 
787 	return 0;
788 
789 nla_put_failure:
790 	genlmsg_cancel(msg, hdr);
791 	return -EMSGSIZE;
792 }
793 
794 static struct net_dm_hw_metadata *
795 net_dm_hw_metadata_clone(const struct net_dm_hw_metadata *hw_metadata)
796 {
797 	struct net_dm_hw_metadata *n_hw_metadata;
798 	const char *trap_group_name;
799 	const char *trap_name;
800 
801 	n_hw_metadata = kmalloc(sizeof(*hw_metadata), GFP_ATOMIC);
802 	if (!n_hw_metadata)
803 		return NULL;
804 
805 	trap_group_name = kstrdup(hw_metadata->trap_group_name, GFP_ATOMIC);
806 	if (!trap_group_name)
807 		goto free_hw_metadata;
808 	n_hw_metadata->trap_group_name = trap_group_name;
809 
810 	trap_name = kstrdup(hw_metadata->trap_name, GFP_ATOMIC);
811 	if (!trap_name)
812 		goto free_trap_group;
813 	n_hw_metadata->trap_name = trap_name;
814 
815 	n_hw_metadata->input_dev = hw_metadata->input_dev;
816 	if (n_hw_metadata->input_dev)
817 		dev_hold(n_hw_metadata->input_dev);
818 
819 	return n_hw_metadata;
820 
821 free_trap_group:
822 	kfree(trap_group_name);
823 free_hw_metadata:
824 	kfree(n_hw_metadata);
825 	return NULL;
826 }
827 
828 static void
829 net_dm_hw_metadata_free(const struct net_dm_hw_metadata *hw_metadata)
830 {
831 	if (hw_metadata->input_dev)
832 		dev_put(hw_metadata->input_dev);
833 	kfree(hw_metadata->trap_name);
834 	kfree(hw_metadata->trap_group_name);
835 	kfree(hw_metadata);
836 }
837 
838 static void net_dm_hw_packet_report(struct sk_buff *skb)
839 {
840 	struct net_dm_hw_metadata *hw_metadata;
841 	struct sk_buff *msg;
842 	size_t payload_len;
843 	int rc;
844 
845 	if (skb->data > skb_mac_header(skb))
846 		skb_push(skb, skb->data - skb_mac_header(skb));
847 	else
848 		skb_pull(skb, skb_mac_header(skb) - skb->data);
849 
850 	payload_len = min_t(size_t, skb->len, NET_DM_MAX_PACKET_SIZE);
851 	if (net_dm_trunc_len)
852 		payload_len = min_t(size_t, net_dm_trunc_len, payload_len);
853 
854 	hw_metadata = NET_DM_SKB_CB(skb)->hw_metadata;
855 	msg = nlmsg_new(net_dm_hw_packet_report_size(payload_len, hw_metadata),
856 			GFP_KERNEL);
857 	if (!msg)
858 		goto out;
859 
860 	rc = net_dm_hw_packet_report_fill(msg, skb, payload_len);
861 	if (rc) {
862 		nlmsg_free(msg);
863 		goto out;
864 	}
865 
866 	genlmsg_multicast(&net_drop_monitor_family, msg, 0, 0, GFP_KERNEL);
867 
868 out:
869 	net_dm_hw_metadata_free(NET_DM_SKB_CB(skb)->hw_metadata);
870 	consume_skb(skb);
871 }
872 
873 static void net_dm_hw_packet_work(struct work_struct *work)
874 {
875 	struct per_cpu_dm_data *hw_data;
876 	struct sk_buff_head list;
877 	struct sk_buff *skb;
878 	unsigned long flags;
879 
880 	hw_data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
881 
882 	__skb_queue_head_init(&list);
883 
884 	spin_lock_irqsave(&hw_data->drop_queue.lock, flags);
885 	skb_queue_splice_tail_init(&hw_data->drop_queue, &list);
886 	spin_unlock_irqrestore(&hw_data->drop_queue.lock, flags);
887 
888 	while ((skb = __skb_dequeue(&list)))
889 		net_dm_hw_packet_report(skb);
890 }
891 
892 static void
893 net_dm_hw_packet_probe(struct sk_buff *skb,
894 		       const struct net_dm_hw_metadata *hw_metadata)
895 {
896 	struct net_dm_hw_metadata *n_hw_metadata;
897 	ktime_t tstamp = ktime_get_real();
898 	struct per_cpu_dm_data *hw_data;
899 	struct sk_buff *nskb;
900 	unsigned long flags;
901 
902 	if (!skb_mac_header_was_set(skb))
903 		return;
904 
905 	nskb = skb_clone(skb, GFP_ATOMIC);
906 	if (!nskb)
907 		return;
908 
909 	n_hw_metadata = net_dm_hw_metadata_clone(hw_metadata);
910 	if (!n_hw_metadata)
911 		goto free;
912 
913 	NET_DM_SKB_CB(nskb)->hw_metadata = n_hw_metadata;
914 	nskb->tstamp = tstamp;
915 
916 	hw_data = this_cpu_ptr(&dm_hw_cpu_data);
917 
918 	spin_lock_irqsave(&hw_data->drop_queue.lock, flags);
919 	if (skb_queue_len(&hw_data->drop_queue) < net_dm_queue_len)
920 		__skb_queue_tail(&hw_data->drop_queue, nskb);
921 	else
922 		goto unlock_free;
923 	spin_unlock_irqrestore(&hw_data->drop_queue.lock, flags);
924 
925 	schedule_work(&hw_data->dm_alert_work);
926 
927 	return;
928 
929 unlock_free:
930 	spin_unlock_irqrestore(&hw_data->drop_queue.lock, flags);
931 	u64_stats_update_begin(&hw_data->stats.syncp);
932 	hw_data->stats.dropped++;
933 	u64_stats_update_end(&hw_data->stats.syncp);
934 	net_dm_hw_metadata_free(n_hw_metadata);
935 free:
936 	consume_skb(nskb);
937 }
938 
939 static const struct net_dm_alert_ops net_dm_alert_packet_ops = {
940 	.kfree_skb_probe	= net_dm_packet_trace_kfree_skb_hit,
941 	.napi_poll_probe	= net_dm_packet_trace_napi_poll_hit,
942 	.work_item_func		= net_dm_packet_work,
943 	.hw_work_item_func	= net_dm_hw_packet_work,
944 	.hw_probe		= net_dm_hw_packet_probe,
945 };
946 
947 static const struct net_dm_alert_ops *net_dm_alert_ops_arr[] = {
948 	[NET_DM_ALERT_MODE_SUMMARY]	= &net_dm_alert_summary_ops,
949 	[NET_DM_ALERT_MODE_PACKET]	= &net_dm_alert_packet_ops,
950 };
951 
952 void net_dm_hw_report(struct sk_buff *skb,
953 		      const struct net_dm_hw_metadata *hw_metadata)
954 {
955 	rcu_read_lock();
956 
957 	if (!monitor_hw)
958 		goto out;
959 
960 	net_dm_alert_ops_arr[net_dm_alert_mode]->hw_probe(skb, hw_metadata);
961 
962 out:
963 	rcu_read_unlock();
964 }
965 EXPORT_SYMBOL_GPL(net_dm_hw_report);
966 
967 static int net_dm_hw_monitor_start(struct netlink_ext_ack *extack)
968 {
969 	const struct net_dm_alert_ops *ops;
970 	int cpu;
971 
972 	if (monitor_hw) {
973 		NL_SET_ERR_MSG_MOD(extack, "Hardware monitoring already enabled");
974 		return -EAGAIN;
975 	}
976 
977 	ops = net_dm_alert_ops_arr[net_dm_alert_mode];
978 
979 	if (!try_module_get(THIS_MODULE)) {
980 		NL_SET_ERR_MSG_MOD(extack, "Failed to take reference on module");
981 		return -ENODEV;
982 	}
983 
984 	for_each_possible_cpu(cpu) {
985 		struct per_cpu_dm_data *hw_data = &per_cpu(dm_hw_cpu_data, cpu);
986 		struct net_dm_hw_entries *hw_entries;
987 
988 		INIT_WORK(&hw_data->dm_alert_work, ops->hw_work_item_func);
989 		timer_setup(&hw_data->send_timer, sched_send_work, 0);
990 		hw_entries = net_dm_hw_reset_per_cpu_data(hw_data);
991 		kfree(hw_entries);
992 	}
993 
994 	monitor_hw = true;
995 
996 	return 0;
997 }
998 
999 static void net_dm_hw_monitor_stop(struct netlink_ext_ack *extack)
1000 {
1001 	int cpu;
1002 
1003 	if (!monitor_hw) {
1004 		NL_SET_ERR_MSG_MOD(extack, "Hardware monitoring already disabled");
1005 		return;
1006 	}
1007 
1008 	monitor_hw = false;
1009 
1010 	/* After this call returns we are guaranteed that no CPU is processing
1011 	 * any hardware drops.
1012 	 */
1013 	synchronize_rcu();
1014 
1015 	for_each_possible_cpu(cpu) {
1016 		struct per_cpu_dm_data *hw_data = &per_cpu(dm_hw_cpu_data, cpu);
1017 		struct sk_buff *skb;
1018 
1019 		del_timer_sync(&hw_data->send_timer);
1020 		cancel_work_sync(&hw_data->dm_alert_work);
1021 		while ((skb = __skb_dequeue(&hw_data->drop_queue))) {
1022 			struct net_dm_hw_metadata *hw_metadata;
1023 
1024 			hw_metadata = NET_DM_SKB_CB(skb)->hw_metadata;
1025 			net_dm_hw_metadata_free(hw_metadata);
1026 			consume_skb(skb);
1027 		}
1028 	}
1029 
1030 	module_put(THIS_MODULE);
1031 }
1032 
1033 static int net_dm_trace_on_set(struct netlink_ext_ack *extack)
1034 {
1035 	const struct net_dm_alert_ops *ops;
1036 	int cpu, rc;
1037 
1038 	ops = net_dm_alert_ops_arr[net_dm_alert_mode];
1039 
1040 	if (!try_module_get(THIS_MODULE)) {
1041 		NL_SET_ERR_MSG_MOD(extack, "Failed to take reference on module");
1042 		return -ENODEV;
1043 	}
1044 
1045 	for_each_possible_cpu(cpu) {
1046 		struct per_cpu_dm_data *data = &per_cpu(dm_cpu_data, cpu);
1047 		struct sk_buff *skb;
1048 
1049 		INIT_WORK(&data->dm_alert_work, ops->work_item_func);
1050 		timer_setup(&data->send_timer, sched_send_work, 0);
1051 		/* Allocate a new per-CPU skb for the summary alert message and
1052 		 * free the old one which might contain stale data from
1053 		 * previous tracing.
1054 		 */
1055 		skb = reset_per_cpu_data(data);
1056 		consume_skb(skb);
1057 	}
1058 
1059 	rc = register_trace_kfree_skb(ops->kfree_skb_probe, NULL);
1060 	if (rc) {
1061 		NL_SET_ERR_MSG_MOD(extack, "Failed to connect probe to kfree_skb() tracepoint");
1062 		goto err_module_put;
1063 	}
1064 
1065 	rc = register_trace_napi_poll(ops->napi_poll_probe, NULL);
1066 	if (rc) {
1067 		NL_SET_ERR_MSG_MOD(extack, "Failed to connect probe to napi_poll() tracepoint");
1068 		goto err_unregister_trace;
1069 	}
1070 
1071 	return 0;
1072 
1073 err_unregister_trace:
1074 	unregister_trace_kfree_skb(ops->kfree_skb_probe, NULL);
1075 err_module_put:
1076 	module_put(THIS_MODULE);
1077 	return rc;
1078 }
1079 
1080 static void net_dm_trace_off_set(void)
1081 {
1082 	struct dm_hw_stat_delta *new_stat, *temp;
1083 	const struct net_dm_alert_ops *ops;
1084 	int cpu;
1085 
1086 	ops = net_dm_alert_ops_arr[net_dm_alert_mode];
1087 
1088 	unregister_trace_napi_poll(ops->napi_poll_probe, NULL);
1089 	unregister_trace_kfree_skb(ops->kfree_skb_probe, NULL);
1090 
1091 	tracepoint_synchronize_unregister();
1092 
1093 	/* Make sure we do not send notifications to user space after request
1094 	 * to stop tracing returns.
1095 	 */
1096 	for_each_possible_cpu(cpu) {
1097 		struct per_cpu_dm_data *data = &per_cpu(dm_cpu_data, cpu);
1098 		struct sk_buff *skb;
1099 
1100 		del_timer_sync(&data->send_timer);
1101 		cancel_work_sync(&data->dm_alert_work);
1102 		while ((skb = __skb_dequeue(&data->drop_queue)))
1103 			consume_skb(skb);
1104 	}
1105 
1106 	list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
1107 		if (new_stat->dev == NULL) {
1108 			list_del_rcu(&new_stat->list);
1109 			kfree_rcu(new_stat, rcu);
1110 		}
1111 	}
1112 
1113 	module_put(THIS_MODULE);
1114 }
1115 
1116 static int set_all_monitor_traces(int state, struct netlink_ext_ack *extack)
1117 {
1118 	int rc = 0;
1119 
1120 	if (state == trace_state) {
1121 		NL_SET_ERR_MSG_MOD(extack, "Trace state already set to requested state");
1122 		return -EAGAIN;
1123 	}
1124 
1125 	switch (state) {
1126 	case TRACE_ON:
1127 		rc = net_dm_trace_on_set(extack);
1128 		break;
1129 	case TRACE_OFF:
1130 		net_dm_trace_off_set();
1131 		break;
1132 	default:
1133 		rc = 1;
1134 		break;
1135 	}
1136 
1137 	if (!rc)
1138 		trace_state = state;
1139 	else
1140 		rc = -EINPROGRESS;
1141 
1142 	return rc;
1143 }
1144 
1145 static bool net_dm_is_monitoring(void)
1146 {
1147 	return trace_state == TRACE_ON || monitor_hw;
1148 }
1149 
1150 static int net_dm_alert_mode_get_from_info(struct genl_info *info,
1151 					   enum net_dm_alert_mode *p_alert_mode)
1152 {
1153 	u8 val;
1154 
1155 	val = nla_get_u8(info->attrs[NET_DM_ATTR_ALERT_MODE]);
1156 
1157 	switch (val) {
1158 	case NET_DM_ALERT_MODE_SUMMARY: /* fall-through */
1159 	case NET_DM_ALERT_MODE_PACKET:
1160 		*p_alert_mode = val;
1161 		break;
1162 	default:
1163 		return -EINVAL;
1164 	}
1165 
1166 	return 0;
1167 }
1168 
1169 static int net_dm_alert_mode_set(struct genl_info *info)
1170 {
1171 	struct netlink_ext_ack *extack = info->extack;
1172 	enum net_dm_alert_mode alert_mode;
1173 	int rc;
1174 
1175 	if (!info->attrs[NET_DM_ATTR_ALERT_MODE])
1176 		return 0;
1177 
1178 	rc = net_dm_alert_mode_get_from_info(info, &alert_mode);
1179 	if (rc) {
1180 		NL_SET_ERR_MSG_MOD(extack, "Invalid alert mode");
1181 		return -EINVAL;
1182 	}
1183 
1184 	net_dm_alert_mode = alert_mode;
1185 
1186 	return 0;
1187 }
1188 
1189 static void net_dm_trunc_len_set(struct genl_info *info)
1190 {
1191 	if (!info->attrs[NET_DM_ATTR_TRUNC_LEN])
1192 		return;
1193 
1194 	net_dm_trunc_len = nla_get_u32(info->attrs[NET_DM_ATTR_TRUNC_LEN]);
1195 }
1196 
1197 static void net_dm_queue_len_set(struct genl_info *info)
1198 {
1199 	if (!info->attrs[NET_DM_ATTR_QUEUE_LEN])
1200 		return;
1201 
1202 	net_dm_queue_len = nla_get_u32(info->attrs[NET_DM_ATTR_QUEUE_LEN]);
1203 }
1204 
1205 static int net_dm_cmd_config(struct sk_buff *skb,
1206 			struct genl_info *info)
1207 {
1208 	struct netlink_ext_ack *extack = info->extack;
1209 	int rc;
1210 
1211 	if (net_dm_is_monitoring()) {
1212 		NL_SET_ERR_MSG_MOD(extack, "Cannot configure drop monitor during monitoring");
1213 		return -EBUSY;
1214 	}
1215 
1216 	rc = net_dm_alert_mode_set(info);
1217 	if (rc)
1218 		return rc;
1219 
1220 	net_dm_trunc_len_set(info);
1221 
1222 	net_dm_queue_len_set(info);
1223 
1224 	return 0;
1225 }
1226 
1227 static int net_dm_monitor_start(bool set_sw, bool set_hw,
1228 				struct netlink_ext_ack *extack)
1229 {
1230 	bool sw_set = false;
1231 	int rc;
1232 
1233 	if (set_sw) {
1234 		rc = set_all_monitor_traces(TRACE_ON, extack);
1235 		if (rc)
1236 			return rc;
1237 		sw_set = true;
1238 	}
1239 
1240 	if (set_hw) {
1241 		rc = net_dm_hw_monitor_start(extack);
1242 		if (rc)
1243 			goto err_monitor_hw;
1244 	}
1245 
1246 	return 0;
1247 
1248 err_monitor_hw:
1249 	if (sw_set)
1250 		set_all_monitor_traces(TRACE_OFF, extack);
1251 	return rc;
1252 }
1253 
1254 static void net_dm_monitor_stop(bool set_sw, bool set_hw,
1255 				struct netlink_ext_ack *extack)
1256 {
1257 	if (set_hw)
1258 		net_dm_hw_monitor_stop(extack);
1259 	if (set_sw)
1260 		set_all_monitor_traces(TRACE_OFF, extack);
1261 }
1262 
1263 static int net_dm_cmd_trace(struct sk_buff *skb,
1264 			struct genl_info *info)
1265 {
1266 	bool set_sw = !!info->attrs[NET_DM_ATTR_SW_DROPS];
1267 	bool set_hw = !!info->attrs[NET_DM_ATTR_HW_DROPS];
1268 	struct netlink_ext_ack *extack = info->extack;
1269 
1270 	/* To maintain backward compatibility, we start / stop monitoring of
1271 	 * software drops if no flag is specified.
1272 	 */
1273 	if (!set_sw && !set_hw)
1274 		set_sw = true;
1275 
1276 	switch (info->genlhdr->cmd) {
1277 	case NET_DM_CMD_START:
1278 		return net_dm_monitor_start(set_sw, set_hw, extack);
1279 	case NET_DM_CMD_STOP:
1280 		net_dm_monitor_stop(set_sw, set_hw, extack);
1281 		return 0;
1282 	}
1283 
1284 	return -EOPNOTSUPP;
1285 }
1286 
1287 static int net_dm_config_fill(struct sk_buff *msg, struct genl_info *info)
1288 {
1289 	void *hdr;
1290 
1291 	hdr = genlmsg_put(msg, info->snd_portid, info->snd_seq,
1292 			  &net_drop_monitor_family, 0, NET_DM_CMD_CONFIG_NEW);
1293 	if (!hdr)
1294 		return -EMSGSIZE;
1295 
1296 	if (nla_put_u8(msg, NET_DM_ATTR_ALERT_MODE, net_dm_alert_mode))
1297 		goto nla_put_failure;
1298 
1299 	if (nla_put_u32(msg, NET_DM_ATTR_TRUNC_LEN, net_dm_trunc_len))
1300 		goto nla_put_failure;
1301 
1302 	if (nla_put_u32(msg, NET_DM_ATTR_QUEUE_LEN, net_dm_queue_len))
1303 		goto nla_put_failure;
1304 
1305 	genlmsg_end(msg, hdr);
1306 
1307 	return 0;
1308 
1309 nla_put_failure:
1310 	genlmsg_cancel(msg, hdr);
1311 	return -EMSGSIZE;
1312 }
1313 
1314 static int net_dm_cmd_config_get(struct sk_buff *skb, struct genl_info *info)
1315 {
1316 	struct sk_buff *msg;
1317 	int rc;
1318 
1319 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1320 	if (!msg)
1321 		return -ENOMEM;
1322 
1323 	rc = net_dm_config_fill(msg, info);
1324 	if (rc)
1325 		goto free_msg;
1326 
1327 	return genlmsg_reply(msg, info);
1328 
1329 free_msg:
1330 	nlmsg_free(msg);
1331 	return rc;
1332 }
1333 
1334 static void net_dm_stats_read(struct net_dm_stats *stats)
1335 {
1336 	int cpu;
1337 
1338 	memset(stats, 0, sizeof(*stats));
1339 	for_each_possible_cpu(cpu) {
1340 		struct per_cpu_dm_data *data = &per_cpu(dm_cpu_data, cpu);
1341 		struct net_dm_stats *cpu_stats = &data->stats;
1342 		unsigned int start;
1343 		u64 dropped;
1344 
1345 		do {
1346 			start = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
1347 			dropped = cpu_stats->dropped;
1348 		} while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, start));
1349 
1350 		stats->dropped += dropped;
1351 	}
1352 }
1353 
1354 static int net_dm_stats_put(struct sk_buff *msg)
1355 {
1356 	struct net_dm_stats stats;
1357 	struct nlattr *attr;
1358 
1359 	net_dm_stats_read(&stats);
1360 
1361 	attr = nla_nest_start(msg, NET_DM_ATTR_STATS);
1362 	if (!attr)
1363 		return -EMSGSIZE;
1364 
1365 	if (nla_put_u64_64bit(msg, NET_DM_ATTR_STATS_DROPPED,
1366 			      stats.dropped, NET_DM_ATTR_PAD))
1367 		goto nla_put_failure;
1368 
1369 	nla_nest_end(msg, attr);
1370 
1371 	return 0;
1372 
1373 nla_put_failure:
1374 	nla_nest_cancel(msg, attr);
1375 	return -EMSGSIZE;
1376 }
1377 
1378 static void net_dm_hw_stats_read(struct net_dm_stats *stats)
1379 {
1380 	int cpu;
1381 
1382 	memset(stats, 0, sizeof(*stats));
1383 	for_each_possible_cpu(cpu) {
1384 		struct per_cpu_dm_data *hw_data = &per_cpu(dm_hw_cpu_data, cpu);
1385 		struct net_dm_stats *cpu_stats = &hw_data->stats;
1386 		unsigned int start;
1387 		u64 dropped;
1388 
1389 		do {
1390 			start = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
1391 			dropped = cpu_stats->dropped;
1392 		} while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, start));
1393 
1394 		stats->dropped += dropped;
1395 	}
1396 }
1397 
1398 static int net_dm_hw_stats_put(struct sk_buff *msg)
1399 {
1400 	struct net_dm_stats stats;
1401 	struct nlattr *attr;
1402 
1403 	net_dm_hw_stats_read(&stats);
1404 
1405 	attr = nla_nest_start(msg, NET_DM_ATTR_HW_STATS);
1406 	if (!attr)
1407 		return -EMSGSIZE;
1408 
1409 	if (nla_put_u64_64bit(msg, NET_DM_ATTR_STATS_DROPPED,
1410 			      stats.dropped, NET_DM_ATTR_PAD))
1411 		goto nla_put_failure;
1412 
1413 	nla_nest_end(msg, attr);
1414 
1415 	return 0;
1416 
1417 nla_put_failure:
1418 	nla_nest_cancel(msg, attr);
1419 	return -EMSGSIZE;
1420 }
1421 
1422 static int net_dm_stats_fill(struct sk_buff *msg, struct genl_info *info)
1423 {
1424 	void *hdr;
1425 	int rc;
1426 
1427 	hdr = genlmsg_put(msg, info->snd_portid, info->snd_seq,
1428 			  &net_drop_monitor_family, 0, NET_DM_CMD_STATS_NEW);
1429 	if (!hdr)
1430 		return -EMSGSIZE;
1431 
1432 	rc = net_dm_stats_put(msg);
1433 	if (rc)
1434 		goto nla_put_failure;
1435 
1436 	rc = net_dm_hw_stats_put(msg);
1437 	if (rc)
1438 		goto nla_put_failure;
1439 
1440 	genlmsg_end(msg, hdr);
1441 
1442 	return 0;
1443 
1444 nla_put_failure:
1445 	genlmsg_cancel(msg, hdr);
1446 	return -EMSGSIZE;
1447 }
1448 
1449 static int net_dm_cmd_stats_get(struct sk_buff *skb, struct genl_info *info)
1450 {
1451 	struct sk_buff *msg;
1452 	int rc;
1453 
1454 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1455 	if (!msg)
1456 		return -ENOMEM;
1457 
1458 	rc = net_dm_stats_fill(msg, info);
1459 	if (rc)
1460 		goto free_msg;
1461 
1462 	return genlmsg_reply(msg, info);
1463 
1464 free_msg:
1465 	nlmsg_free(msg);
1466 	return rc;
1467 }
1468 
1469 static int dropmon_net_event(struct notifier_block *ev_block,
1470 			     unsigned long event, void *ptr)
1471 {
1472 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1473 	struct dm_hw_stat_delta *new_stat = NULL;
1474 	struct dm_hw_stat_delta *tmp;
1475 
1476 	switch (event) {
1477 	case NETDEV_REGISTER:
1478 		new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
1479 
1480 		if (!new_stat)
1481 			goto out;
1482 
1483 		new_stat->dev = dev;
1484 		new_stat->last_rx = jiffies;
1485 		mutex_lock(&net_dm_mutex);
1486 		list_add_rcu(&new_stat->list, &hw_stats_list);
1487 		mutex_unlock(&net_dm_mutex);
1488 		break;
1489 	case NETDEV_UNREGISTER:
1490 		mutex_lock(&net_dm_mutex);
1491 		list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
1492 			if (new_stat->dev == dev) {
1493 				new_stat->dev = NULL;
1494 				if (trace_state == TRACE_OFF) {
1495 					list_del_rcu(&new_stat->list);
1496 					kfree_rcu(new_stat, rcu);
1497 					break;
1498 				}
1499 			}
1500 		}
1501 		mutex_unlock(&net_dm_mutex);
1502 		break;
1503 	}
1504 out:
1505 	return NOTIFY_DONE;
1506 }
1507 
1508 static const struct nla_policy net_dm_nl_policy[NET_DM_ATTR_MAX + 1] = {
1509 	[NET_DM_ATTR_UNSPEC] = { .strict_start_type = NET_DM_ATTR_UNSPEC + 1 },
1510 	[NET_DM_ATTR_ALERT_MODE] = { .type = NLA_U8 },
1511 	[NET_DM_ATTR_TRUNC_LEN] = { .type = NLA_U32 },
1512 	[NET_DM_ATTR_QUEUE_LEN] = { .type = NLA_U32 },
1513 	[NET_DM_ATTR_SW_DROPS]	= {. type = NLA_FLAG },
1514 	[NET_DM_ATTR_HW_DROPS]	= {. type = NLA_FLAG },
1515 };
1516 
1517 static const struct genl_ops dropmon_ops[] = {
1518 	{
1519 		.cmd = NET_DM_CMD_CONFIG,
1520 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
1521 		.doit = net_dm_cmd_config,
1522 		.flags = GENL_ADMIN_PERM,
1523 	},
1524 	{
1525 		.cmd = NET_DM_CMD_START,
1526 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
1527 		.doit = net_dm_cmd_trace,
1528 	},
1529 	{
1530 		.cmd = NET_DM_CMD_STOP,
1531 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
1532 		.doit = net_dm_cmd_trace,
1533 	},
1534 	{
1535 		.cmd = NET_DM_CMD_CONFIG_GET,
1536 		.doit = net_dm_cmd_config_get,
1537 	},
1538 	{
1539 		.cmd = NET_DM_CMD_STATS_GET,
1540 		.doit = net_dm_cmd_stats_get,
1541 	},
1542 };
1543 
1544 static int net_dm_nl_pre_doit(const struct genl_ops *ops,
1545 			      struct sk_buff *skb, struct genl_info *info)
1546 {
1547 	mutex_lock(&net_dm_mutex);
1548 
1549 	return 0;
1550 }
1551 
1552 static void net_dm_nl_post_doit(const struct genl_ops *ops,
1553 				struct sk_buff *skb, struct genl_info *info)
1554 {
1555 	mutex_unlock(&net_dm_mutex);
1556 }
1557 
1558 static struct genl_family net_drop_monitor_family __ro_after_init = {
1559 	.hdrsize        = 0,
1560 	.name           = "NET_DM",
1561 	.version        = 2,
1562 	.maxattr	= NET_DM_ATTR_MAX,
1563 	.policy		= net_dm_nl_policy,
1564 	.pre_doit	= net_dm_nl_pre_doit,
1565 	.post_doit	= net_dm_nl_post_doit,
1566 	.module		= THIS_MODULE,
1567 	.ops		= dropmon_ops,
1568 	.n_ops		= ARRAY_SIZE(dropmon_ops),
1569 	.mcgrps		= dropmon_mcgrps,
1570 	.n_mcgrps	= ARRAY_SIZE(dropmon_mcgrps),
1571 };
1572 
1573 static struct notifier_block dropmon_net_notifier = {
1574 	.notifier_call = dropmon_net_event
1575 };
1576 
1577 static void __net_dm_cpu_data_init(struct per_cpu_dm_data *data)
1578 {
1579 	spin_lock_init(&data->lock);
1580 	skb_queue_head_init(&data->drop_queue);
1581 	u64_stats_init(&data->stats.syncp);
1582 }
1583 
1584 static void __net_dm_cpu_data_fini(struct per_cpu_dm_data *data)
1585 {
1586 	WARN_ON(!skb_queue_empty(&data->drop_queue));
1587 }
1588 
1589 static void net_dm_cpu_data_init(int cpu)
1590 {
1591 	struct per_cpu_dm_data *data;
1592 
1593 	data = &per_cpu(dm_cpu_data, cpu);
1594 	__net_dm_cpu_data_init(data);
1595 }
1596 
1597 static void net_dm_cpu_data_fini(int cpu)
1598 {
1599 	struct per_cpu_dm_data *data;
1600 
1601 	data = &per_cpu(dm_cpu_data, cpu);
1602 	/* At this point, we should have exclusive access
1603 	 * to this struct and can free the skb inside it.
1604 	 */
1605 	consume_skb(data->skb);
1606 	__net_dm_cpu_data_fini(data);
1607 }
1608 
1609 static void net_dm_hw_cpu_data_init(int cpu)
1610 {
1611 	struct per_cpu_dm_data *hw_data;
1612 
1613 	hw_data = &per_cpu(dm_hw_cpu_data, cpu);
1614 	__net_dm_cpu_data_init(hw_data);
1615 }
1616 
1617 static void net_dm_hw_cpu_data_fini(int cpu)
1618 {
1619 	struct per_cpu_dm_data *hw_data;
1620 
1621 	hw_data = &per_cpu(dm_hw_cpu_data, cpu);
1622 	kfree(hw_data->hw_entries);
1623 	__net_dm_cpu_data_fini(hw_data);
1624 }
1625 
1626 static int __init init_net_drop_monitor(void)
1627 {
1628 	int cpu, rc;
1629 
1630 	pr_info("Initializing network drop monitor service\n");
1631 
1632 	if (sizeof(void *) > 8) {
1633 		pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
1634 		return -ENOSPC;
1635 	}
1636 
1637 	rc = genl_register_family(&net_drop_monitor_family);
1638 	if (rc) {
1639 		pr_err("Could not create drop monitor netlink family\n");
1640 		return rc;
1641 	}
1642 	WARN_ON(net_drop_monitor_family.mcgrp_offset != NET_DM_GRP_ALERT);
1643 
1644 	rc = register_netdevice_notifier(&dropmon_net_notifier);
1645 	if (rc < 0) {
1646 		pr_crit("Failed to register netdevice notifier\n");
1647 		goto out_unreg;
1648 	}
1649 
1650 	rc = 0;
1651 
1652 	for_each_possible_cpu(cpu) {
1653 		net_dm_cpu_data_init(cpu);
1654 		net_dm_hw_cpu_data_init(cpu);
1655 	}
1656 
1657 	goto out;
1658 
1659 out_unreg:
1660 	genl_unregister_family(&net_drop_monitor_family);
1661 out:
1662 	return rc;
1663 }
1664 
1665 static void exit_net_drop_monitor(void)
1666 {
1667 	int cpu;
1668 
1669 	BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
1670 
1671 	/*
1672 	 * Because of the module_get/put we do in the trace state change path
1673 	 * we are guarnateed not to have any current users when we get here
1674 	 */
1675 
1676 	for_each_possible_cpu(cpu) {
1677 		net_dm_hw_cpu_data_fini(cpu);
1678 		net_dm_cpu_data_fini(cpu);
1679 	}
1680 
1681 	BUG_ON(genl_unregister_family(&net_drop_monitor_family));
1682 }
1683 
1684 module_init(init_net_drop_monitor);
1685 module_exit(exit_net_drop_monitor);
1686 
1687 MODULE_LICENSE("GPL v2");
1688 MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
1689 MODULE_ALIAS_GENL_FAMILY("NET_DM");
1690