xref: /linux/net/core/netpoll.c (revision bcefe12eff5dca6fdfa94ed85e5bee66380d5cd9)
1 /*
2  * Common framework for low-level network console, dump, and debugger code
3  *
4  * Sep 8 2003  Matt Mackall <mpm@selenic.com>
5  *
6  * based on the netconsole code from:
7  *
8  * Copyright (C) 2001  Ingo Molnar <mingo@redhat.com>
9  * Copyright (C) 2002  Red Hat, Inc.
10  */
11 
12 #include <linux/moduleparam.h>
13 #include <linux/netdevice.h>
14 #include <linux/etherdevice.h>
15 #include <linux/string.h>
16 #include <linux/if_arp.h>
17 #include <linux/inetdevice.h>
18 #include <linux/inet.h>
19 #include <linux/interrupt.h>
20 #include <linux/netpoll.h>
21 #include <linux/sched.h>
22 #include <linux/delay.h>
23 #include <linux/rcupdate.h>
24 #include <linux/workqueue.h>
25 #include <net/tcp.h>
26 #include <net/udp.h>
27 #include <asm/unaligned.h>
28 #include <trace/events/napi.h>
29 
30 /*
31  * We maintain a small pool of fully-sized skbs, to make sure the
32  * message gets out even in extreme OOM situations.
33  */
34 
35 #define MAX_UDP_CHUNK 1460
36 #define MAX_SKBS 32
37 #define MAX_QUEUE_DEPTH (MAX_SKBS / 2)
38 
39 static struct sk_buff_head skb_pool;
40 
41 static atomic_t trapped;
42 
43 #define USEC_PER_POLL	50
44 #define NETPOLL_RX_ENABLED  1
45 #define NETPOLL_RX_DROP     2
46 
47 #define MAX_SKB_SIZE \
48 		(MAX_UDP_CHUNK + sizeof(struct udphdr) + \
49 				sizeof(struct iphdr) + sizeof(struct ethhdr))
50 
51 static void zap_completion_queue(void);
52 static void arp_reply(struct sk_buff *skb);
53 
54 static unsigned int carrier_timeout = 4;
55 module_param(carrier_timeout, uint, 0644);
56 
57 static void queue_process(struct work_struct *work)
58 {
59 	struct netpoll_info *npinfo =
60 		container_of(work, struct netpoll_info, tx_work.work);
61 	struct sk_buff *skb;
62 	unsigned long flags;
63 
64 	while ((skb = skb_dequeue(&npinfo->txq))) {
65 		struct net_device *dev = skb->dev;
66 		const struct net_device_ops *ops = dev->netdev_ops;
67 		struct netdev_queue *txq;
68 
69 		if (!netif_device_present(dev) || !netif_running(dev)) {
70 			__kfree_skb(skb);
71 			continue;
72 		}
73 
74 		txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
75 
76 		local_irq_save(flags);
77 		__netif_tx_lock(txq, smp_processor_id());
78 		if (netif_tx_queue_stopped(txq) ||
79 		    netif_tx_queue_frozen(txq) ||
80 		    ops->ndo_start_xmit(skb, dev) != NETDEV_TX_OK) {
81 			skb_queue_head(&npinfo->txq, skb);
82 			__netif_tx_unlock(txq);
83 			local_irq_restore(flags);
84 
85 			schedule_delayed_work(&npinfo->tx_work, HZ/10);
86 			return;
87 		}
88 		__netif_tx_unlock(txq);
89 		local_irq_restore(flags);
90 	}
91 }
92 
93 static __sum16 checksum_udp(struct sk_buff *skb, struct udphdr *uh,
94 			    unsigned short ulen, __be32 saddr, __be32 daddr)
95 {
96 	__wsum psum;
97 
98 	if (uh->check == 0 || skb_csum_unnecessary(skb))
99 		return 0;
100 
101 	psum = csum_tcpudp_nofold(saddr, daddr, ulen, IPPROTO_UDP, 0);
102 
103 	if (skb->ip_summed == CHECKSUM_COMPLETE &&
104 	    !csum_fold(csum_add(psum, skb->csum)))
105 		return 0;
106 
107 	skb->csum = psum;
108 
109 	return __skb_checksum_complete(skb);
110 }
111 
112 /*
113  * Check whether delayed processing was scheduled for our NIC. If so,
114  * we attempt to grab the poll lock and use ->poll() to pump the card.
115  * If this fails, either we've recursed in ->poll() or it's already
116  * running on another CPU.
117  *
118  * Note: we don't mask interrupts with this lock because we're using
119  * trylock here and interrupts are already disabled in the softirq
120  * case. Further, we test the poll_owner to avoid recursion on UP
121  * systems where the lock doesn't exist.
122  *
123  * In cases where there is bi-directional communications, reading only
124  * one message at a time can lead to packets being dropped by the
125  * network adapter, forcing superfluous retries and possibly timeouts.
126  * Thus, we set our budget to greater than 1.
127  */
128 static int poll_one_napi(struct netpoll_info *npinfo,
129 			 struct napi_struct *napi, int budget)
130 {
131 	int work;
132 
133 	/* net_rx_action's ->poll() invocations and our's are
134 	 * synchronized by this test which is only made while
135 	 * holding the napi->poll_lock.
136 	 */
137 	if (!test_bit(NAPI_STATE_SCHED, &napi->state))
138 		return budget;
139 
140 	npinfo->rx_flags |= NETPOLL_RX_DROP;
141 	atomic_inc(&trapped);
142 	set_bit(NAPI_STATE_NPSVC, &napi->state);
143 
144 	work = napi->poll(napi, budget);
145 	trace_napi_poll(napi);
146 
147 	clear_bit(NAPI_STATE_NPSVC, &napi->state);
148 	atomic_dec(&trapped);
149 	npinfo->rx_flags &= ~NETPOLL_RX_DROP;
150 
151 	return budget - work;
152 }
153 
154 static void poll_napi(struct net_device *dev)
155 {
156 	struct napi_struct *napi;
157 	int budget = 16;
158 
159 	list_for_each_entry(napi, &dev->napi_list, dev_list) {
160 		if (napi->poll_owner != smp_processor_id() &&
161 		    spin_trylock(&napi->poll_lock)) {
162 			budget = poll_one_napi(dev->npinfo, napi, budget);
163 			spin_unlock(&napi->poll_lock);
164 
165 			if (!budget)
166 				break;
167 		}
168 	}
169 }
170 
171 static void service_arp_queue(struct netpoll_info *npi)
172 {
173 	if (npi) {
174 		struct sk_buff *skb;
175 
176 		while ((skb = skb_dequeue(&npi->arp_tx)))
177 			arp_reply(skb);
178 	}
179 }
180 
181 void netpoll_poll(struct netpoll *np)
182 {
183 	struct net_device *dev = np->dev;
184 	const struct net_device_ops *ops;
185 
186 	if (!dev || !netif_running(dev))
187 		return;
188 
189 	ops = dev->netdev_ops;
190 	if (!ops->ndo_poll_controller)
191 		return;
192 
193 	/* Process pending work on NIC */
194 	ops->ndo_poll_controller(dev);
195 
196 	poll_napi(dev);
197 
198 	service_arp_queue(dev->npinfo);
199 
200 	zap_completion_queue();
201 }
202 
203 static void refill_skbs(void)
204 {
205 	struct sk_buff *skb;
206 	unsigned long flags;
207 
208 	spin_lock_irqsave(&skb_pool.lock, flags);
209 	while (skb_pool.qlen < MAX_SKBS) {
210 		skb = alloc_skb(MAX_SKB_SIZE, GFP_ATOMIC);
211 		if (!skb)
212 			break;
213 
214 		__skb_queue_tail(&skb_pool, skb);
215 	}
216 	spin_unlock_irqrestore(&skb_pool.lock, flags);
217 }
218 
219 static void zap_completion_queue(void)
220 {
221 	unsigned long flags;
222 	struct softnet_data *sd = &get_cpu_var(softnet_data);
223 
224 	if (sd->completion_queue) {
225 		struct sk_buff *clist;
226 
227 		local_irq_save(flags);
228 		clist = sd->completion_queue;
229 		sd->completion_queue = NULL;
230 		local_irq_restore(flags);
231 
232 		while (clist != NULL) {
233 			struct sk_buff *skb = clist;
234 			clist = clist->next;
235 			if (skb->destructor) {
236 				atomic_inc(&skb->users);
237 				dev_kfree_skb_any(skb); /* put this one back */
238 			} else {
239 				__kfree_skb(skb);
240 			}
241 		}
242 	}
243 
244 	put_cpu_var(softnet_data);
245 }
246 
247 static struct sk_buff *find_skb(struct netpoll *np, int len, int reserve)
248 {
249 	int count = 0;
250 	struct sk_buff *skb;
251 
252 	zap_completion_queue();
253 	refill_skbs();
254 repeat:
255 
256 	skb = alloc_skb(len, GFP_ATOMIC);
257 	if (!skb)
258 		skb = skb_dequeue(&skb_pool);
259 
260 	if (!skb) {
261 		if (++count < 10) {
262 			netpoll_poll(np);
263 			goto repeat;
264 		}
265 		return NULL;
266 	}
267 
268 	atomic_set(&skb->users, 1);
269 	skb_reserve(skb, reserve);
270 	return skb;
271 }
272 
273 static int netpoll_owner_active(struct net_device *dev)
274 {
275 	struct napi_struct *napi;
276 
277 	list_for_each_entry(napi, &dev->napi_list, dev_list) {
278 		if (napi->poll_owner == smp_processor_id())
279 			return 1;
280 	}
281 	return 0;
282 }
283 
284 static void netpoll_send_skb(struct netpoll *np, struct sk_buff *skb)
285 {
286 	int status = NETDEV_TX_BUSY;
287 	unsigned long tries;
288 	struct net_device *dev = np->dev;
289 	const struct net_device_ops *ops = dev->netdev_ops;
290 	struct netpoll_info *npinfo = np->dev->npinfo;
291 
292 	if (!npinfo || !netif_running(dev) || !netif_device_present(dev)) {
293 		__kfree_skb(skb);
294 		return;
295 	}
296 
297 	/* don't get messages out of order, and no recursion */
298 	if (skb_queue_len(&npinfo->txq) == 0 && !netpoll_owner_active(dev)) {
299 		struct netdev_queue *txq;
300 		unsigned long flags;
301 
302 		txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
303 
304 		local_irq_save(flags);
305 		/* try until next clock tick */
306 		for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
307 		     tries > 0; --tries) {
308 			if (__netif_tx_trylock(txq)) {
309 				if (!netif_tx_queue_stopped(txq)) {
310 					status = ops->ndo_start_xmit(skb, dev);
311 					if (status == NETDEV_TX_OK)
312 						txq_trans_update(txq);
313 				}
314 				__netif_tx_unlock(txq);
315 
316 				if (status == NETDEV_TX_OK)
317 					break;
318 
319 			}
320 
321 			/* tickle device maybe there is some cleanup */
322 			netpoll_poll(np);
323 
324 			udelay(USEC_PER_POLL);
325 		}
326 
327 		WARN_ONCE(!irqs_disabled(),
328 			"netpoll_send_skb(): %s enabled interrupts in poll (%pF)\n",
329 			dev->name, ops->ndo_start_xmit);
330 
331 		local_irq_restore(flags);
332 	}
333 
334 	if (status != NETDEV_TX_OK) {
335 		skb_queue_tail(&npinfo->txq, skb);
336 		schedule_delayed_work(&npinfo->tx_work,0);
337 	}
338 }
339 
340 void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
341 {
342 	int total_len, eth_len, ip_len, udp_len;
343 	struct sk_buff *skb;
344 	struct udphdr *udph;
345 	struct iphdr *iph;
346 	struct ethhdr *eth;
347 
348 	udp_len = len + sizeof(*udph);
349 	ip_len = eth_len = udp_len + sizeof(*iph);
350 	total_len = eth_len + ETH_HLEN + NET_IP_ALIGN;
351 
352 	skb = find_skb(np, total_len, total_len - len);
353 	if (!skb)
354 		return;
355 
356 	skb_copy_to_linear_data(skb, msg, len);
357 	skb->len += len;
358 
359 	skb_push(skb, sizeof(*udph));
360 	skb_reset_transport_header(skb);
361 	udph = udp_hdr(skb);
362 	udph->source = htons(np->local_port);
363 	udph->dest = htons(np->remote_port);
364 	udph->len = htons(udp_len);
365 	udph->check = 0;
366 	udph->check = csum_tcpudp_magic(np->local_ip,
367 					np->remote_ip,
368 					udp_len, IPPROTO_UDP,
369 					csum_partial(udph, udp_len, 0));
370 	if (udph->check == 0)
371 		udph->check = CSUM_MANGLED_0;
372 
373 	skb_push(skb, sizeof(*iph));
374 	skb_reset_network_header(skb);
375 	iph = ip_hdr(skb);
376 
377 	/* iph->version = 4; iph->ihl = 5; */
378 	put_unaligned(0x45, (unsigned char *)iph);
379 	iph->tos      = 0;
380 	put_unaligned(htons(ip_len), &(iph->tot_len));
381 	iph->id       = 0;
382 	iph->frag_off = 0;
383 	iph->ttl      = 64;
384 	iph->protocol = IPPROTO_UDP;
385 	iph->check    = 0;
386 	put_unaligned(np->local_ip, &(iph->saddr));
387 	put_unaligned(np->remote_ip, &(iph->daddr));
388 	iph->check    = ip_fast_csum((unsigned char *)iph, iph->ihl);
389 
390 	eth = (struct ethhdr *) skb_push(skb, ETH_HLEN);
391 	skb_reset_mac_header(skb);
392 	skb->protocol = eth->h_proto = htons(ETH_P_IP);
393 	memcpy(eth->h_source, np->dev->dev_addr, ETH_ALEN);
394 	memcpy(eth->h_dest, np->remote_mac, ETH_ALEN);
395 
396 	skb->dev = np->dev;
397 
398 	netpoll_send_skb(np, skb);
399 }
400 
401 static void arp_reply(struct sk_buff *skb)
402 {
403 	struct netpoll_info *npinfo = skb->dev->npinfo;
404 	struct arphdr *arp;
405 	unsigned char *arp_ptr;
406 	int size, type = ARPOP_REPLY, ptype = ETH_P_ARP;
407 	__be32 sip, tip;
408 	unsigned char *sha;
409 	struct sk_buff *send_skb;
410 	struct netpoll *np = NULL;
411 
412 	if (npinfo->rx_np && npinfo->rx_np->dev == skb->dev)
413 		np = npinfo->rx_np;
414 	if (!np)
415 		return;
416 
417 	/* No arp on this interface */
418 	if (skb->dev->flags & IFF_NOARP)
419 		return;
420 
421 	if (!pskb_may_pull(skb, arp_hdr_len(skb->dev)))
422 		return;
423 
424 	skb_reset_network_header(skb);
425 	skb_reset_transport_header(skb);
426 	arp = arp_hdr(skb);
427 
428 	if ((arp->ar_hrd != htons(ARPHRD_ETHER) &&
429 	     arp->ar_hrd != htons(ARPHRD_IEEE802)) ||
430 	    arp->ar_pro != htons(ETH_P_IP) ||
431 	    arp->ar_op != htons(ARPOP_REQUEST))
432 		return;
433 
434 	arp_ptr = (unsigned char *)(arp+1);
435 	/* save the location of the src hw addr */
436 	sha = arp_ptr;
437 	arp_ptr += skb->dev->addr_len;
438 	memcpy(&sip, arp_ptr, 4);
439 	arp_ptr += 4;
440 	/* if we actually cared about dst hw addr, it would get copied here */
441 	arp_ptr += skb->dev->addr_len;
442 	memcpy(&tip, arp_ptr, 4);
443 
444 	/* Should we ignore arp? */
445 	if (tip != np->local_ip ||
446 	    ipv4_is_loopback(tip) || ipv4_is_multicast(tip))
447 		return;
448 
449 	size = arp_hdr_len(skb->dev);
450 	send_skb = find_skb(np, size + LL_ALLOCATED_SPACE(np->dev),
451 			    LL_RESERVED_SPACE(np->dev));
452 
453 	if (!send_skb)
454 		return;
455 
456 	skb_reset_network_header(send_skb);
457 	arp = (struct arphdr *) skb_put(send_skb, size);
458 	send_skb->dev = skb->dev;
459 	send_skb->protocol = htons(ETH_P_ARP);
460 
461 	/* Fill the device header for the ARP frame */
462 	if (dev_hard_header(send_skb, skb->dev, ptype,
463 			    sha, np->dev->dev_addr,
464 			    send_skb->len) < 0) {
465 		kfree_skb(send_skb);
466 		return;
467 	}
468 
469 	/*
470 	 * Fill out the arp protocol part.
471 	 *
472 	 * we only support ethernet device type,
473 	 * which (according to RFC 1390) should always equal 1 (Ethernet).
474 	 */
475 
476 	arp->ar_hrd = htons(np->dev->type);
477 	arp->ar_pro = htons(ETH_P_IP);
478 	arp->ar_hln = np->dev->addr_len;
479 	arp->ar_pln = 4;
480 	arp->ar_op = htons(type);
481 
482 	arp_ptr=(unsigned char *)(arp + 1);
483 	memcpy(arp_ptr, np->dev->dev_addr, np->dev->addr_len);
484 	arp_ptr += np->dev->addr_len;
485 	memcpy(arp_ptr, &tip, 4);
486 	arp_ptr += 4;
487 	memcpy(arp_ptr, sha, np->dev->addr_len);
488 	arp_ptr += np->dev->addr_len;
489 	memcpy(arp_ptr, &sip, 4);
490 
491 	netpoll_send_skb(np, send_skb);
492 }
493 
494 int __netpoll_rx(struct sk_buff *skb)
495 {
496 	int proto, len, ulen;
497 	struct iphdr *iph;
498 	struct udphdr *uh;
499 	struct netpoll_info *npi = skb->dev->npinfo;
500 	struct netpoll *np = npi->rx_np;
501 
502 	if (!np)
503 		goto out;
504 	if (skb->dev->type != ARPHRD_ETHER)
505 		goto out;
506 
507 	/* check if netpoll clients need ARP */
508 	if (skb->protocol == htons(ETH_P_ARP) &&
509 	    atomic_read(&trapped)) {
510 		skb_queue_tail(&npi->arp_tx, skb);
511 		return 1;
512 	}
513 
514 	proto = ntohs(eth_hdr(skb)->h_proto);
515 	if (proto != ETH_P_IP)
516 		goto out;
517 	if (skb->pkt_type == PACKET_OTHERHOST)
518 		goto out;
519 	if (skb_shared(skb))
520 		goto out;
521 
522 	iph = (struct iphdr *)skb->data;
523 	if (!pskb_may_pull(skb, sizeof(struct iphdr)))
524 		goto out;
525 	if (iph->ihl < 5 || iph->version != 4)
526 		goto out;
527 	if (!pskb_may_pull(skb, iph->ihl*4))
528 		goto out;
529 	if (ip_fast_csum((u8 *)iph, iph->ihl) != 0)
530 		goto out;
531 
532 	len = ntohs(iph->tot_len);
533 	if (skb->len < len || len < iph->ihl*4)
534 		goto out;
535 
536 	/*
537 	 * Our transport medium may have padded the buffer out.
538 	 * Now We trim to the true length of the frame.
539 	 */
540 	if (pskb_trim_rcsum(skb, len))
541 		goto out;
542 
543 	if (iph->protocol != IPPROTO_UDP)
544 		goto out;
545 
546 	len -= iph->ihl*4;
547 	uh = (struct udphdr *)(((char *)iph) + iph->ihl*4);
548 	ulen = ntohs(uh->len);
549 
550 	if (ulen != len)
551 		goto out;
552 	if (checksum_udp(skb, uh, ulen, iph->saddr, iph->daddr))
553 		goto out;
554 	if (np->local_ip && np->local_ip != iph->daddr)
555 		goto out;
556 	if (np->remote_ip && np->remote_ip != iph->saddr)
557 		goto out;
558 	if (np->local_port && np->local_port != ntohs(uh->dest))
559 		goto out;
560 
561 	np->rx_hook(np, ntohs(uh->source),
562 		    (char *)(uh+1),
563 		    ulen - sizeof(struct udphdr));
564 
565 	kfree_skb(skb);
566 	return 1;
567 
568 out:
569 	if (atomic_read(&trapped)) {
570 		kfree_skb(skb);
571 		return 1;
572 	}
573 
574 	return 0;
575 }
576 
577 void netpoll_print_options(struct netpoll *np)
578 {
579 	printk(KERN_INFO "%s: local port %d\n",
580 			 np->name, np->local_port);
581 	printk(KERN_INFO "%s: local IP %pI4\n",
582 			 np->name, &np->local_ip);
583 	printk(KERN_INFO "%s: interface %s\n",
584 			 np->name, np->dev_name);
585 	printk(KERN_INFO "%s: remote port %d\n",
586 			 np->name, np->remote_port);
587 	printk(KERN_INFO "%s: remote IP %pI4\n",
588 			 np->name, &np->remote_ip);
589 	printk(KERN_INFO "%s: remote ethernet address %pM\n",
590 	                 np->name, np->remote_mac);
591 }
592 
593 int netpoll_parse_options(struct netpoll *np, char *opt)
594 {
595 	char *cur=opt, *delim;
596 
597 	if (*cur != '@') {
598 		if ((delim = strchr(cur, '@')) == NULL)
599 			goto parse_failed;
600 		*delim = 0;
601 		np->local_port = simple_strtol(cur, NULL, 10);
602 		cur = delim;
603 	}
604 	cur++;
605 
606 	if (*cur != '/') {
607 		if ((delim = strchr(cur, '/')) == NULL)
608 			goto parse_failed;
609 		*delim = 0;
610 		np->local_ip = in_aton(cur);
611 		cur = delim;
612 	}
613 	cur++;
614 
615 	if (*cur != ',') {
616 		/* parse out dev name */
617 		if ((delim = strchr(cur, ',')) == NULL)
618 			goto parse_failed;
619 		*delim = 0;
620 		strlcpy(np->dev_name, cur, sizeof(np->dev_name));
621 		cur = delim;
622 	}
623 	cur++;
624 
625 	if (*cur != '@') {
626 		/* dst port */
627 		if ((delim = strchr(cur, '@')) == NULL)
628 			goto parse_failed;
629 		*delim = 0;
630 		np->remote_port = simple_strtol(cur, NULL, 10);
631 		cur = delim;
632 	}
633 	cur++;
634 
635 	/* dst ip */
636 	if ((delim = strchr(cur, '/')) == NULL)
637 		goto parse_failed;
638 	*delim = 0;
639 	np->remote_ip = in_aton(cur);
640 	cur = delim + 1;
641 
642 	if (*cur != 0) {
643 		/* MAC address */
644 		if ((delim = strchr(cur, ':')) == NULL)
645 			goto parse_failed;
646 		*delim = 0;
647 		np->remote_mac[0] = simple_strtol(cur, NULL, 16);
648 		cur = delim + 1;
649 		if ((delim = strchr(cur, ':')) == NULL)
650 			goto parse_failed;
651 		*delim = 0;
652 		np->remote_mac[1] = simple_strtol(cur, NULL, 16);
653 		cur = delim + 1;
654 		if ((delim = strchr(cur, ':')) == NULL)
655 			goto parse_failed;
656 		*delim = 0;
657 		np->remote_mac[2] = simple_strtol(cur, NULL, 16);
658 		cur = delim + 1;
659 		if ((delim = strchr(cur, ':')) == NULL)
660 			goto parse_failed;
661 		*delim = 0;
662 		np->remote_mac[3] = simple_strtol(cur, NULL, 16);
663 		cur = delim + 1;
664 		if ((delim = strchr(cur, ':')) == NULL)
665 			goto parse_failed;
666 		*delim = 0;
667 		np->remote_mac[4] = simple_strtol(cur, NULL, 16);
668 		cur = delim + 1;
669 		np->remote_mac[5] = simple_strtol(cur, NULL, 16);
670 	}
671 
672 	netpoll_print_options(np);
673 
674 	return 0;
675 
676  parse_failed:
677 	printk(KERN_INFO "%s: couldn't parse config at %s!\n",
678 	       np->name, cur);
679 	return -1;
680 }
681 
682 int netpoll_setup(struct netpoll *np)
683 {
684 	struct net_device *ndev = NULL;
685 	struct in_device *in_dev;
686 	struct netpoll_info *npinfo;
687 	unsigned long flags;
688 	int err;
689 
690 	if (np->dev_name)
691 		ndev = dev_get_by_name(&init_net, np->dev_name);
692 	if (!ndev) {
693 		printk(KERN_ERR "%s: %s doesn't exist, aborting.\n",
694 		       np->name, np->dev_name);
695 		return -ENODEV;
696 	}
697 
698 	np->dev = ndev;
699 	if (!ndev->npinfo) {
700 		npinfo = kmalloc(sizeof(*npinfo), GFP_KERNEL);
701 		if (!npinfo) {
702 			err = -ENOMEM;
703 			goto release;
704 		}
705 
706 		npinfo->rx_flags = 0;
707 		npinfo->rx_np = NULL;
708 
709 		spin_lock_init(&npinfo->rx_lock);
710 		skb_queue_head_init(&npinfo->arp_tx);
711 		skb_queue_head_init(&npinfo->txq);
712 		INIT_DELAYED_WORK(&npinfo->tx_work, queue_process);
713 
714 		atomic_set(&npinfo->refcnt, 1);
715 	} else {
716 		npinfo = ndev->npinfo;
717 		atomic_inc(&npinfo->refcnt);
718 	}
719 
720 	if (!ndev->netdev_ops->ndo_poll_controller) {
721 		printk(KERN_ERR "%s: %s doesn't support polling, aborting.\n",
722 		       np->name, np->dev_name);
723 		err = -ENOTSUPP;
724 		goto release;
725 	}
726 
727 	if (!netif_running(ndev)) {
728 		unsigned long atmost, atleast;
729 
730 		printk(KERN_INFO "%s: device %s not up yet, forcing it\n",
731 		       np->name, np->dev_name);
732 
733 		rtnl_lock();
734 		err = dev_open(ndev);
735 		rtnl_unlock();
736 
737 		if (err) {
738 			printk(KERN_ERR "%s: failed to open %s\n",
739 			       np->name, ndev->name);
740 			goto release;
741 		}
742 
743 		atleast = jiffies + HZ/10;
744 		atmost = jiffies + carrier_timeout * HZ;
745 		while (!netif_carrier_ok(ndev)) {
746 			if (time_after(jiffies, atmost)) {
747 				printk(KERN_NOTICE
748 				       "%s: timeout waiting for carrier\n",
749 				       np->name);
750 				break;
751 			}
752 			msleep(1);
753 		}
754 
755 		/* If carrier appears to come up instantly, we don't
756 		 * trust it and pause so that we don't pump all our
757 		 * queued console messages into the bitbucket.
758 		 */
759 
760 		if (time_before(jiffies, atleast)) {
761 			printk(KERN_NOTICE "%s: carrier detect appears"
762 			       " untrustworthy, waiting 4 seconds\n",
763 			       np->name);
764 			msleep(4000);
765 		}
766 	}
767 
768 	if (!np->local_ip) {
769 		rcu_read_lock();
770 		in_dev = __in_dev_get_rcu(ndev);
771 
772 		if (!in_dev || !in_dev->ifa_list) {
773 			rcu_read_unlock();
774 			printk(KERN_ERR "%s: no IP address for %s, aborting\n",
775 			       np->name, np->dev_name);
776 			err = -EDESTADDRREQ;
777 			goto release;
778 		}
779 
780 		np->local_ip = in_dev->ifa_list->ifa_local;
781 		rcu_read_unlock();
782 		printk(KERN_INFO "%s: local IP %pI4\n", np->name, &np->local_ip);
783 	}
784 
785 	if (np->rx_hook) {
786 		spin_lock_irqsave(&npinfo->rx_lock, flags);
787 		npinfo->rx_flags |= NETPOLL_RX_ENABLED;
788 		npinfo->rx_np = np;
789 		spin_unlock_irqrestore(&npinfo->rx_lock, flags);
790 	}
791 
792 	/* fill up the skb queue */
793 	refill_skbs();
794 
795 	/* last thing to do is link it to the net device structure */
796 	ndev->npinfo = npinfo;
797 
798 	/* avoid racing with NAPI reading npinfo */
799 	synchronize_rcu();
800 
801 	return 0;
802 
803  release:
804 	if (!ndev->npinfo)
805 		kfree(npinfo);
806 	np->dev = NULL;
807 	dev_put(ndev);
808 	return err;
809 }
810 
811 static int __init netpoll_init(void)
812 {
813 	skb_queue_head_init(&skb_pool);
814 	return 0;
815 }
816 core_initcall(netpoll_init);
817 
818 void netpoll_cleanup(struct netpoll *np)
819 {
820 	struct netpoll_info *npinfo;
821 	unsigned long flags;
822 
823 	if (np->dev) {
824 		npinfo = np->dev->npinfo;
825 		if (npinfo) {
826 			if (npinfo->rx_np == np) {
827 				spin_lock_irqsave(&npinfo->rx_lock, flags);
828 				npinfo->rx_np = NULL;
829 				npinfo->rx_flags &= ~NETPOLL_RX_ENABLED;
830 				spin_unlock_irqrestore(&npinfo->rx_lock, flags);
831 			}
832 
833 			if (atomic_dec_and_test(&npinfo->refcnt)) {
834 				skb_queue_purge(&npinfo->arp_tx);
835 				skb_queue_purge(&npinfo->txq);
836 				cancel_rearming_delayed_work(&npinfo->tx_work);
837 
838 				/* clean after last, unfinished work */
839 				__skb_queue_purge(&npinfo->txq);
840 				kfree(npinfo);
841 				np->dev->npinfo = NULL;
842 			}
843 		}
844 
845 		dev_put(np->dev);
846 	}
847 
848 	np->dev = NULL;
849 }
850 
851 int netpoll_trap(void)
852 {
853 	return atomic_read(&trapped);
854 }
855 
856 void netpoll_set_trap(int trap)
857 {
858 	if (trap)
859 		atomic_inc(&trapped);
860 	else
861 		atomic_dec(&trapped);
862 }
863 
864 EXPORT_SYMBOL(netpoll_set_trap);
865 EXPORT_SYMBOL(netpoll_trap);
866 EXPORT_SYMBOL(netpoll_print_options);
867 EXPORT_SYMBOL(netpoll_parse_options);
868 EXPORT_SYMBOL(netpoll_setup);
869 EXPORT_SYMBOL(netpoll_cleanup);
870 EXPORT_SYMBOL(netpoll_send_udp);
871 EXPORT_SYMBOL(netpoll_poll);
872