netpoll: fix race on poll_list resulting in garbage entry
[safe/jmp/linux-2.6] / net / core / netpoll.c
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                 struct netdev_queue *txq;
62
63                 if (!netif_device_present(dev) || !netif_running(dev)) {
64                         __kfree_skb(skb);
65                         continue;
66                 }
67
68                 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
69
70                 local_irq_save(flags);
71                 __netif_tx_lock(txq, smp_processor_id());
72                 if (netif_tx_queue_stopped(txq) ||
73                     netif_tx_queue_frozen(txq) ||
74                     dev->hard_start_xmit(skb, dev) != NETDEV_TX_OK) {
75                         skb_queue_head(&npinfo->txq, skb);
76                         __netif_tx_unlock(txq);
77                         local_irq_restore(flags);
78
79                         schedule_delayed_work(&npinfo->tx_work, HZ/10);
80                         return;
81                 }
82                 __netif_tx_unlock(txq);
83                 local_irq_restore(flags);
84         }
85 }
86
87 static __sum16 checksum_udp(struct sk_buff *skb, struct udphdr *uh,
88                             unsigned short ulen, __be32 saddr, __be32 daddr)
89 {
90         __wsum psum;
91
92         if (uh->check == 0 || skb_csum_unnecessary(skb))
93                 return 0;
94
95         psum = csum_tcpudp_nofold(saddr, daddr, ulen, IPPROTO_UDP, 0);
96
97         if (skb->ip_summed == CHECKSUM_COMPLETE &&
98             !csum_fold(csum_add(psum, skb->csum)))
99                 return 0;
100
101         skb->csum = psum;
102
103         return __skb_checksum_complete(skb);
104 }
105
106 /*
107  * Check whether delayed processing was scheduled for our NIC. If so,
108  * we attempt to grab the poll lock and use ->poll() to pump the card.
109  * If this fails, either we've recursed in ->poll() or it's already
110  * running on another CPU.
111  *
112  * Note: we don't mask interrupts with this lock because we're using
113  * trylock here and interrupts are already disabled in the softirq
114  * case. Further, we test the poll_owner to avoid recursion on UP
115  * systems where the lock doesn't exist.
116  *
117  * In cases where there is bi-directional communications, reading only
118  * one message at a time can lead to packets being dropped by the
119  * network adapter, forcing superfluous retries and possibly timeouts.
120  * Thus, we set our budget to greater than 1.
121  */
122 static int poll_one_napi(struct netpoll_info *npinfo,
123                          struct napi_struct *napi, int budget)
124 {
125         int work;
126
127         /* net_rx_action's ->poll() invocations and our's are
128          * synchronized by this test which is only made while
129          * holding the napi->poll_lock.
130          */
131         if (!test_bit(NAPI_STATE_SCHED, &napi->state))
132                 return budget;
133
134         npinfo->rx_flags |= NETPOLL_RX_DROP;
135         atomic_inc(&trapped);
136         set_bit(NAPI_STATE_NPSVC, &napi->state);
137
138         work = napi->poll(napi, budget);
139
140         clear_bit(NAPI_STATE_NPSVC, &napi->state);
141         atomic_dec(&trapped);
142         npinfo->rx_flags &= ~NETPOLL_RX_DROP;
143
144         return budget - work;
145 }
146
147 static void poll_napi(struct net_device *dev)
148 {
149         struct napi_struct *napi;
150         int budget = 16;
151
152         list_for_each_entry(napi, &dev->napi_list, dev_list) {
153                 if (napi->poll_owner != smp_processor_id() &&
154                     spin_trylock(&napi->poll_lock)) {
155                         budget = poll_one_napi(dev->npinfo, napi, budget);
156                         spin_unlock(&napi->poll_lock);
157
158                         if (!budget)
159                                 break;
160                 }
161         }
162 }
163
164 static void service_arp_queue(struct netpoll_info *npi)
165 {
166         if (npi) {
167                 struct sk_buff *skb;
168
169                 while ((skb = skb_dequeue(&npi->arp_tx)))
170                         arp_reply(skb);
171         }
172 }
173
174 void netpoll_poll(struct netpoll *np)
175 {
176         struct net_device *dev = np->dev;
177
178         if (!dev || !netif_running(dev) || !dev->poll_controller)
179                 return;
180
181         /* Process pending work on NIC */
182         dev->poll_controller(dev);
183
184         poll_napi(dev);
185
186         service_arp_queue(dev->npinfo);
187
188         zap_completion_queue();
189 }
190
191 static void refill_skbs(void)
192 {
193         struct sk_buff *skb;
194         unsigned long flags;
195
196         spin_lock_irqsave(&skb_pool.lock, flags);
197         while (skb_pool.qlen < MAX_SKBS) {
198                 skb = alloc_skb(MAX_SKB_SIZE, GFP_ATOMIC);
199                 if (!skb)
200                         break;
201
202                 __skb_queue_tail(&skb_pool, skb);
203         }
204         spin_unlock_irqrestore(&skb_pool.lock, flags);
205 }
206
207 static void zap_completion_queue(void)
208 {
209         unsigned long flags;
210         struct softnet_data *sd = &get_cpu_var(softnet_data);
211
212         if (sd->completion_queue) {
213                 struct sk_buff *clist;
214
215                 local_irq_save(flags);
216                 clist = sd->completion_queue;
217                 sd->completion_queue = NULL;
218                 local_irq_restore(flags);
219
220                 while (clist != NULL) {
221                         struct sk_buff *skb = clist;
222                         clist = clist->next;
223                         if (skb->destructor) {
224                                 atomic_inc(&skb->users);
225                                 dev_kfree_skb_any(skb); /* put this one back */
226                         } else {
227                                 __kfree_skb(skb);
228                         }
229                 }
230         }
231
232         put_cpu_var(softnet_data);
233 }
234
235 static struct sk_buff *find_skb(struct netpoll *np, int len, int reserve)
236 {
237         int count = 0;
238         struct sk_buff *skb;
239
240         zap_completion_queue();
241         refill_skbs();
242 repeat:
243
244         skb = alloc_skb(len, GFP_ATOMIC);
245         if (!skb)
246                 skb = skb_dequeue(&skb_pool);
247
248         if (!skb) {
249                 if (++count < 10) {
250                         netpoll_poll(np);
251                         goto repeat;
252                 }
253                 return NULL;
254         }
255
256         atomic_set(&skb->users, 1);
257         skb_reserve(skb, reserve);
258         return skb;
259 }
260
261 static int netpoll_owner_active(struct net_device *dev)
262 {
263         struct napi_struct *napi;
264
265         list_for_each_entry(napi, &dev->napi_list, dev_list) {
266                 if (napi->poll_owner == smp_processor_id())
267                         return 1;
268         }
269         return 0;
270 }
271
272 static void netpoll_send_skb(struct netpoll *np, struct sk_buff *skb)
273 {
274         int status = NETDEV_TX_BUSY;
275         unsigned long tries;
276         struct net_device *dev = np->dev;
277         struct netpoll_info *npinfo = np->dev->npinfo;
278
279         if (!npinfo || !netif_running(dev) || !netif_device_present(dev)) {
280                 __kfree_skb(skb);
281                 return;
282         }
283
284         /* don't get messages out of order, and no recursion */
285         if (skb_queue_len(&npinfo->txq) == 0 && !netpoll_owner_active(dev)) {
286                 struct netdev_queue *txq;
287                 unsigned long flags;
288
289                 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
290
291                 local_irq_save(flags);
292                 /* try until next clock tick */
293                 for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
294                      tries > 0; --tries) {
295                         if (__netif_tx_trylock(txq)) {
296                                 if (!netif_tx_queue_stopped(txq))
297                                         status = dev->hard_start_xmit(skb, dev);
298                                 __netif_tx_unlock(txq);
299
300                                 if (status == NETDEV_TX_OK)
301                                         break;
302
303                         }
304
305                         /* tickle device maybe there is some cleanup */
306                         netpoll_poll(np);
307
308                         udelay(USEC_PER_POLL);
309                 }
310                 local_irq_restore(flags);
311         }
312
313         if (status != NETDEV_TX_OK) {
314                 skb_queue_tail(&npinfo->txq, skb);
315                 schedule_delayed_work(&npinfo->tx_work,0);
316         }
317 }
318
319 void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
320 {
321         int total_len, eth_len, ip_len, udp_len;
322         struct sk_buff *skb;
323         struct udphdr *udph;
324         struct iphdr *iph;
325         struct ethhdr *eth;
326
327         udp_len = len + sizeof(*udph);
328         ip_len = eth_len = udp_len + sizeof(*iph);
329         total_len = eth_len + ETH_HLEN + NET_IP_ALIGN;
330
331         skb = find_skb(np, total_len, total_len - len);
332         if (!skb)
333                 return;
334
335         skb_copy_to_linear_data(skb, msg, len);
336         skb->len += len;
337
338         skb_push(skb, sizeof(*udph));
339         skb_reset_transport_header(skb);
340         udph = udp_hdr(skb);
341         udph->source = htons(np->local_port);
342         udph->dest = htons(np->remote_port);
343         udph->len = htons(udp_len);
344         udph->check = 0;
345         udph->check = csum_tcpudp_magic(htonl(np->local_ip),
346                                         htonl(np->remote_ip),
347                                         udp_len, IPPROTO_UDP,
348                                         csum_partial((unsigned char *)udph, udp_len, 0));
349         if (udph->check == 0)
350                 udph->check = CSUM_MANGLED_0;
351
352         skb_push(skb, sizeof(*iph));
353         skb_reset_network_header(skb);
354         iph = ip_hdr(skb);
355
356         /* iph->version = 4; iph->ihl = 5; */
357         put_unaligned(0x45, (unsigned char *)iph);
358         iph->tos      = 0;
359         put_unaligned(htons(ip_len), &(iph->tot_len));
360         iph->id       = 0;
361         iph->frag_off = 0;
362         iph->ttl      = 64;
363         iph->protocol = IPPROTO_UDP;
364         iph->check    = 0;
365         put_unaligned(htonl(np->local_ip), &(iph->saddr));
366         put_unaligned(htonl(np->remote_ip), &(iph->daddr));
367         iph->check    = ip_fast_csum((unsigned char *)iph, iph->ihl);
368
369         eth = (struct ethhdr *) skb_push(skb, ETH_HLEN);
370         skb_reset_mac_header(skb);
371         skb->protocol = eth->h_proto = htons(ETH_P_IP);
372         memcpy(eth->h_source, np->dev->dev_addr, ETH_ALEN);
373         memcpy(eth->h_dest, np->remote_mac, ETH_ALEN);
374
375         skb->dev = np->dev;
376
377         netpoll_send_skb(np, skb);
378 }
379
380 static void arp_reply(struct sk_buff *skb)
381 {
382         struct netpoll_info *npinfo = skb->dev->npinfo;
383         struct arphdr *arp;
384         unsigned char *arp_ptr;
385         int size, type = ARPOP_REPLY, ptype = ETH_P_ARP;
386         __be32 sip, tip;
387         unsigned char *sha;
388         struct sk_buff *send_skb;
389         struct netpoll *np = NULL;
390
391         if (npinfo->rx_np && npinfo->rx_np->dev == skb->dev)
392                 np = npinfo->rx_np;
393         if (!np)
394                 return;
395
396         /* No arp on this interface */
397         if (skb->dev->flags & IFF_NOARP)
398                 return;
399
400         if (!pskb_may_pull(skb, arp_hdr_len(skb->dev)))
401                 return;
402
403         skb_reset_network_header(skb);
404         skb_reset_transport_header(skb);
405         arp = arp_hdr(skb);
406
407         if ((arp->ar_hrd != htons(ARPHRD_ETHER) &&
408              arp->ar_hrd != htons(ARPHRD_IEEE802)) ||
409             arp->ar_pro != htons(ETH_P_IP) ||
410             arp->ar_op != htons(ARPOP_REQUEST))
411                 return;
412
413         arp_ptr = (unsigned char *)(arp+1);
414         /* save the location of the src hw addr */
415         sha = arp_ptr;
416         arp_ptr += skb->dev->addr_len;
417         memcpy(&sip, arp_ptr, 4);
418         arp_ptr += 4;
419         /* if we actually cared about dst hw addr, it would get copied here */
420         arp_ptr += skb->dev->addr_len;
421         memcpy(&tip, arp_ptr, 4);
422
423         /* Should we ignore arp? */
424         if (tip != htonl(np->local_ip) ||
425             ipv4_is_loopback(tip) || ipv4_is_multicast(tip))
426                 return;
427
428         size = arp_hdr_len(skb->dev);
429         send_skb = find_skb(np, size + LL_ALLOCATED_SPACE(np->dev),
430                             LL_RESERVED_SPACE(np->dev));
431
432         if (!send_skb)
433                 return;
434
435         skb_reset_network_header(send_skb);
436         arp = (struct arphdr *) skb_put(send_skb, size);
437         send_skb->dev = skb->dev;
438         send_skb->protocol = htons(ETH_P_ARP);
439
440         /* Fill the device header for the ARP frame */
441         if (dev_hard_header(send_skb, skb->dev, ptype,
442                             sha, np->dev->dev_addr,
443                             send_skb->len) < 0) {
444                 kfree_skb(send_skb);
445                 return;
446         }
447
448         /*
449          * Fill out the arp protocol part.
450          *
451          * we only support ethernet device type,
452          * which (according to RFC 1390) should always equal 1 (Ethernet).
453          */
454
455         arp->ar_hrd = htons(np->dev->type);
456         arp->ar_pro = htons(ETH_P_IP);
457         arp->ar_hln = np->dev->addr_len;
458         arp->ar_pln = 4;
459         arp->ar_op = htons(type);
460
461         arp_ptr=(unsigned char *)(arp + 1);
462         memcpy(arp_ptr, np->dev->dev_addr, np->dev->addr_len);
463         arp_ptr += np->dev->addr_len;
464         memcpy(arp_ptr, &tip, 4);
465         arp_ptr += 4;
466         memcpy(arp_ptr, sha, np->dev->addr_len);
467         arp_ptr += np->dev->addr_len;
468         memcpy(arp_ptr, &sip, 4);
469
470         netpoll_send_skb(np, send_skb);
471 }
472
473 int __netpoll_rx(struct sk_buff *skb)
474 {
475         int proto, len, ulen;
476         struct iphdr *iph;
477         struct udphdr *uh;
478         struct netpoll_info *npi = skb->dev->npinfo;
479         struct netpoll *np = npi->rx_np;
480
481         if (!np)
482                 goto out;
483         if (skb->dev->type != ARPHRD_ETHER)
484                 goto out;
485
486         /* check if netpoll clients need ARP */
487         if (skb->protocol == htons(ETH_P_ARP) &&
488             atomic_read(&trapped)) {
489                 skb_queue_tail(&npi->arp_tx, skb);
490                 return 1;
491         }
492
493         proto = ntohs(eth_hdr(skb)->h_proto);
494         if (proto != ETH_P_IP)
495                 goto out;
496         if (skb->pkt_type == PACKET_OTHERHOST)
497                 goto out;
498         if (skb_shared(skb))
499                 goto out;
500
501         iph = (struct iphdr *)skb->data;
502         if (!pskb_may_pull(skb, sizeof(struct iphdr)))
503                 goto out;
504         if (iph->ihl < 5 || iph->version != 4)
505                 goto out;
506         if (!pskb_may_pull(skb, iph->ihl*4))
507                 goto out;
508         if (ip_fast_csum((u8 *)iph, iph->ihl) != 0)
509                 goto out;
510
511         len = ntohs(iph->tot_len);
512         if (skb->len < len || len < iph->ihl*4)
513                 goto out;
514
515         /*
516          * Our transport medium may have padded the buffer out.
517          * Now We trim to the true length of the frame.
518          */
519         if (pskb_trim_rcsum(skb, len))
520                 goto out;
521
522         if (iph->protocol != IPPROTO_UDP)
523                 goto out;
524
525         len -= iph->ihl*4;
526         uh = (struct udphdr *)(((char *)iph) + iph->ihl*4);
527         ulen = ntohs(uh->len);
528
529         if (ulen != len)
530                 goto out;
531         if (checksum_udp(skb, uh, ulen, iph->saddr, iph->daddr))
532                 goto out;
533         if (np->local_ip && np->local_ip != ntohl(iph->daddr))
534                 goto out;
535         if (np->remote_ip && np->remote_ip != ntohl(iph->saddr))
536                 goto out;
537         if (np->local_port && np->local_port != ntohs(uh->dest))
538                 goto out;
539
540         np->rx_hook(np, ntohs(uh->source),
541                     (char *)(uh+1),
542                     ulen - sizeof(struct udphdr));
543
544         kfree_skb(skb);
545         return 1;
546
547 out:
548         if (atomic_read(&trapped)) {
549                 kfree_skb(skb);
550                 return 1;
551         }
552
553         return 0;
554 }
555
556 void netpoll_print_options(struct netpoll *np)
557 {
558         DECLARE_MAC_BUF(mac);
559         printk(KERN_INFO "%s: local port %d\n",
560                          np->name, np->local_port);
561         printk(KERN_INFO "%s: local IP %d.%d.%d.%d\n",
562                          np->name, HIPQUAD(np->local_ip));
563         printk(KERN_INFO "%s: interface %s\n",
564                          np->name, np->dev_name);
565         printk(KERN_INFO "%s: remote port %d\n",
566                          np->name, np->remote_port);
567         printk(KERN_INFO "%s: remote IP %d.%d.%d.%d\n",
568                          np->name, HIPQUAD(np->remote_ip));
569         printk(KERN_INFO "%s: remote ethernet address %s\n",
570                          np->name, print_mac(mac, np->remote_mac));
571 }
572
573 int netpoll_parse_options(struct netpoll *np, char *opt)
574 {
575         char *cur=opt, *delim;
576
577         if (*cur != '@') {
578                 if ((delim = strchr(cur, '@')) == NULL)
579                         goto parse_failed;
580                 *delim = 0;
581                 np->local_port = simple_strtol(cur, NULL, 10);
582                 cur = delim;
583         }
584         cur++;
585
586         if (*cur != '/') {
587                 if ((delim = strchr(cur, '/')) == NULL)
588                         goto parse_failed;
589                 *delim = 0;
590                 np->local_ip = ntohl(in_aton(cur));
591                 cur = delim;
592         }
593         cur++;
594
595         if (*cur != ',') {
596                 /* parse out dev name */
597                 if ((delim = strchr(cur, ',')) == NULL)
598                         goto parse_failed;
599                 *delim = 0;
600                 strlcpy(np->dev_name, cur, sizeof(np->dev_name));
601                 cur = delim;
602         }
603         cur++;
604
605         if (*cur != '@') {
606                 /* dst port */
607                 if ((delim = strchr(cur, '@')) == NULL)
608                         goto parse_failed;
609                 *delim = 0;
610                 np->remote_port = simple_strtol(cur, NULL, 10);
611                 cur = delim;
612         }
613         cur++;
614
615         /* dst ip */
616         if ((delim = strchr(cur, '/')) == NULL)
617                 goto parse_failed;
618         *delim = 0;
619         np->remote_ip = ntohl(in_aton(cur));
620         cur = delim + 1;
621
622         if (*cur != 0) {
623                 /* MAC address */
624                 if ((delim = strchr(cur, ':')) == NULL)
625                         goto parse_failed;
626                 *delim = 0;
627                 np->remote_mac[0] = simple_strtol(cur, NULL, 16);
628                 cur = delim + 1;
629                 if ((delim = strchr(cur, ':')) == NULL)
630                         goto parse_failed;
631                 *delim = 0;
632                 np->remote_mac[1] = simple_strtol(cur, NULL, 16);
633                 cur = delim + 1;
634                 if ((delim = strchr(cur, ':')) == NULL)
635                         goto parse_failed;
636                 *delim = 0;
637                 np->remote_mac[2] = simple_strtol(cur, NULL, 16);
638                 cur = delim + 1;
639                 if ((delim = strchr(cur, ':')) == NULL)
640                         goto parse_failed;
641                 *delim = 0;
642                 np->remote_mac[3] = simple_strtol(cur, NULL, 16);
643                 cur = delim + 1;
644                 if ((delim = strchr(cur, ':')) == NULL)
645                         goto parse_failed;
646                 *delim = 0;
647                 np->remote_mac[4] = simple_strtol(cur, NULL, 16);
648                 cur = delim + 1;
649                 np->remote_mac[5] = simple_strtol(cur, NULL, 16);
650         }
651
652         netpoll_print_options(np);
653
654         return 0;
655
656  parse_failed:
657         printk(KERN_INFO "%s: couldn't parse config at %s!\n",
658                np->name, cur);
659         return -1;
660 }
661
662 int netpoll_setup(struct netpoll *np)
663 {
664         struct net_device *ndev = NULL;
665         struct in_device *in_dev;
666         struct netpoll_info *npinfo;
667         unsigned long flags;
668         int err;
669
670         if (np->dev_name)
671                 ndev = dev_get_by_name(&init_net, np->dev_name);
672         if (!ndev) {
673                 printk(KERN_ERR "%s: %s doesn't exist, aborting.\n",
674                        np->name, np->dev_name);
675                 return -ENODEV;
676         }
677
678         np->dev = ndev;
679         if (!ndev->npinfo) {
680                 npinfo = kmalloc(sizeof(*npinfo), GFP_KERNEL);
681                 if (!npinfo) {
682                         err = -ENOMEM;
683                         goto release;
684                 }
685
686                 npinfo->rx_flags = 0;
687                 npinfo->rx_np = NULL;
688
689                 spin_lock_init(&npinfo->rx_lock);
690                 skb_queue_head_init(&npinfo->arp_tx);
691                 skb_queue_head_init(&npinfo->txq);
692                 INIT_DELAYED_WORK(&npinfo->tx_work, queue_process);
693
694                 atomic_set(&npinfo->refcnt, 1);
695         } else {
696                 npinfo = ndev->npinfo;
697                 atomic_inc(&npinfo->refcnt);
698         }
699
700         if (!ndev->poll_controller) {
701                 printk(KERN_ERR "%s: %s doesn't support polling, aborting.\n",
702                        np->name, np->dev_name);
703                 err = -ENOTSUPP;
704                 goto release;
705         }
706
707         if (!netif_running(ndev)) {
708                 unsigned long atmost, atleast;
709
710                 printk(KERN_INFO "%s: device %s not up yet, forcing it\n",
711                        np->name, np->dev_name);
712
713                 rtnl_lock();
714                 err = dev_open(ndev);
715                 rtnl_unlock();
716
717                 if (err) {
718                         printk(KERN_ERR "%s: failed to open %s\n",
719                                np->name, ndev->name);
720                         goto release;
721                 }
722
723                 atleast = jiffies + HZ/10;
724                 atmost = jiffies + 4*HZ;
725                 while (!netif_carrier_ok(ndev)) {
726                         if (time_after(jiffies, atmost)) {
727                                 printk(KERN_NOTICE
728                                        "%s: timeout waiting for carrier\n",
729                                        np->name);
730                                 break;
731                         }
732                         cond_resched();
733                 }
734
735                 /* If carrier appears to come up instantly, we don't
736                  * trust it and pause so that we don't pump all our
737                  * queued console messages into the bitbucket.
738                  */
739
740                 if (time_before(jiffies, atleast)) {
741                         printk(KERN_NOTICE "%s: carrier detect appears"
742                                " untrustworthy, waiting 4 seconds\n",
743                                np->name);
744                         msleep(4000);
745                 }
746         }
747
748         if (!np->local_ip) {
749                 rcu_read_lock();
750                 in_dev = __in_dev_get_rcu(ndev);
751
752                 if (!in_dev || !in_dev->ifa_list) {
753                         rcu_read_unlock();
754                         printk(KERN_ERR "%s: no IP address for %s, aborting\n",
755                                np->name, np->dev_name);
756                         err = -EDESTADDRREQ;
757                         goto release;
758                 }
759
760                 np->local_ip = ntohl(in_dev->ifa_list->ifa_local);
761                 rcu_read_unlock();
762                 printk(KERN_INFO "%s: local IP %d.%d.%d.%d\n",
763                        np->name, HIPQUAD(np->local_ip));
764         }
765
766         if (np->rx_hook) {
767                 spin_lock_irqsave(&npinfo->rx_lock, flags);
768                 npinfo->rx_flags |= NETPOLL_RX_ENABLED;
769                 npinfo->rx_np = np;
770                 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
771         }
772
773         /* fill up the skb queue */
774         refill_skbs();
775
776         /* last thing to do is link it to the net device structure */
777         ndev->npinfo = npinfo;
778
779         /* avoid racing with NAPI reading npinfo */
780         synchronize_rcu();
781
782         return 0;
783
784  release:
785         if (!ndev->npinfo)
786                 kfree(npinfo);
787         np->dev = NULL;
788         dev_put(ndev);
789         return err;
790 }
791
792 static int __init netpoll_init(void)
793 {
794         skb_queue_head_init(&skb_pool);
795         return 0;
796 }
797 core_initcall(netpoll_init);
798
799 void netpoll_cleanup(struct netpoll *np)
800 {
801         struct netpoll_info *npinfo;
802         unsigned long flags;
803
804         if (np->dev) {
805                 npinfo = np->dev->npinfo;
806                 if (npinfo) {
807                         if (npinfo->rx_np == np) {
808                                 spin_lock_irqsave(&npinfo->rx_lock, flags);
809                                 npinfo->rx_np = NULL;
810                                 npinfo->rx_flags &= ~NETPOLL_RX_ENABLED;
811                                 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
812                         }
813
814                         if (atomic_dec_and_test(&npinfo->refcnt)) {
815                                 skb_queue_purge(&npinfo->arp_tx);
816                                 skb_queue_purge(&npinfo->txq);
817                                 cancel_rearming_delayed_work(&npinfo->tx_work);
818
819                                 /* clean after last, unfinished work */
820                                 __skb_queue_purge(&npinfo->txq);
821                                 kfree(npinfo);
822                                 np->dev->npinfo = NULL;
823                         }
824                 }
825
826                 dev_put(np->dev);
827         }
828
829         np->dev = NULL;
830 }
831
832 int netpoll_trap(void)
833 {
834         return atomic_read(&trapped);
835 }
836
837 void netpoll_set_trap(int trap)
838 {
839         if (trap)
840                 atomic_inc(&trapped);
841         else
842                 atomic_dec(&trapped);
843 }
844
845 EXPORT_SYMBOL(netpoll_set_trap);
846 EXPORT_SYMBOL(netpoll_trap);
847 EXPORT_SYMBOL(netpoll_print_options);
848 EXPORT_SYMBOL(netpoll_parse_options);
849 EXPORT_SYMBOL(netpoll_setup);
850 EXPORT_SYMBOL(netpoll_cleanup);
851 EXPORT_SYMBOL(netpoll_send_udp);
852 EXPORT_SYMBOL(netpoll_poll);