xref: /openbmc/linux/net/core/netpoll.c (revision b627b4ed)
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/netdevice.h>
13 #include <linux/etherdevice.h>
14 #include <linux/string.h>
15 #include <linux/if_arp.h>
16 #include <linux/inetdevice.h>
17 #include <linux/inet.h>
18 #include <linux/interrupt.h>
19 #include <linux/netpoll.h>
20 #include <linux/sched.h>
21 #include <linux/delay.h>
22 #include <linux/rcupdate.h>
23 #include <linux/workqueue.h>
24 #include <net/tcp.h>
25 #include <net/udp.h>
26 #include <asm/unaligned.h>
27 
28 /*
29  * We maintain a small pool of fully-sized skbs, to make sure the
30  * message gets out even in extreme OOM situations.
31  */
32 
33 #define MAX_UDP_CHUNK 1460
34 #define MAX_SKBS 32
35 #define MAX_QUEUE_DEPTH (MAX_SKBS / 2)
36 
37 static struct sk_buff_head skb_pool;
38 
39 static atomic_t trapped;
40 
41 #define USEC_PER_POLL	50
42 #define NETPOLL_RX_ENABLED  1
43 #define NETPOLL_RX_DROP     2
44 
45 #define MAX_SKB_SIZE \
46 		(MAX_UDP_CHUNK + sizeof(struct udphdr) + \
47 				sizeof(struct iphdr) + sizeof(struct ethhdr))
48 
49 static void zap_completion_queue(void);
50 static void arp_reply(struct sk_buff *skb);
51 
52 static void queue_process(struct work_struct *work)
53 {
54 	struct netpoll_info *npinfo =
55 		container_of(work, struct netpoll_info, tx_work.work);
56 	struct sk_buff *skb;
57 	unsigned long flags;
58 
59 	while ((skb = skb_dequeue(&npinfo->txq))) {
60 		struct net_device *dev = skb->dev;
61 		const struct net_device_ops *ops = dev->netdev_ops;
62 		struct netdev_queue *txq;
63 
64 		if (!netif_device_present(dev) || !netif_running(dev)) {
65 			__kfree_skb(skb);
66 			continue;
67 		}
68 
69 		txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
70 
71 		local_irq_save(flags);
72 		__netif_tx_lock(txq, smp_processor_id());
73 		if (netif_tx_queue_stopped(txq) ||
74 		    netif_tx_queue_frozen(txq) ||
75 		    ops->ndo_start_xmit(skb, dev) != NETDEV_TX_OK) {
76 			skb_queue_head(&npinfo->txq, skb);
77 			__netif_tx_unlock(txq);
78 			local_irq_restore(flags);
79 
80 			schedule_delayed_work(&npinfo->tx_work, HZ/10);
81 			return;
82 		}
83 		__netif_tx_unlock(txq);
84 		local_irq_restore(flags);
85 	}
86 }
87 
88 static __sum16 checksum_udp(struct sk_buff *skb, struct udphdr *uh,
89 			    unsigned short ulen, __be32 saddr, __be32 daddr)
90 {
91 	__wsum psum;
92 
93 	if (uh->check == 0 || skb_csum_unnecessary(skb))
94 		return 0;
95 
96 	psum = csum_tcpudp_nofold(saddr, daddr, ulen, IPPROTO_UDP, 0);
97 
98 	if (skb->ip_summed == CHECKSUM_COMPLETE &&
99 	    !csum_fold(csum_add(psum, skb->csum)))
100 		return 0;
101 
102 	skb->csum = psum;
103 
104 	return __skb_checksum_complete(skb);
105 }
106 
107 /*
108  * Check whether delayed processing was scheduled for our NIC. If so,
109  * we attempt to grab the poll lock and use ->poll() to pump the card.
110  * If this fails, either we've recursed in ->poll() or it's already
111  * running on another CPU.
112  *
113  * Note: we don't mask interrupts with this lock because we're using
114  * trylock here and interrupts are already disabled in the softirq
115  * case. Further, we test the poll_owner to avoid recursion on UP
116  * systems where the lock doesn't exist.
117  *
118  * In cases where there is bi-directional communications, reading only
119  * one message at a time can lead to packets being dropped by the
120  * network adapter, forcing superfluous retries and possibly timeouts.
121  * Thus, we set our budget to greater than 1.
122  */
123 static int poll_one_napi(struct netpoll_info *npinfo,
124 			 struct napi_struct *napi, int budget)
125 {
126 	int work;
127 
128 	/* net_rx_action's ->poll() invocations and our's are
129 	 * synchronized by this test which is only made while
130 	 * holding the napi->poll_lock.
131 	 */
132 	if (!test_bit(NAPI_STATE_SCHED, &napi->state))
133 		return budget;
134 
135 	npinfo->rx_flags |= NETPOLL_RX_DROP;
136 	atomic_inc(&trapped);
137 	set_bit(NAPI_STATE_NPSVC, &napi->state);
138 
139 	work = napi->poll(napi, budget);
140 
141 	clear_bit(NAPI_STATE_NPSVC, &napi->state);
142 	atomic_dec(&trapped);
143 	npinfo->rx_flags &= ~NETPOLL_RX_DROP;
144 
145 	return budget - work;
146 }
147 
148 static void poll_napi(struct net_device *dev)
149 {
150 	struct napi_struct *napi;
151 	int budget = 16;
152 
153 	list_for_each_entry(napi, &dev->napi_list, dev_list) {
154 		if (napi->poll_owner != smp_processor_id() &&
155 		    spin_trylock(&napi->poll_lock)) {
156 			budget = poll_one_napi(dev->npinfo, napi, budget);
157 			spin_unlock(&napi->poll_lock);
158 
159 			if (!budget)
160 				break;
161 		}
162 	}
163 }
164 
165 static void service_arp_queue(struct netpoll_info *npi)
166 {
167 	if (npi) {
168 		struct sk_buff *skb;
169 
170 		while ((skb = skb_dequeue(&npi->arp_tx)))
171 			arp_reply(skb);
172 	}
173 }
174 
175 void netpoll_poll(struct netpoll *np)
176 {
177 	struct net_device *dev = np->dev;
178 	const struct net_device_ops *ops = dev->netdev_ops;
179 
180 	if (!dev || !netif_running(dev) || !ops->ndo_poll_controller)
181 		return;
182 
183 	/* Process pending work on NIC */
184 	ops->ndo_poll_controller(dev);
185 
186 	poll_napi(dev);
187 
188 	service_arp_queue(dev->npinfo);
189 
190 	zap_completion_queue();
191 }
192 
193 static void refill_skbs(void)
194 {
195 	struct sk_buff *skb;
196 	unsigned long flags;
197 
198 	spin_lock_irqsave(&skb_pool.lock, flags);
199 	while (skb_pool.qlen < MAX_SKBS) {
200 		skb = alloc_skb(MAX_SKB_SIZE, GFP_ATOMIC);
201 		if (!skb)
202 			break;
203 
204 		__skb_queue_tail(&skb_pool, skb);
205 	}
206 	spin_unlock_irqrestore(&skb_pool.lock, flags);
207 }
208 
209 static void zap_completion_queue(void)
210 {
211 	unsigned long flags;
212 	struct softnet_data *sd = &get_cpu_var(softnet_data);
213 
214 	if (sd->completion_queue) {
215 		struct sk_buff *clist;
216 
217 		local_irq_save(flags);
218 		clist = sd->completion_queue;
219 		sd->completion_queue = NULL;
220 		local_irq_restore(flags);
221 
222 		while (clist != NULL) {
223 			struct sk_buff *skb = clist;
224 			clist = clist->next;
225 			if (skb->destructor) {
226 				atomic_inc(&skb->users);
227 				dev_kfree_skb_any(skb); /* put this one back */
228 			} else {
229 				__kfree_skb(skb);
230 			}
231 		}
232 	}
233 
234 	put_cpu_var(softnet_data);
235 }
236 
237 static struct sk_buff *find_skb(struct netpoll *np, int len, int reserve)
238 {
239 	int count = 0;
240 	struct sk_buff *skb;
241 
242 	zap_completion_queue();
243 	refill_skbs();
244 repeat:
245 
246 	skb = alloc_skb(len, GFP_ATOMIC);
247 	if (!skb)
248 		skb = skb_dequeue(&skb_pool);
249 
250 	if (!skb) {
251 		if (++count < 10) {
252 			netpoll_poll(np);
253 			goto repeat;
254 		}
255 		return NULL;
256 	}
257 
258 	atomic_set(&skb->users, 1);
259 	skb_reserve(skb, reserve);
260 	return skb;
261 }
262 
263 static int netpoll_owner_active(struct net_device *dev)
264 {
265 	struct napi_struct *napi;
266 
267 	list_for_each_entry(napi, &dev->napi_list, dev_list) {
268 		if (napi->poll_owner == smp_processor_id())
269 			return 1;
270 	}
271 	return 0;
272 }
273 
274 static void netpoll_send_skb(struct netpoll *np, struct sk_buff *skb)
275 {
276 	int status = NETDEV_TX_BUSY;
277 	unsigned long tries;
278 	struct net_device *dev = np->dev;
279 	const struct net_device_ops *ops = dev->netdev_ops;
280 	struct netpoll_info *npinfo = np->dev->npinfo;
281 
282 	if (!npinfo || !netif_running(dev) || !netif_device_present(dev)) {
283 		__kfree_skb(skb);
284 		return;
285 	}
286 
287 	/* don't get messages out of order, and no recursion */
288 	if (skb_queue_len(&npinfo->txq) == 0 && !netpoll_owner_active(dev)) {
289 		struct netdev_queue *txq;
290 		unsigned long flags;
291 
292 		txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
293 
294 		local_irq_save(flags);
295 		/* try until next clock tick */
296 		for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
297 		     tries > 0; --tries) {
298 			if (__netif_tx_trylock(txq)) {
299 				if (!netif_tx_queue_stopped(txq))
300 					status = ops->ndo_start_xmit(skb, dev);
301 				__netif_tx_unlock(txq);
302 
303 				if (status == NETDEV_TX_OK)
304 					break;
305 
306 			}
307 
308 			/* tickle device maybe there is some cleanup */
309 			netpoll_poll(np);
310 
311 			udelay(USEC_PER_POLL);
312 		}
313 		local_irq_restore(flags);
314 	}
315 
316 	if (status != NETDEV_TX_OK) {
317 		skb_queue_tail(&npinfo->txq, skb);
318 		schedule_delayed_work(&npinfo->tx_work,0);
319 	}
320 }
321 
322 void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
323 {
324 	int total_len, eth_len, ip_len, udp_len;
325 	struct sk_buff *skb;
326 	struct udphdr *udph;
327 	struct iphdr *iph;
328 	struct ethhdr *eth;
329 
330 	udp_len = len + sizeof(*udph);
331 	ip_len = eth_len = udp_len + sizeof(*iph);
332 	total_len = eth_len + ETH_HLEN + NET_IP_ALIGN;
333 
334 	skb = find_skb(np, total_len, total_len - len);
335 	if (!skb)
336 		return;
337 
338 	skb_copy_to_linear_data(skb, msg, len);
339 	skb->len += len;
340 
341 	skb_push(skb, sizeof(*udph));
342 	skb_reset_transport_header(skb);
343 	udph = udp_hdr(skb);
344 	udph->source = htons(np->local_port);
345 	udph->dest = htons(np->remote_port);
346 	udph->len = htons(udp_len);
347 	udph->check = 0;
348 	udph->check = csum_tcpudp_magic(np->local_ip,
349 					np->remote_ip,
350 					udp_len, IPPROTO_UDP,
351 					csum_partial(udph, udp_len, 0));
352 	if (udph->check == 0)
353 		udph->check = CSUM_MANGLED_0;
354 
355 	skb_push(skb, sizeof(*iph));
356 	skb_reset_network_header(skb);
357 	iph = ip_hdr(skb);
358 
359 	/* iph->version = 4; iph->ihl = 5; */
360 	put_unaligned(0x45, (unsigned char *)iph);
361 	iph->tos      = 0;
362 	put_unaligned(htons(ip_len), &(iph->tot_len));
363 	iph->id       = 0;
364 	iph->frag_off = 0;
365 	iph->ttl      = 64;
366 	iph->protocol = IPPROTO_UDP;
367 	iph->check    = 0;
368 	put_unaligned(np->local_ip, &(iph->saddr));
369 	put_unaligned(np->remote_ip, &(iph->daddr));
370 	iph->check    = ip_fast_csum((unsigned char *)iph, iph->ihl);
371 
372 	eth = (struct ethhdr *) skb_push(skb, ETH_HLEN);
373 	skb_reset_mac_header(skb);
374 	skb->protocol = eth->h_proto = htons(ETH_P_IP);
375 	memcpy(eth->h_source, np->dev->dev_addr, ETH_ALEN);
376 	memcpy(eth->h_dest, np->remote_mac, ETH_ALEN);
377 
378 	skb->dev = np->dev;
379 
380 	netpoll_send_skb(np, skb);
381 }
382 
383 static void arp_reply(struct sk_buff *skb)
384 {
385 	struct netpoll_info *npinfo = skb->dev->npinfo;
386 	struct arphdr *arp;
387 	unsigned char *arp_ptr;
388 	int size, type = ARPOP_REPLY, ptype = ETH_P_ARP;
389 	__be32 sip, tip;
390 	unsigned char *sha;
391 	struct sk_buff *send_skb;
392 	struct netpoll *np = NULL;
393 
394 	if (npinfo->rx_np && npinfo->rx_np->dev == skb->dev)
395 		np = npinfo->rx_np;
396 	if (!np)
397 		return;
398 
399 	/* No arp on this interface */
400 	if (skb->dev->flags & IFF_NOARP)
401 		return;
402 
403 	if (!pskb_may_pull(skb, arp_hdr_len(skb->dev)))
404 		return;
405 
406 	skb_reset_network_header(skb);
407 	skb_reset_transport_header(skb);
408 	arp = arp_hdr(skb);
409 
410 	if ((arp->ar_hrd != htons(ARPHRD_ETHER) &&
411 	     arp->ar_hrd != htons(ARPHRD_IEEE802)) ||
412 	    arp->ar_pro != htons(ETH_P_IP) ||
413 	    arp->ar_op != htons(ARPOP_REQUEST))
414 		return;
415 
416 	arp_ptr = (unsigned char *)(arp+1);
417 	/* save the location of the src hw addr */
418 	sha = arp_ptr;
419 	arp_ptr += skb->dev->addr_len;
420 	memcpy(&sip, arp_ptr, 4);
421 	arp_ptr += 4;
422 	/* if we actually cared about dst hw addr, it would get copied here */
423 	arp_ptr += skb->dev->addr_len;
424 	memcpy(&tip, arp_ptr, 4);
425 
426 	/* Should we ignore arp? */
427 	if (tip != np->local_ip ||
428 	    ipv4_is_loopback(tip) || ipv4_is_multicast(tip))
429 		return;
430 
431 	size = arp_hdr_len(skb->dev);
432 	send_skb = find_skb(np, size + LL_ALLOCATED_SPACE(np->dev),
433 			    LL_RESERVED_SPACE(np->dev));
434 
435 	if (!send_skb)
436 		return;
437 
438 	skb_reset_network_header(send_skb);
439 	arp = (struct arphdr *) skb_put(send_skb, size);
440 	send_skb->dev = skb->dev;
441 	send_skb->protocol = htons(ETH_P_ARP);
442 
443 	/* Fill the device header for the ARP frame */
444 	if (dev_hard_header(send_skb, skb->dev, ptype,
445 			    sha, np->dev->dev_addr,
446 			    send_skb->len) < 0) {
447 		kfree_skb(send_skb);
448 		return;
449 	}
450 
451 	/*
452 	 * Fill out the arp protocol part.
453 	 *
454 	 * we only support ethernet device type,
455 	 * which (according to RFC 1390) should always equal 1 (Ethernet).
456 	 */
457 
458 	arp->ar_hrd = htons(np->dev->type);
459 	arp->ar_pro = htons(ETH_P_IP);
460 	arp->ar_hln = np->dev->addr_len;
461 	arp->ar_pln = 4;
462 	arp->ar_op = htons(type);
463 
464 	arp_ptr=(unsigned char *)(arp + 1);
465 	memcpy(arp_ptr, np->dev->dev_addr, np->dev->addr_len);
466 	arp_ptr += np->dev->addr_len;
467 	memcpy(arp_ptr, &tip, 4);
468 	arp_ptr += 4;
469 	memcpy(arp_ptr, sha, np->dev->addr_len);
470 	arp_ptr += np->dev->addr_len;
471 	memcpy(arp_ptr, &sip, 4);
472 
473 	netpoll_send_skb(np, send_skb);
474 }
475 
476 int __netpoll_rx(struct sk_buff *skb)
477 {
478 	int proto, len, ulen;
479 	struct iphdr *iph;
480 	struct udphdr *uh;
481 	struct netpoll_info *npi = skb->dev->npinfo;
482 	struct netpoll *np = npi->rx_np;
483 
484 	if (!np)
485 		goto out;
486 	if (skb->dev->type != ARPHRD_ETHER)
487 		goto out;
488 
489 	/* check if netpoll clients need ARP */
490 	if (skb->protocol == htons(ETH_P_ARP) &&
491 	    atomic_read(&trapped)) {
492 		skb_queue_tail(&npi->arp_tx, skb);
493 		return 1;
494 	}
495 
496 	proto = ntohs(eth_hdr(skb)->h_proto);
497 	if (proto != ETH_P_IP)
498 		goto out;
499 	if (skb->pkt_type == PACKET_OTHERHOST)
500 		goto out;
501 	if (skb_shared(skb))
502 		goto out;
503 
504 	iph = (struct iphdr *)skb->data;
505 	if (!pskb_may_pull(skb, sizeof(struct iphdr)))
506 		goto out;
507 	if (iph->ihl < 5 || iph->version != 4)
508 		goto out;
509 	if (!pskb_may_pull(skb, iph->ihl*4))
510 		goto out;
511 	if (ip_fast_csum((u8 *)iph, iph->ihl) != 0)
512 		goto out;
513 
514 	len = ntohs(iph->tot_len);
515 	if (skb->len < len || len < iph->ihl*4)
516 		goto out;
517 
518 	/*
519 	 * Our transport medium may have padded the buffer out.
520 	 * Now We trim to the true length of the frame.
521 	 */
522 	if (pskb_trim_rcsum(skb, len))
523 		goto out;
524 
525 	if (iph->protocol != IPPROTO_UDP)
526 		goto out;
527 
528 	len -= iph->ihl*4;
529 	uh = (struct udphdr *)(((char *)iph) + iph->ihl*4);
530 	ulen = ntohs(uh->len);
531 
532 	if (ulen != len)
533 		goto out;
534 	if (checksum_udp(skb, uh, ulen, iph->saddr, iph->daddr))
535 		goto out;
536 	if (np->local_ip && np->local_ip != iph->daddr)
537 		goto out;
538 	if (np->remote_ip && np->remote_ip != iph->saddr)
539 		goto out;
540 	if (np->local_port && np->local_port != ntohs(uh->dest))
541 		goto out;
542 
543 	np->rx_hook(np, ntohs(uh->source),
544 		    (char *)(uh+1),
545 		    ulen - sizeof(struct udphdr));
546 
547 	kfree_skb(skb);
548 	return 1;
549 
550 out:
551 	if (atomic_read(&trapped)) {
552 		kfree_skb(skb);
553 		return 1;
554 	}
555 
556 	return 0;
557 }
558 
559 void netpoll_print_options(struct netpoll *np)
560 {
561 	printk(KERN_INFO "%s: local port %d\n",
562 			 np->name, np->local_port);
563 	printk(KERN_INFO "%s: local IP %pI4\n",
564 			 np->name, &np->local_ip);
565 	printk(KERN_INFO "%s: interface %s\n",
566 			 np->name, np->dev_name);
567 	printk(KERN_INFO "%s: remote port %d\n",
568 			 np->name, np->remote_port);
569 	printk(KERN_INFO "%s: remote IP %pI4\n",
570 			 np->name, &np->remote_ip);
571 	printk(KERN_INFO "%s: remote ethernet address %pM\n",
572 	                 np->name, np->remote_mac);
573 }
574 
575 int netpoll_parse_options(struct netpoll *np, char *opt)
576 {
577 	char *cur=opt, *delim;
578 
579 	if (*cur != '@') {
580 		if ((delim = strchr(cur, '@')) == NULL)
581 			goto parse_failed;
582 		*delim = 0;
583 		np->local_port = simple_strtol(cur, NULL, 10);
584 		cur = delim;
585 	}
586 	cur++;
587 
588 	if (*cur != '/') {
589 		if ((delim = strchr(cur, '/')) == NULL)
590 			goto parse_failed;
591 		*delim = 0;
592 		np->local_ip = in_aton(cur);
593 		cur = delim;
594 	}
595 	cur++;
596 
597 	if (*cur != ',') {
598 		/* parse out dev name */
599 		if ((delim = strchr(cur, ',')) == NULL)
600 			goto parse_failed;
601 		*delim = 0;
602 		strlcpy(np->dev_name, cur, sizeof(np->dev_name));
603 		cur = delim;
604 	}
605 	cur++;
606 
607 	if (*cur != '@') {
608 		/* dst port */
609 		if ((delim = strchr(cur, '@')) == NULL)
610 			goto parse_failed;
611 		*delim = 0;
612 		np->remote_port = simple_strtol(cur, NULL, 10);
613 		cur = delim;
614 	}
615 	cur++;
616 
617 	/* dst ip */
618 	if ((delim = strchr(cur, '/')) == NULL)
619 		goto parse_failed;
620 	*delim = 0;
621 	np->remote_ip = in_aton(cur);
622 	cur = delim + 1;
623 
624 	if (*cur != 0) {
625 		/* MAC address */
626 		if ((delim = strchr(cur, ':')) == NULL)
627 			goto parse_failed;
628 		*delim = 0;
629 		np->remote_mac[0] = simple_strtol(cur, NULL, 16);
630 		cur = delim + 1;
631 		if ((delim = strchr(cur, ':')) == NULL)
632 			goto parse_failed;
633 		*delim = 0;
634 		np->remote_mac[1] = simple_strtol(cur, NULL, 16);
635 		cur = delim + 1;
636 		if ((delim = strchr(cur, ':')) == NULL)
637 			goto parse_failed;
638 		*delim = 0;
639 		np->remote_mac[2] = simple_strtol(cur, NULL, 16);
640 		cur = delim + 1;
641 		if ((delim = strchr(cur, ':')) == NULL)
642 			goto parse_failed;
643 		*delim = 0;
644 		np->remote_mac[3] = simple_strtol(cur, NULL, 16);
645 		cur = delim + 1;
646 		if ((delim = strchr(cur, ':')) == NULL)
647 			goto parse_failed;
648 		*delim = 0;
649 		np->remote_mac[4] = simple_strtol(cur, NULL, 16);
650 		cur = delim + 1;
651 		np->remote_mac[5] = simple_strtol(cur, NULL, 16);
652 	}
653 
654 	netpoll_print_options(np);
655 
656 	return 0;
657 
658  parse_failed:
659 	printk(KERN_INFO "%s: couldn't parse config at %s!\n",
660 	       np->name, cur);
661 	return -1;
662 }
663 
664 int netpoll_setup(struct netpoll *np)
665 {
666 	struct net_device *ndev = NULL;
667 	struct in_device *in_dev;
668 	struct netpoll_info *npinfo;
669 	unsigned long flags;
670 	int err;
671 
672 	if (np->dev_name)
673 		ndev = dev_get_by_name(&init_net, np->dev_name);
674 	if (!ndev) {
675 		printk(KERN_ERR "%s: %s doesn't exist, aborting.\n",
676 		       np->name, np->dev_name);
677 		return -ENODEV;
678 	}
679 
680 	np->dev = ndev;
681 	if (!ndev->npinfo) {
682 		npinfo = kmalloc(sizeof(*npinfo), GFP_KERNEL);
683 		if (!npinfo) {
684 			err = -ENOMEM;
685 			goto release;
686 		}
687 
688 		npinfo->rx_flags = 0;
689 		npinfo->rx_np = NULL;
690 
691 		spin_lock_init(&npinfo->rx_lock);
692 		skb_queue_head_init(&npinfo->arp_tx);
693 		skb_queue_head_init(&npinfo->txq);
694 		INIT_DELAYED_WORK(&npinfo->tx_work, queue_process);
695 
696 		atomic_set(&npinfo->refcnt, 1);
697 	} else {
698 		npinfo = ndev->npinfo;
699 		atomic_inc(&npinfo->refcnt);
700 	}
701 
702 	if (!ndev->netdev_ops->ndo_poll_controller) {
703 		printk(KERN_ERR "%s: %s doesn't support polling, aborting.\n",
704 		       np->name, np->dev_name);
705 		err = -ENOTSUPP;
706 		goto release;
707 	}
708 
709 	if (!netif_running(ndev)) {
710 		unsigned long atmost, atleast;
711 
712 		printk(KERN_INFO "%s: device %s not up yet, forcing it\n",
713 		       np->name, np->dev_name);
714 
715 		rtnl_lock();
716 		err = dev_open(ndev);
717 		rtnl_unlock();
718 
719 		if (err) {
720 			printk(KERN_ERR "%s: failed to open %s\n",
721 			       np->name, ndev->name);
722 			goto release;
723 		}
724 
725 		atleast = jiffies + HZ/10;
726 		atmost = jiffies + 4*HZ;
727 		while (!netif_carrier_ok(ndev)) {
728 			if (time_after(jiffies, atmost)) {
729 				printk(KERN_NOTICE
730 				       "%s: timeout waiting for carrier\n",
731 				       np->name);
732 				break;
733 			}
734 			cond_resched();
735 		}
736 
737 		/* If carrier appears to come up instantly, we don't
738 		 * trust it and pause so that we don't pump all our
739 		 * queued console messages into the bitbucket.
740 		 */
741 
742 		if (time_before(jiffies, atleast)) {
743 			printk(KERN_NOTICE "%s: carrier detect appears"
744 			       " untrustworthy, waiting 4 seconds\n",
745 			       np->name);
746 			msleep(4000);
747 		}
748 	}
749 
750 	if (!np->local_ip) {
751 		rcu_read_lock();
752 		in_dev = __in_dev_get_rcu(ndev);
753 
754 		if (!in_dev || !in_dev->ifa_list) {
755 			rcu_read_unlock();
756 			printk(KERN_ERR "%s: no IP address for %s, aborting\n",
757 			       np->name, np->dev_name);
758 			err = -EDESTADDRREQ;
759 			goto release;
760 		}
761 
762 		np->local_ip = in_dev->ifa_list->ifa_local;
763 		rcu_read_unlock();
764 		printk(KERN_INFO "%s: local IP %pI4\n", np->name, &np->local_ip);
765 	}
766 
767 	if (np->rx_hook) {
768 		spin_lock_irqsave(&npinfo->rx_lock, flags);
769 		npinfo->rx_flags |= NETPOLL_RX_ENABLED;
770 		npinfo->rx_np = np;
771 		spin_unlock_irqrestore(&npinfo->rx_lock, flags);
772 	}
773 
774 	/* fill up the skb queue */
775 	refill_skbs();
776 
777 	/* last thing to do is link it to the net device structure */
778 	ndev->npinfo = npinfo;
779 
780 	/* avoid racing with NAPI reading npinfo */
781 	synchronize_rcu();
782 
783 	return 0;
784 
785  release:
786 	if (!ndev->npinfo)
787 		kfree(npinfo);
788 	np->dev = NULL;
789 	dev_put(ndev);
790 	return err;
791 }
792 
793 static int __init netpoll_init(void)
794 {
795 	skb_queue_head_init(&skb_pool);
796 	return 0;
797 }
798 core_initcall(netpoll_init);
799 
800 void netpoll_cleanup(struct netpoll *np)
801 {
802 	struct netpoll_info *npinfo;
803 	unsigned long flags;
804 
805 	if (np->dev) {
806 		npinfo = np->dev->npinfo;
807 		if (npinfo) {
808 			if (npinfo->rx_np == np) {
809 				spin_lock_irqsave(&npinfo->rx_lock, flags);
810 				npinfo->rx_np = NULL;
811 				npinfo->rx_flags &= ~NETPOLL_RX_ENABLED;
812 				spin_unlock_irqrestore(&npinfo->rx_lock, flags);
813 			}
814 
815 			if (atomic_dec_and_test(&npinfo->refcnt)) {
816 				skb_queue_purge(&npinfo->arp_tx);
817 				skb_queue_purge(&npinfo->txq);
818 				cancel_rearming_delayed_work(&npinfo->tx_work);
819 
820 				/* clean after last, unfinished work */
821 				__skb_queue_purge(&npinfo->txq);
822 				kfree(npinfo);
823 				np->dev->npinfo = NULL;
824 			}
825 		}
826 
827 		dev_put(np->dev);
828 	}
829 
830 	np->dev = NULL;
831 }
832 
833 int netpoll_trap(void)
834 {
835 	return atomic_read(&trapped);
836 }
837 
838 void netpoll_set_trap(int trap)
839 {
840 	if (trap)
841 		atomic_inc(&trapped);
842 	else
843 		atomic_dec(&trapped);
844 }
845 
846 EXPORT_SYMBOL(netpoll_set_trap);
847 EXPORT_SYMBOL(netpoll_trap);
848 EXPORT_SYMBOL(netpoll_print_options);
849 EXPORT_SYMBOL(netpoll_parse_options);
850 EXPORT_SYMBOL(netpoll_setup);
851 EXPORT_SYMBOL(netpoll_cleanup);
852 EXPORT_SYMBOL(netpoll_send_udp);
853 EXPORT_SYMBOL(netpoll_poll);
854