xref: /openbmc/linux/net/core/drop_monitor.c (revision 5a158981)
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 = kmemdup(hw_metadata->trap_group_name,
806 				  strlen(hw_metadata->trap_group_name) + 1,
807 				  GFP_ATOMIC | __GFP_ZERO);
808 	if (!trap_group_name)
809 		goto free_hw_metadata;
810 	n_hw_metadata->trap_group_name = trap_group_name;
811 
812 	trap_name = kmemdup(hw_metadata->trap_name,
813 			    strlen(hw_metadata->trap_name) + 1,
814 			    GFP_ATOMIC | __GFP_ZERO);
815 	if (!trap_name)
816 		goto free_trap_group;
817 	n_hw_metadata->trap_name = trap_name;
818 
819 	n_hw_metadata->input_dev = hw_metadata->input_dev;
820 	if (n_hw_metadata->input_dev)
821 		dev_hold(n_hw_metadata->input_dev);
822 
823 	return n_hw_metadata;
824 
825 free_trap_group:
826 	kfree(trap_group_name);
827 free_hw_metadata:
828 	kfree(n_hw_metadata);
829 	return NULL;
830 }
831 
832 static void
833 net_dm_hw_metadata_free(const struct net_dm_hw_metadata *hw_metadata)
834 {
835 	if (hw_metadata->input_dev)
836 		dev_put(hw_metadata->input_dev);
837 	kfree(hw_metadata->trap_name);
838 	kfree(hw_metadata->trap_group_name);
839 	kfree(hw_metadata);
840 }
841 
842 static void net_dm_hw_packet_report(struct sk_buff *skb)
843 {
844 	struct net_dm_hw_metadata *hw_metadata;
845 	struct sk_buff *msg;
846 	size_t payload_len;
847 	int rc;
848 
849 	if (skb->data > skb_mac_header(skb))
850 		skb_push(skb, skb->data - skb_mac_header(skb));
851 	else
852 		skb_pull(skb, skb_mac_header(skb) - skb->data);
853 
854 	payload_len = min_t(size_t, skb->len, NET_DM_MAX_PACKET_SIZE);
855 	if (net_dm_trunc_len)
856 		payload_len = min_t(size_t, net_dm_trunc_len, payload_len);
857 
858 	hw_metadata = NET_DM_SKB_CB(skb)->hw_metadata;
859 	msg = nlmsg_new(net_dm_hw_packet_report_size(payload_len, hw_metadata),
860 			GFP_KERNEL);
861 	if (!msg)
862 		goto out;
863 
864 	rc = net_dm_hw_packet_report_fill(msg, skb, payload_len);
865 	if (rc) {
866 		nlmsg_free(msg);
867 		goto out;
868 	}
869 
870 	genlmsg_multicast(&net_drop_monitor_family, msg, 0, 0, GFP_KERNEL);
871 
872 out:
873 	net_dm_hw_metadata_free(NET_DM_SKB_CB(skb)->hw_metadata);
874 	consume_skb(skb);
875 }
876 
877 static void net_dm_hw_packet_work(struct work_struct *work)
878 {
879 	struct per_cpu_dm_data *hw_data;
880 	struct sk_buff_head list;
881 	struct sk_buff *skb;
882 	unsigned long flags;
883 
884 	hw_data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
885 
886 	__skb_queue_head_init(&list);
887 
888 	spin_lock_irqsave(&hw_data->drop_queue.lock, flags);
889 	skb_queue_splice_tail_init(&hw_data->drop_queue, &list);
890 	spin_unlock_irqrestore(&hw_data->drop_queue.lock, flags);
891 
892 	while ((skb = __skb_dequeue(&list)))
893 		net_dm_hw_packet_report(skb);
894 }
895 
896 static void
897 net_dm_hw_packet_probe(struct sk_buff *skb,
898 		       const struct net_dm_hw_metadata *hw_metadata)
899 {
900 	struct net_dm_hw_metadata *n_hw_metadata;
901 	ktime_t tstamp = ktime_get_real();
902 	struct per_cpu_dm_data *hw_data;
903 	struct sk_buff *nskb;
904 	unsigned long flags;
905 
906 	if (!skb_mac_header_was_set(skb))
907 		return;
908 
909 	nskb = skb_clone(skb, GFP_ATOMIC);
910 	if (!nskb)
911 		return;
912 
913 	n_hw_metadata = net_dm_hw_metadata_clone(hw_metadata);
914 	if (!n_hw_metadata)
915 		goto free;
916 
917 	NET_DM_SKB_CB(nskb)->hw_metadata = n_hw_metadata;
918 	nskb->tstamp = tstamp;
919 
920 	hw_data = this_cpu_ptr(&dm_hw_cpu_data);
921 
922 	spin_lock_irqsave(&hw_data->drop_queue.lock, flags);
923 	if (skb_queue_len(&hw_data->drop_queue) < net_dm_queue_len)
924 		__skb_queue_tail(&hw_data->drop_queue, nskb);
925 	else
926 		goto unlock_free;
927 	spin_unlock_irqrestore(&hw_data->drop_queue.lock, flags);
928 
929 	schedule_work(&hw_data->dm_alert_work);
930 
931 	return;
932 
933 unlock_free:
934 	spin_unlock_irqrestore(&hw_data->drop_queue.lock, flags);
935 	u64_stats_update_begin(&hw_data->stats.syncp);
936 	hw_data->stats.dropped++;
937 	u64_stats_update_end(&hw_data->stats.syncp);
938 	net_dm_hw_metadata_free(n_hw_metadata);
939 free:
940 	consume_skb(nskb);
941 }
942 
943 static const struct net_dm_alert_ops net_dm_alert_packet_ops = {
944 	.kfree_skb_probe	= net_dm_packet_trace_kfree_skb_hit,
945 	.napi_poll_probe	= net_dm_packet_trace_napi_poll_hit,
946 	.work_item_func		= net_dm_packet_work,
947 	.hw_work_item_func	= net_dm_hw_packet_work,
948 	.hw_probe		= net_dm_hw_packet_probe,
949 };
950 
951 static const struct net_dm_alert_ops *net_dm_alert_ops_arr[] = {
952 	[NET_DM_ALERT_MODE_SUMMARY]	= &net_dm_alert_summary_ops,
953 	[NET_DM_ALERT_MODE_PACKET]	= &net_dm_alert_packet_ops,
954 };
955 
956 void net_dm_hw_report(struct sk_buff *skb,
957 		      const struct net_dm_hw_metadata *hw_metadata)
958 {
959 	rcu_read_lock();
960 
961 	if (!monitor_hw)
962 		goto out;
963 
964 	net_dm_alert_ops_arr[net_dm_alert_mode]->hw_probe(skb, hw_metadata);
965 
966 out:
967 	rcu_read_unlock();
968 }
969 EXPORT_SYMBOL_GPL(net_dm_hw_report);
970 
971 static int net_dm_hw_monitor_start(struct netlink_ext_ack *extack)
972 {
973 	const struct net_dm_alert_ops *ops;
974 	int cpu;
975 
976 	if (monitor_hw) {
977 		NL_SET_ERR_MSG_MOD(extack, "Hardware monitoring already enabled");
978 		return -EAGAIN;
979 	}
980 
981 	ops = net_dm_alert_ops_arr[net_dm_alert_mode];
982 
983 	if (!try_module_get(THIS_MODULE)) {
984 		NL_SET_ERR_MSG_MOD(extack, "Failed to take reference on module");
985 		return -ENODEV;
986 	}
987 
988 	for_each_possible_cpu(cpu) {
989 		struct per_cpu_dm_data *hw_data = &per_cpu(dm_hw_cpu_data, cpu);
990 		struct net_dm_hw_entries *hw_entries;
991 
992 		INIT_WORK(&hw_data->dm_alert_work, ops->hw_work_item_func);
993 		timer_setup(&hw_data->send_timer, sched_send_work, 0);
994 		hw_entries = net_dm_hw_reset_per_cpu_data(hw_data);
995 		kfree(hw_entries);
996 	}
997 
998 	monitor_hw = true;
999 
1000 	return 0;
1001 }
1002 
1003 static void net_dm_hw_monitor_stop(struct netlink_ext_ack *extack)
1004 {
1005 	int cpu;
1006 
1007 	if (!monitor_hw)
1008 		NL_SET_ERR_MSG_MOD(extack, "Hardware monitoring already disabled");
1009 
1010 	monitor_hw = false;
1011 
1012 	/* After this call returns we are guaranteed that no CPU is processing
1013 	 * any hardware drops.
1014 	 */
1015 	synchronize_rcu();
1016 
1017 	for_each_possible_cpu(cpu) {
1018 		struct per_cpu_dm_data *hw_data = &per_cpu(dm_hw_cpu_data, cpu);
1019 		struct sk_buff *skb;
1020 
1021 		del_timer_sync(&hw_data->send_timer);
1022 		cancel_work_sync(&hw_data->dm_alert_work);
1023 		while ((skb = __skb_dequeue(&hw_data->drop_queue))) {
1024 			struct net_dm_hw_metadata *hw_metadata;
1025 
1026 			hw_metadata = NET_DM_SKB_CB(skb)->hw_metadata;
1027 			net_dm_hw_metadata_free(hw_metadata);
1028 			consume_skb(skb);
1029 		}
1030 	}
1031 
1032 	module_put(THIS_MODULE);
1033 }
1034 
1035 static int net_dm_trace_on_set(struct netlink_ext_ack *extack)
1036 {
1037 	const struct net_dm_alert_ops *ops;
1038 	int cpu, rc;
1039 
1040 	ops = net_dm_alert_ops_arr[net_dm_alert_mode];
1041 
1042 	if (!try_module_get(THIS_MODULE)) {
1043 		NL_SET_ERR_MSG_MOD(extack, "Failed to take reference on module");
1044 		return -ENODEV;
1045 	}
1046 
1047 	for_each_possible_cpu(cpu) {
1048 		struct per_cpu_dm_data *data = &per_cpu(dm_cpu_data, cpu);
1049 		struct sk_buff *skb;
1050 
1051 		INIT_WORK(&data->dm_alert_work, ops->work_item_func);
1052 		timer_setup(&data->send_timer, sched_send_work, 0);
1053 		/* Allocate a new per-CPU skb for the summary alert message and
1054 		 * free the old one which might contain stale data from
1055 		 * previous tracing.
1056 		 */
1057 		skb = reset_per_cpu_data(data);
1058 		consume_skb(skb);
1059 	}
1060 
1061 	rc = register_trace_kfree_skb(ops->kfree_skb_probe, NULL);
1062 	if (rc) {
1063 		NL_SET_ERR_MSG_MOD(extack, "Failed to connect probe to kfree_skb() tracepoint");
1064 		goto err_module_put;
1065 	}
1066 
1067 	rc = register_trace_napi_poll(ops->napi_poll_probe, NULL);
1068 	if (rc) {
1069 		NL_SET_ERR_MSG_MOD(extack, "Failed to connect probe to napi_poll() tracepoint");
1070 		goto err_unregister_trace;
1071 	}
1072 
1073 	return 0;
1074 
1075 err_unregister_trace:
1076 	unregister_trace_kfree_skb(ops->kfree_skb_probe, NULL);
1077 err_module_put:
1078 	module_put(THIS_MODULE);
1079 	return rc;
1080 }
1081 
1082 static void net_dm_trace_off_set(void)
1083 {
1084 	struct dm_hw_stat_delta *new_stat, *temp;
1085 	const struct net_dm_alert_ops *ops;
1086 	int cpu;
1087 
1088 	ops = net_dm_alert_ops_arr[net_dm_alert_mode];
1089 
1090 	unregister_trace_napi_poll(ops->napi_poll_probe, NULL);
1091 	unregister_trace_kfree_skb(ops->kfree_skb_probe, NULL);
1092 
1093 	tracepoint_synchronize_unregister();
1094 
1095 	/* Make sure we do not send notifications to user space after request
1096 	 * to stop tracing returns.
1097 	 */
1098 	for_each_possible_cpu(cpu) {
1099 		struct per_cpu_dm_data *data = &per_cpu(dm_cpu_data, cpu);
1100 		struct sk_buff *skb;
1101 
1102 		del_timer_sync(&data->send_timer);
1103 		cancel_work_sync(&data->dm_alert_work);
1104 		while ((skb = __skb_dequeue(&data->drop_queue)))
1105 			consume_skb(skb);
1106 	}
1107 
1108 	list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
1109 		if (new_stat->dev == NULL) {
1110 			list_del_rcu(&new_stat->list);
1111 			kfree_rcu(new_stat, rcu);
1112 		}
1113 	}
1114 
1115 	module_put(THIS_MODULE);
1116 }
1117 
1118 static int set_all_monitor_traces(int state, struct netlink_ext_ack *extack)
1119 {
1120 	int rc = 0;
1121 
1122 	if (state == trace_state) {
1123 		NL_SET_ERR_MSG_MOD(extack, "Trace state already set to requested state");
1124 		return -EAGAIN;
1125 	}
1126 
1127 	switch (state) {
1128 	case TRACE_ON:
1129 		rc = net_dm_trace_on_set(extack);
1130 		break;
1131 	case TRACE_OFF:
1132 		net_dm_trace_off_set();
1133 		break;
1134 	default:
1135 		rc = 1;
1136 		break;
1137 	}
1138 
1139 	if (!rc)
1140 		trace_state = state;
1141 	else
1142 		rc = -EINPROGRESS;
1143 
1144 	return rc;
1145 }
1146 
1147 static bool net_dm_is_monitoring(void)
1148 {
1149 	return trace_state == TRACE_ON || monitor_hw;
1150 }
1151 
1152 static int net_dm_alert_mode_get_from_info(struct genl_info *info,
1153 					   enum net_dm_alert_mode *p_alert_mode)
1154 {
1155 	u8 val;
1156 
1157 	val = nla_get_u8(info->attrs[NET_DM_ATTR_ALERT_MODE]);
1158 
1159 	switch (val) {
1160 	case NET_DM_ALERT_MODE_SUMMARY: /* fall-through */
1161 	case NET_DM_ALERT_MODE_PACKET:
1162 		*p_alert_mode = val;
1163 		break;
1164 	default:
1165 		return -EINVAL;
1166 	}
1167 
1168 	return 0;
1169 }
1170 
1171 static int net_dm_alert_mode_set(struct genl_info *info)
1172 {
1173 	struct netlink_ext_ack *extack = info->extack;
1174 	enum net_dm_alert_mode alert_mode;
1175 	int rc;
1176 
1177 	if (!info->attrs[NET_DM_ATTR_ALERT_MODE])
1178 		return 0;
1179 
1180 	rc = net_dm_alert_mode_get_from_info(info, &alert_mode);
1181 	if (rc) {
1182 		NL_SET_ERR_MSG_MOD(extack, "Invalid alert mode");
1183 		return -EINVAL;
1184 	}
1185 
1186 	net_dm_alert_mode = alert_mode;
1187 
1188 	return 0;
1189 }
1190 
1191 static void net_dm_trunc_len_set(struct genl_info *info)
1192 {
1193 	if (!info->attrs[NET_DM_ATTR_TRUNC_LEN])
1194 		return;
1195 
1196 	net_dm_trunc_len = nla_get_u32(info->attrs[NET_DM_ATTR_TRUNC_LEN]);
1197 }
1198 
1199 static void net_dm_queue_len_set(struct genl_info *info)
1200 {
1201 	if (!info->attrs[NET_DM_ATTR_QUEUE_LEN])
1202 		return;
1203 
1204 	net_dm_queue_len = nla_get_u32(info->attrs[NET_DM_ATTR_QUEUE_LEN]);
1205 }
1206 
1207 static int net_dm_cmd_config(struct sk_buff *skb,
1208 			struct genl_info *info)
1209 {
1210 	struct netlink_ext_ack *extack = info->extack;
1211 	int rc;
1212 
1213 	if (net_dm_is_monitoring()) {
1214 		NL_SET_ERR_MSG_MOD(extack, "Cannot configure drop monitor during monitoring");
1215 		return -EBUSY;
1216 	}
1217 
1218 	rc = net_dm_alert_mode_set(info);
1219 	if (rc)
1220 		return rc;
1221 
1222 	net_dm_trunc_len_set(info);
1223 
1224 	net_dm_queue_len_set(info);
1225 
1226 	return 0;
1227 }
1228 
1229 static int net_dm_monitor_start(bool set_sw, bool set_hw,
1230 				struct netlink_ext_ack *extack)
1231 {
1232 	bool sw_set = false;
1233 	int rc;
1234 
1235 	if (set_sw) {
1236 		rc = set_all_monitor_traces(TRACE_ON, extack);
1237 		if (rc)
1238 			return rc;
1239 		sw_set = true;
1240 	}
1241 
1242 	if (set_hw) {
1243 		rc = net_dm_hw_monitor_start(extack);
1244 		if (rc)
1245 			goto err_monitor_hw;
1246 	}
1247 
1248 	return 0;
1249 
1250 err_monitor_hw:
1251 	if (sw_set)
1252 		set_all_monitor_traces(TRACE_OFF, extack);
1253 	return rc;
1254 }
1255 
1256 static void net_dm_monitor_stop(bool set_sw, bool set_hw,
1257 				struct netlink_ext_ack *extack)
1258 {
1259 	if (set_hw)
1260 		net_dm_hw_monitor_stop(extack);
1261 	if (set_sw)
1262 		set_all_monitor_traces(TRACE_OFF, extack);
1263 }
1264 
1265 static int net_dm_cmd_trace(struct sk_buff *skb,
1266 			struct genl_info *info)
1267 {
1268 	bool set_sw = !!info->attrs[NET_DM_ATTR_SW_DROPS];
1269 	bool set_hw = !!info->attrs[NET_DM_ATTR_HW_DROPS];
1270 	struct netlink_ext_ack *extack = info->extack;
1271 
1272 	/* To maintain backward compatibility, we start / stop monitoring of
1273 	 * software drops if no flag is specified.
1274 	 */
1275 	if (!set_sw && !set_hw)
1276 		set_sw = true;
1277 
1278 	switch (info->genlhdr->cmd) {
1279 	case NET_DM_CMD_START:
1280 		return net_dm_monitor_start(set_sw, set_hw, extack);
1281 	case NET_DM_CMD_STOP:
1282 		net_dm_monitor_stop(set_sw, set_hw, extack);
1283 		return 0;
1284 	}
1285 
1286 	return -EOPNOTSUPP;
1287 }
1288 
1289 static int net_dm_config_fill(struct sk_buff *msg, struct genl_info *info)
1290 {
1291 	void *hdr;
1292 
1293 	hdr = genlmsg_put(msg, info->snd_portid, info->snd_seq,
1294 			  &net_drop_monitor_family, 0, NET_DM_CMD_CONFIG_NEW);
1295 	if (!hdr)
1296 		return -EMSGSIZE;
1297 
1298 	if (nla_put_u8(msg, NET_DM_ATTR_ALERT_MODE, net_dm_alert_mode))
1299 		goto nla_put_failure;
1300 
1301 	if (nla_put_u32(msg, NET_DM_ATTR_TRUNC_LEN, net_dm_trunc_len))
1302 		goto nla_put_failure;
1303 
1304 	if (nla_put_u32(msg, NET_DM_ATTR_QUEUE_LEN, net_dm_queue_len))
1305 		goto nla_put_failure;
1306 
1307 	genlmsg_end(msg, hdr);
1308 
1309 	return 0;
1310 
1311 nla_put_failure:
1312 	genlmsg_cancel(msg, hdr);
1313 	return -EMSGSIZE;
1314 }
1315 
1316 static int net_dm_cmd_config_get(struct sk_buff *skb, struct genl_info *info)
1317 {
1318 	struct sk_buff *msg;
1319 	int rc;
1320 
1321 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1322 	if (!msg)
1323 		return -ENOMEM;
1324 
1325 	rc = net_dm_config_fill(msg, info);
1326 	if (rc)
1327 		goto free_msg;
1328 
1329 	return genlmsg_reply(msg, info);
1330 
1331 free_msg:
1332 	nlmsg_free(msg);
1333 	return rc;
1334 }
1335 
1336 static void net_dm_stats_read(struct net_dm_stats *stats)
1337 {
1338 	int cpu;
1339 
1340 	memset(stats, 0, sizeof(*stats));
1341 	for_each_possible_cpu(cpu) {
1342 		struct per_cpu_dm_data *data = &per_cpu(dm_cpu_data, cpu);
1343 		struct net_dm_stats *cpu_stats = &data->stats;
1344 		unsigned int start;
1345 		u64 dropped;
1346 
1347 		do {
1348 			start = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
1349 			dropped = cpu_stats->dropped;
1350 		} while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, start));
1351 
1352 		stats->dropped += dropped;
1353 	}
1354 }
1355 
1356 static int net_dm_stats_put(struct sk_buff *msg)
1357 {
1358 	struct net_dm_stats stats;
1359 	struct nlattr *attr;
1360 
1361 	net_dm_stats_read(&stats);
1362 
1363 	attr = nla_nest_start(msg, NET_DM_ATTR_STATS);
1364 	if (!attr)
1365 		return -EMSGSIZE;
1366 
1367 	if (nla_put_u64_64bit(msg, NET_DM_ATTR_STATS_DROPPED,
1368 			      stats.dropped, NET_DM_ATTR_PAD))
1369 		goto nla_put_failure;
1370 
1371 	nla_nest_end(msg, attr);
1372 
1373 	return 0;
1374 
1375 nla_put_failure:
1376 	nla_nest_cancel(msg, attr);
1377 	return -EMSGSIZE;
1378 }
1379 
1380 static void net_dm_hw_stats_read(struct net_dm_stats *stats)
1381 {
1382 	int cpu;
1383 
1384 	memset(stats, 0, sizeof(*stats));
1385 	for_each_possible_cpu(cpu) {
1386 		struct per_cpu_dm_data *hw_data = &per_cpu(dm_hw_cpu_data, cpu);
1387 		struct net_dm_stats *cpu_stats = &hw_data->stats;
1388 		unsigned int start;
1389 		u64 dropped;
1390 
1391 		do {
1392 			start = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
1393 			dropped = cpu_stats->dropped;
1394 		} while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, start));
1395 
1396 		stats->dropped += dropped;
1397 	}
1398 }
1399 
1400 static int net_dm_hw_stats_put(struct sk_buff *msg)
1401 {
1402 	struct net_dm_stats stats;
1403 	struct nlattr *attr;
1404 
1405 	net_dm_hw_stats_read(&stats);
1406 
1407 	attr = nla_nest_start(msg, NET_DM_ATTR_HW_STATS);
1408 	if (!attr)
1409 		return -EMSGSIZE;
1410 
1411 	if (nla_put_u64_64bit(msg, NET_DM_ATTR_STATS_DROPPED,
1412 			      stats.dropped, NET_DM_ATTR_PAD))
1413 		goto nla_put_failure;
1414 
1415 	nla_nest_end(msg, attr);
1416 
1417 	return 0;
1418 
1419 nla_put_failure:
1420 	nla_nest_cancel(msg, attr);
1421 	return -EMSGSIZE;
1422 }
1423 
1424 static int net_dm_stats_fill(struct sk_buff *msg, struct genl_info *info)
1425 {
1426 	void *hdr;
1427 	int rc;
1428 
1429 	hdr = genlmsg_put(msg, info->snd_portid, info->snd_seq,
1430 			  &net_drop_monitor_family, 0, NET_DM_CMD_STATS_NEW);
1431 	if (!hdr)
1432 		return -EMSGSIZE;
1433 
1434 	rc = net_dm_stats_put(msg);
1435 	if (rc)
1436 		goto nla_put_failure;
1437 
1438 	rc = net_dm_hw_stats_put(msg);
1439 	if (rc)
1440 		goto nla_put_failure;
1441 
1442 	genlmsg_end(msg, hdr);
1443 
1444 	return 0;
1445 
1446 nla_put_failure:
1447 	genlmsg_cancel(msg, hdr);
1448 	return -EMSGSIZE;
1449 }
1450 
1451 static int net_dm_cmd_stats_get(struct sk_buff *skb, struct genl_info *info)
1452 {
1453 	struct sk_buff *msg;
1454 	int rc;
1455 
1456 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1457 	if (!msg)
1458 		return -ENOMEM;
1459 
1460 	rc = net_dm_stats_fill(msg, info);
1461 	if (rc)
1462 		goto free_msg;
1463 
1464 	return genlmsg_reply(msg, info);
1465 
1466 free_msg:
1467 	nlmsg_free(msg);
1468 	return rc;
1469 }
1470 
1471 static int dropmon_net_event(struct notifier_block *ev_block,
1472 			     unsigned long event, void *ptr)
1473 {
1474 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1475 	struct dm_hw_stat_delta *new_stat = NULL;
1476 	struct dm_hw_stat_delta *tmp;
1477 
1478 	switch (event) {
1479 	case NETDEV_REGISTER:
1480 		new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
1481 
1482 		if (!new_stat)
1483 			goto out;
1484 
1485 		new_stat->dev = dev;
1486 		new_stat->last_rx = jiffies;
1487 		mutex_lock(&net_dm_mutex);
1488 		list_add_rcu(&new_stat->list, &hw_stats_list);
1489 		mutex_unlock(&net_dm_mutex);
1490 		break;
1491 	case NETDEV_UNREGISTER:
1492 		mutex_lock(&net_dm_mutex);
1493 		list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
1494 			if (new_stat->dev == dev) {
1495 				new_stat->dev = NULL;
1496 				if (trace_state == TRACE_OFF) {
1497 					list_del_rcu(&new_stat->list);
1498 					kfree_rcu(new_stat, rcu);
1499 					break;
1500 				}
1501 			}
1502 		}
1503 		mutex_unlock(&net_dm_mutex);
1504 		break;
1505 	}
1506 out:
1507 	return NOTIFY_DONE;
1508 }
1509 
1510 static const struct nla_policy net_dm_nl_policy[NET_DM_ATTR_MAX + 1] = {
1511 	[NET_DM_ATTR_UNSPEC] = { .strict_start_type = NET_DM_ATTR_UNSPEC + 1 },
1512 	[NET_DM_ATTR_ALERT_MODE] = { .type = NLA_U8 },
1513 	[NET_DM_ATTR_TRUNC_LEN] = { .type = NLA_U32 },
1514 	[NET_DM_ATTR_QUEUE_LEN] = { .type = NLA_U32 },
1515 	[NET_DM_ATTR_SW_DROPS]	= {. type = NLA_FLAG },
1516 	[NET_DM_ATTR_HW_DROPS]	= {. type = NLA_FLAG },
1517 };
1518 
1519 static const struct genl_ops dropmon_ops[] = {
1520 	{
1521 		.cmd = NET_DM_CMD_CONFIG,
1522 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
1523 		.doit = net_dm_cmd_config,
1524 		.flags = GENL_ADMIN_PERM,
1525 	},
1526 	{
1527 		.cmd = NET_DM_CMD_START,
1528 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
1529 		.doit = net_dm_cmd_trace,
1530 	},
1531 	{
1532 		.cmd = NET_DM_CMD_STOP,
1533 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
1534 		.doit = net_dm_cmd_trace,
1535 	},
1536 	{
1537 		.cmd = NET_DM_CMD_CONFIG_GET,
1538 		.doit = net_dm_cmd_config_get,
1539 	},
1540 	{
1541 		.cmd = NET_DM_CMD_STATS_GET,
1542 		.doit = net_dm_cmd_stats_get,
1543 	},
1544 };
1545 
1546 static int net_dm_nl_pre_doit(const struct genl_ops *ops,
1547 			      struct sk_buff *skb, struct genl_info *info)
1548 {
1549 	mutex_lock(&net_dm_mutex);
1550 
1551 	return 0;
1552 }
1553 
1554 static void net_dm_nl_post_doit(const struct genl_ops *ops,
1555 				struct sk_buff *skb, struct genl_info *info)
1556 {
1557 	mutex_unlock(&net_dm_mutex);
1558 }
1559 
1560 static struct genl_family net_drop_monitor_family __ro_after_init = {
1561 	.hdrsize        = 0,
1562 	.name           = "NET_DM",
1563 	.version        = 2,
1564 	.maxattr	= NET_DM_ATTR_MAX,
1565 	.policy		= net_dm_nl_policy,
1566 	.pre_doit	= net_dm_nl_pre_doit,
1567 	.post_doit	= net_dm_nl_post_doit,
1568 	.module		= THIS_MODULE,
1569 	.ops		= dropmon_ops,
1570 	.n_ops		= ARRAY_SIZE(dropmon_ops),
1571 	.mcgrps		= dropmon_mcgrps,
1572 	.n_mcgrps	= ARRAY_SIZE(dropmon_mcgrps),
1573 };
1574 
1575 static struct notifier_block dropmon_net_notifier = {
1576 	.notifier_call = dropmon_net_event
1577 };
1578 
1579 static void __net_dm_cpu_data_init(struct per_cpu_dm_data *data)
1580 {
1581 	spin_lock_init(&data->lock);
1582 	skb_queue_head_init(&data->drop_queue);
1583 	u64_stats_init(&data->stats.syncp);
1584 }
1585 
1586 static void __net_dm_cpu_data_fini(struct per_cpu_dm_data *data)
1587 {
1588 	WARN_ON(!skb_queue_empty(&data->drop_queue));
1589 }
1590 
1591 static void net_dm_cpu_data_init(int cpu)
1592 {
1593 	struct per_cpu_dm_data *data;
1594 
1595 	data = &per_cpu(dm_cpu_data, cpu);
1596 	__net_dm_cpu_data_init(data);
1597 }
1598 
1599 static void net_dm_cpu_data_fini(int cpu)
1600 {
1601 	struct per_cpu_dm_data *data;
1602 
1603 	data = &per_cpu(dm_cpu_data, cpu);
1604 	/* At this point, we should have exclusive access
1605 	 * to this struct and can free the skb inside it.
1606 	 */
1607 	consume_skb(data->skb);
1608 	__net_dm_cpu_data_fini(data);
1609 }
1610 
1611 static void net_dm_hw_cpu_data_init(int cpu)
1612 {
1613 	struct per_cpu_dm_data *hw_data;
1614 
1615 	hw_data = &per_cpu(dm_hw_cpu_data, cpu);
1616 	__net_dm_cpu_data_init(hw_data);
1617 }
1618 
1619 static void net_dm_hw_cpu_data_fini(int cpu)
1620 {
1621 	struct per_cpu_dm_data *hw_data;
1622 
1623 	hw_data = &per_cpu(dm_hw_cpu_data, cpu);
1624 	kfree(hw_data->hw_entries);
1625 	__net_dm_cpu_data_fini(hw_data);
1626 }
1627 
1628 static int __init init_net_drop_monitor(void)
1629 {
1630 	int cpu, rc;
1631 
1632 	pr_info("Initializing network drop monitor service\n");
1633 
1634 	if (sizeof(void *) > 8) {
1635 		pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
1636 		return -ENOSPC;
1637 	}
1638 
1639 	rc = genl_register_family(&net_drop_monitor_family);
1640 	if (rc) {
1641 		pr_err("Could not create drop monitor netlink family\n");
1642 		return rc;
1643 	}
1644 	WARN_ON(net_drop_monitor_family.mcgrp_offset != NET_DM_GRP_ALERT);
1645 
1646 	rc = register_netdevice_notifier(&dropmon_net_notifier);
1647 	if (rc < 0) {
1648 		pr_crit("Failed to register netdevice notifier\n");
1649 		goto out_unreg;
1650 	}
1651 
1652 	rc = 0;
1653 
1654 	for_each_possible_cpu(cpu) {
1655 		net_dm_cpu_data_init(cpu);
1656 		net_dm_hw_cpu_data_init(cpu);
1657 	}
1658 
1659 	goto out;
1660 
1661 out_unreg:
1662 	genl_unregister_family(&net_drop_monitor_family);
1663 out:
1664 	return rc;
1665 }
1666 
1667 static void exit_net_drop_monitor(void)
1668 {
1669 	int cpu;
1670 
1671 	BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
1672 
1673 	/*
1674 	 * Because of the module_get/put we do in the trace state change path
1675 	 * we are guarnateed not to have any current users when we get here
1676 	 */
1677 
1678 	for_each_possible_cpu(cpu) {
1679 		net_dm_hw_cpu_data_fini(cpu);
1680 		net_dm_cpu_data_fini(cpu);
1681 	}
1682 
1683 	BUG_ON(genl_unregister_family(&net_drop_monitor_family));
1684 }
1685 
1686 module_init(init_net_drop_monitor);
1687 module_exit(exit_net_drop_monitor);
1688 
1689 MODULE_LICENSE("GPL v2");
1690 MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
1691 MODULE_ALIAS_GENL_FAMILY("NET_DM");
1692