d676a561d983afd62db8cf917816367727ae8b16
[linux-2.6.git] / 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/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 <linux/slab.h>
26 #include <net/tcp.h>
27 #include <net/udp.h>
28 #include <asm/unaligned.h>
29 #include <trace/events/napi.h>
30
31 /*
32  * We maintain a small pool of fully-sized skbs, to make sure the
33  * message gets out even in extreme OOM situations.
34  */
35
36 #define MAX_UDP_CHUNK 1460
37 #define MAX_SKBS 32
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_frozen_or_stopped(txq) ||
79                     ops->ndo_start_xmit(skb, dev) != NETDEV_TX_OK) {
80                         skb_queue_head(&npinfo->txq, skb);
81                         __netif_tx_unlock(txq);
82                         local_irq_restore(flags);
83
84                         schedule_delayed_work(&npinfo->tx_work, HZ/10);
85                         return;
86                 }
87                 __netif_tx_unlock(txq);
88                 local_irq_restore(flags);
89         }
90 }
91
92 static __sum16 checksum_udp(struct sk_buff *skb, struct udphdr *uh,
93                             unsigned short ulen, __be32 saddr, __be32 daddr)
94 {
95         __wsum psum;
96
97         if (uh->check == 0 || skb_csum_unnecessary(skb))
98                 return 0;
99
100         psum = csum_tcpudp_nofold(saddr, daddr, ulen, IPPROTO_UDP, 0);
101
102         if (skb->ip_summed == CHECKSUM_COMPLETE &&
103             !csum_fold(csum_add(psum, skb->csum)))
104                 return 0;
105
106         skb->csum = psum;
107
108         return __skb_checksum_complete(skb);
109 }
110
111 /*
112  * Check whether delayed processing was scheduled for our NIC. If so,
113  * we attempt to grab the poll lock and use ->poll() to pump the card.
114  * If this fails, either we've recursed in ->poll() or it's already
115  * running on another CPU.
116  *
117  * Note: we don't mask interrupts with this lock because we're using
118  * trylock here and interrupts are already disabled in the softirq
119  * case. Further, we test the poll_owner to avoid recursion on UP
120  * systems where the lock doesn't exist.
121  *
122  * In cases where there is bi-directional communications, reading only
123  * one message at a time can lead to packets being dropped by the
124  * network adapter, forcing superfluous retries and possibly timeouts.
125  * Thus, we set our budget to greater than 1.
126  */
127 static int poll_one_napi(struct netpoll_info *npinfo,
128                          struct napi_struct *napi, int budget)
129 {
130         int work;
131
132         /* net_rx_action's ->poll() invocations and our's are
133          * synchronized by this test which is only made while
134          * holding the napi->poll_lock.
135          */
136         if (!test_bit(NAPI_STATE_SCHED, &napi->state))
137                 return budget;
138
139         npinfo->rx_flags |= NETPOLL_RX_DROP;
140         atomic_inc(&trapped);
141         set_bit(NAPI_STATE_NPSVC, &napi->state);
142
143         work = napi->poll(napi, budget);
144         trace_napi_poll(napi);
145
146         clear_bit(NAPI_STATE_NPSVC, &napi->state);
147         atomic_dec(&trapped);
148         npinfo->rx_flags &= ~NETPOLL_RX_DROP;
149
150         return budget - work;
151 }
152
153 static void poll_napi(struct net_device *dev)
154 {
155         struct napi_struct *napi;
156         int budget = 16;
157
158         list_for_each_entry(napi, &dev->napi_list, dev_list) {
159                 if (napi->poll_owner != smp_processor_id() &&
160                     spin_trylock(&napi->poll_lock)) {
161                         budget = poll_one_napi(dev->npinfo, napi, budget);
162                         spin_unlock(&napi->poll_lock);
163
164                         if (!budget)
165                                 break;
166                 }
167         }
168 }
169
170 static void service_arp_queue(struct netpoll_info *npi)
171 {
172         if (npi) {
173                 struct sk_buff *skb;
174
175                 while ((skb = skb_dequeue(&npi->arp_tx)))
176                         arp_reply(skb);
177         }
178 }
179
180 static void netpoll_poll_dev(struct net_device *dev)
181 {
182         const struct net_device_ops *ops;
183
184         if (!dev || !netif_running(dev))
185                 return;
186
187         ops = dev->netdev_ops;
188         if (!ops->ndo_poll_controller)
189                 return;
190
191         /* Process pending work on NIC */
192         ops->ndo_poll_controller(dev);
193
194         poll_napi(dev);
195
196         if (dev->priv_flags & IFF_SLAVE) {
197                 if (dev->npinfo) {
198                         struct net_device *bond_dev = dev->master;
199                         struct sk_buff *skb;
200                         while ((skb = skb_dequeue(&dev->npinfo->arp_tx))) {
201                                 skb->dev = bond_dev;
202                                 skb_queue_tail(&bond_dev->npinfo->arp_tx, skb);
203                         }
204                 }
205         }
206
207         service_arp_queue(dev->npinfo);
208
209         zap_completion_queue();
210 }
211
212 static void refill_skbs(void)
213 {
214         struct sk_buff *skb;
215         unsigned long flags;
216
217         spin_lock_irqsave(&skb_pool.lock, flags);
218         while (skb_pool.qlen < MAX_SKBS) {
219                 skb = alloc_skb(MAX_SKB_SIZE, GFP_ATOMIC);
220                 if (!skb)
221                         break;
222
223                 __skb_queue_tail(&skb_pool, skb);
224         }
225         spin_unlock_irqrestore(&skb_pool.lock, flags);
226 }
227
228 static void zap_completion_queue(void)
229 {
230         unsigned long flags;
231         struct softnet_data *sd = &get_cpu_var(softnet_data);
232
233         if (sd->completion_queue) {
234                 struct sk_buff *clist;
235
236                 local_irq_save(flags);
237                 clist = sd->completion_queue;
238                 sd->completion_queue = NULL;
239                 local_irq_restore(flags);
240
241                 while (clist != NULL) {
242                         struct sk_buff *skb = clist;
243                         clist = clist->next;
244                         if (skb->destructor) {
245                                 atomic_inc(&skb->users);
246                                 dev_kfree_skb_any(skb); /* put this one back */
247                         } else {
248                                 __kfree_skb(skb);
249                         }
250                 }
251         }
252
253         put_cpu_var(softnet_data);
254 }
255
256 static struct sk_buff *find_skb(struct netpoll *np, int len, int reserve)
257 {
258         int count = 0;
259         struct sk_buff *skb;
260
261         zap_completion_queue();
262         refill_skbs();
263 repeat:
264
265         skb = alloc_skb(len, GFP_ATOMIC);
266         if (!skb)
267                 skb = skb_dequeue(&skb_pool);
268
269         if (!skb) {
270                 if (++count < 10) {
271                         netpoll_poll_dev(np->dev);
272                         goto repeat;
273                 }
274                 return NULL;
275         }
276
277         atomic_set(&skb->users, 1);
278         skb_reserve(skb, reserve);
279         return skb;
280 }
281
282 static int netpoll_owner_active(struct net_device *dev)
283 {
284         struct napi_struct *napi;
285
286         list_for_each_entry(napi, &dev->napi_list, dev_list) {
287                 if (napi->poll_owner == smp_processor_id())
288                         return 1;
289         }
290         return 0;
291 }
292
293 void netpoll_send_skb_on_dev(struct netpoll *np, struct sk_buff *skb,
294                              struct net_device *dev)
295 {
296         int status = NETDEV_TX_BUSY;
297         unsigned long tries;
298         const struct net_device_ops *ops = dev->netdev_ops;
299         /* It is up to the caller to keep npinfo alive. */
300         struct netpoll_info *npinfo = np->dev->npinfo;
301
302         if (!npinfo || !netif_running(dev) || !netif_device_present(dev)) {
303                 __kfree_skb(skb);
304                 return;
305         }
306
307         /* don't get messages out of order, and no recursion */
308         if (skb_queue_len(&npinfo->txq) == 0 && !netpoll_owner_active(dev)) {
309                 struct netdev_queue *txq;
310                 unsigned long flags;
311
312                 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
313
314                 local_irq_save(flags);
315                 /* try until next clock tick */
316                 for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
317                      tries > 0; --tries) {
318                         if (__netif_tx_trylock(txq)) {
319                                 if (!netif_tx_queue_stopped(txq)) {
320                                         status = ops->ndo_start_xmit(skb, dev);
321                                         if (status == NETDEV_TX_OK)
322                                                 txq_trans_update(txq);
323                                 }
324                                 __netif_tx_unlock(txq);
325
326                                 if (status == NETDEV_TX_OK)
327                                         break;
328
329                         }
330
331                         /* tickle device maybe there is some cleanup */
332                         netpoll_poll_dev(np->dev);
333
334                         udelay(USEC_PER_POLL);
335                 }
336
337                 WARN_ONCE(!irqs_disabled(),
338                         "netpoll_send_skb(): %s enabled interrupts in poll (%pF)\n",
339                         dev->name, ops->ndo_start_xmit);
340
341                 local_irq_restore(flags);
342         }
343
344         if (status != NETDEV_TX_OK) {
345                 skb_queue_tail(&npinfo->txq, skb);
346                 schedule_delayed_work(&npinfo->tx_work,0);
347         }
348 }
349 EXPORT_SYMBOL(netpoll_send_skb_on_dev);
350
351 void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
352 {
353         int total_len, eth_len, ip_len, udp_len;
354         struct sk_buff *skb;
355         struct udphdr *udph;
356         struct iphdr *iph;
357         struct ethhdr *eth;
358
359         udp_len = len + sizeof(*udph);
360         ip_len = eth_len = udp_len + sizeof(*iph);
361         total_len = eth_len + ETH_HLEN + NET_IP_ALIGN;
362
363         skb = find_skb(np, total_len, total_len - len);
364         if (!skb)
365                 return;
366
367         skb_copy_to_linear_data(skb, msg, len);
368         skb->len += len;
369
370         skb_push(skb, sizeof(*udph));
371         skb_reset_transport_header(skb);
372         udph = udp_hdr(skb);
373         udph->source = htons(np->local_port);
374         udph->dest = htons(np->remote_port);
375         udph->len = htons(udp_len);
376         udph->check = 0;
377         udph->check = csum_tcpudp_magic(np->local_ip,
378                                         np->remote_ip,
379                                         udp_len, IPPROTO_UDP,
380                                         csum_partial(udph, udp_len, 0));
381         if (udph->check == 0)
382                 udph->check = CSUM_MANGLED_0;
383
384         skb_push(skb, sizeof(*iph));
385         skb_reset_network_header(skb);
386         iph = ip_hdr(skb);
387
388         /* iph->version = 4; iph->ihl = 5; */
389         put_unaligned(0x45, (unsigned char *)iph);
390         iph->tos      = 0;
391         put_unaligned(htons(ip_len), &(iph->tot_len));
392         iph->id       = 0;
393         iph->frag_off = 0;
394         iph->ttl      = 64;
395         iph->protocol = IPPROTO_UDP;
396         iph->check    = 0;
397         put_unaligned(np->local_ip, &(iph->saddr));
398         put_unaligned(np->remote_ip, &(iph->daddr));
399         iph->check    = ip_fast_csum((unsigned char *)iph, iph->ihl);
400
401         eth = (struct ethhdr *) skb_push(skb, ETH_HLEN);
402         skb_reset_mac_header(skb);
403         skb->protocol = eth->h_proto = htons(ETH_P_IP);
404         memcpy(eth->h_source, np->dev->dev_addr, ETH_ALEN);
405         memcpy(eth->h_dest, np->remote_mac, ETH_ALEN);
406
407         skb->dev = np->dev;
408
409         netpoll_send_skb(np, skb);
410 }
411 EXPORT_SYMBOL(netpoll_send_udp);
412
413 static void arp_reply(struct sk_buff *skb)
414 {
415         struct netpoll_info *npinfo = skb->dev->npinfo;
416         struct arphdr *arp;
417         unsigned char *arp_ptr;
418         int size, type = ARPOP_REPLY, ptype = ETH_P_ARP;
419         __be32 sip, tip;
420         unsigned char *sha;
421         struct sk_buff *send_skb;
422         struct netpoll *np, *tmp;
423         unsigned long flags;
424         int hits = 0;
425
426         if (list_empty(&npinfo->rx_np))
427                 return;
428
429         /* Before checking the packet, we do some early
430            inspection whether this is interesting at all */
431         spin_lock_irqsave(&npinfo->rx_lock, flags);
432         list_for_each_entry_safe(np, tmp, &npinfo->rx_np, rx) {
433                 if (np->dev == skb->dev)
434                         hits++;
435         }
436         spin_unlock_irqrestore(&npinfo->rx_lock, flags);
437
438         /* No netpoll struct is using this dev */
439         if (!hits)
440                 return;
441
442         /* No arp on this interface */
443         if (skb->dev->flags & IFF_NOARP)
444                 return;
445
446         if (!pskb_may_pull(skb, arp_hdr_len(skb->dev)))
447                 return;
448
449         skb_reset_network_header(skb);
450         skb_reset_transport_header(skb);
451         arp = arp_hdr(skb);
452
453         if ((arp->ar_hrd != htons(ARPHRD_ETHER) &&
454              arp->ar_hrd != htons(ARPHRD_IEEE802)) ||
455             arp->ar_pro != htons(ETH_P_IP) ||
456             arp->ar_op != htons(ARPOP_REQUEST))
457                 return;
458
459         arp_ptr = (unsigned char *)(arp+1);
460         /* save the location of the src hw addr */
461         sha = arp_ptr;
462         arp_ptr += skb->dev->addr_len;
463         memcpy(&sip, arp_ptr, 4);
464         arp_ptr += 4;
465         /* If we actually cared about dst hw addr,
466            it would get copied here */
467         arp_ptr += skb->dev->addr_len;
468         memcpy(&tip, arp_ptr, 4);
469
470         /* Should we ignore arp? */
471         if (ipv4_is_loopback(tip) || ipv4_is_multicast(tip))
472                 return;
473
474         size = arp_hdr_len(skb->dev);
475
476         spin_lock_irqsave(&npinfo->rx_lock, flags);
477         list_for_each_entry_safe(np, tmp, &npinfo->rx_np, rx) {
478                 if (tip != np->local_ip)
479                         continue;
480
481                 send_skb = find_skb(np, size + LL_ALLOCATED_SPACE(np->dev),
482                                     LL_RESERVED_SPACE(np->dev));
483                 if (!send_skb)
484                         continue;
485
486                 skb_reset_network_header(send_skb);
487                 arp = (struct arphdr *) skb_put(send_skb, size);
488                 send_skb->dev = skb->dev;
489                 send_skb->protocol = htons(ETH_P_ARP);
490
491                 /* Fill the device header for the ARP frame */
492                 if (dev_hard_header(send_skb, skb->dev, ptype,
493                                     sha, np->dev->dev_addr,
494                                     send_skb->len) < 0) {
495                         kfree_skb(send_skb);
496                         continue;
497                 }
498
499                 /*
500                  * Fill out the arp protocol part.
501                  *
502                  * we only support ethernet device type,
503                  * which (according to RFC 1390) should
504                  * always equal 1 (Ethernet).
505                  */
506
507                 arp->ar_hrd = htons(np->dev->type);
508                 arp->ar_pro = htons(ETH_P_IP);
509                 arp->ar_hln = np->dev->addr_len;
510                 arp->ar_pln = 4;
511                 arp->ar_op = htons(type);
512
513                 arp_ptr = (unsigned char *)(arp + 1);
514                 memcpy(arp_ptr, np->dev->dev_addr, np->dev->addr_len);
515                 arp_ptr += np->dev->addr_len;
516                 memcpy(arp_ptr, &tip, 4);
517                 arp_ptr += 4;
518                 memcpy(arp_ptr, sha, np->dev->addr_len);
519                 arp_ptr += np->dev->addr_len;
520                 memcpy(arp_ptr, &sip, 4);
521
522                 netpoll_send_skb(np, send_skb);
523
524                 /* If there are several rx_hooks for the same address,
525                    we're fine by sending a single reply */
526                 break;
527         }
528         spin_unlock_irqrestore(&npinfo->rx_lock, flags);
529 }
530
531 int __netpoll_rx(struct sk_buff *skb)
532 {
533         int proto, len, ulen;
534         int hits = 0;
535         const struct iphdr *iph;
536         struct udphdr *uh;
537         struct netpoll_info *npinfo = skb->dev->npinfo;
538         struct netpoll *np, *tmp;
539
540         if (list_empty(&npinfo->rx_np))
541                 goto out;
542
543         if (skb->dev->type != ARPHRD_ETHER)
544                 goto out;
545
546         /* check if netpoll clients need ARP */
547         if (skb->protocol == htons(ETH_P_ARP) &&
548             atomic_read(&trapped)) {
549                 skb_queue_tail(&npinfo->arp_tx, skb);
550                 return 1;
551         }
552
553         proto = ntohs(eth_hdr(skb)->h_proto);
554         if (proto != ETH_P_IP)
555                 goto out;
556         if (skb->pkt_type == PACKET_OTHERHOST)
557                 goto out;
558         if (skb_shared(skb))
559                 goto out;
560
561         iph = (struct iphdr *)skb->data;
562         if (!pskb_may_pull(skb, sizeof(struct iphdr)))
563                 goto out;
564         if (iph->ihl < 5 || iph->version != 4)
565                 goto out;
566         if (!pskb_may_pull(skb, iph->ihl*4))
567                 goto out;
568         if (ip_fast_csum((u8 *)iph, iph->ihl) != 0)
569                 goto out;
570
571         len = ntohs(iph->tot_len);
572         if (skb->len < len || len < iph->ihl*4)
573                 goto out;
574
575         /*
576          * Our transport medium may have padded the buffer out.
577          * Now We trim to the true length of the frame.
578          */
579         if (pskb_trim_rcsum(skb, len))
580                 goto out;
581
582         if (iph->protocol != IPPROTO_UDP)
583                 goto out;
584
585         len -= iph->ihl*4;
586         uh = (struct udphdr *)(((char *)iph) + iph->ihl*4);
587         ulen = ntohs(uh->len);
588
589         if (ulen != len)
590                 goto out;
591         if (checksum_udp(skb, uh, ulen, iph->saddr, iph->daddr))
592                 goto out;
593
594         list_for_each_entry_safe(np, tmp, &npinfo->rx_np, rx) {
595                 if (np->local_ip && np->local_ip != iph->daddr)
596                         continue;
597                 if (np->remote_ip && np->remote_ip != iph->saddr)
598                         continue;
599                 if (np->local_port && np->local_port != ntohs(uh->dest))
600                         continue;
601
602                 np->rx_hook(np, ntohs(uh->source),
603                                (char *)(uh+1),
604                                ulen - sizeof(struct udphdr));
605                 hits++;
606         }
607
608         if (!hits)
609                 goto out;
610
611         kfree_skb(skb);
612         return 1;
613
614 out:
615         if (atomic_read(&trapped)) {
616                 kfree_skb(skb);
617                 return 1;
618         }
619
620         return 0;
621 }
622
623 void netpoll_print_options(struct netpoll *np)
624 {
625         printk(KERN_INFO "%s: local port %d\n",
626                          np->name, np->local_port);
627         printk(KERN_INFO "%s: local IP %pI4\n",
628                          np->name, &np->local_ip);
629         printk(KERN_INFO "%s: interface '%s'\n",
630                          np->name, np->dev_name);
631         printk(KERN_INFO "%s: remote port %d\n",
632                          np->name, np->remote_port);
633         printk(KERN_INFO "%s: remote IP %pI4\n",
634                          np->name, &np->remote_ip);
635         printk(KERN_INFO "%s: remote ethernet address %pM\n",
636                          np->name, np->remote_mac);
637 }
638 EXPORT_SYMBOL(netpoll_print_options);
639
640 int netpoll_parse_options(struct netpoll *np, char *opt)
641 {
642         char *cur=opt, *delim;
643
644         if (*cur != '@') {
645                 if ((delim = strchr(cur, '@')) == NULL)
646                         goto parse_failed;
647                 *delim = 0;
648                 np->local_port = simple_strtol(cur, NULL, 10);
649                 cur = delim;
650         }
651         cur++;
652
653         if (*cur != '/') {
654                 if ((delim = strchr(cur, '/')) == NULL)
655                         goto parse_failed;
656                 *delim = 0;
657                 np->local_ip = in_aton(cur);
658                 cur = delim;
659         }
660         cur++;
661
662         if (*cur != ',') {
663                 /* parse out dev name */
664                 if ((delim = strchr(cur, ',')) == NULL)
665                         goto parse_failed;
666                 *delim = 0;
667                 strlcpy(np->dev_name, cur, sizeof(np->dev_name));
668                 cur = delim;
669         }
670         cur++;
671
672         if (*cur != '@') {
673                 /* dst port */
674                 if ((delim = strchr(cur, '@')) == NULL)
675                         goto parse_failed;
676                 *delim = 0;
677                 if (*cur == ' ' || *cur == '\t')
678                         printk(KERN_INFO "%s: warning: whitespace"
679                                         "is not allowed\n", np->name);
680                 np->remote_port = simple_strtol(cur, NULL, 10);
681                 cur = delim;
682         }
683         cur++;
684
685         /* dst ip */
686         if ((delim = strchr(cur, '/')) == NULL)
687                 goto parse_failed;
688         *delim = 0;
689         np->remote_ip = in_aton(cur);
690         cur = delim + 1;
691
692         if (*cur != 0) {
693                 /* MAC address */
694                 if (!mac_pton(cur, np->remote_mac))
695                         goto parse_failed;
696         }
697
698         netpoll_print_options(np);
699
700         return 0;
701
702  parse_failed:
703         printk(KERN_INFO "%s: couldn't parse config at '%s'!\n",
704                np->name, cur);
705         return -1;
706 }
707 EXPORT_SYMBOL(netpoll_parse_options);
708
709 int __netpoll_setup(struct netpoll *np)
710 {
711         struct net_device *ndev = np->dev;
712         struct netpoll_info *npinfo;
713         const struct net_device_ops *ops;
714         unsigned long flags;
715         int err;
716
717         if ((ndev->priv_flags & IFF_DISABLE_NETPOLL) ||
718             !ndev->netdev_ops->ndo_poll_controller) {
719                 printk(KERN_ERR "%s: %s doesn't support polling, aborting.\n",
720                        np->name, np->dev_name);
721                 err = -ENOTSUPP;
722                 goto out;
723         }
724
725         if (!ndev->npinfo) {
726                 npinfo = kmalloc(sizeof(*npinfo), GFP_KERNEL);
727                 if (!npinfo) {
728                         err = -ENOMEM;
729                         goto out;
730                 }
731
732                 npinfo->rx_flags = 0;
733                 INIT_LIST_HEAD(&npinfo->rx_np);
734
735                 spin_lock_init(&npinfo->rx_lock);
736                 skb_queue_head_init(&npinfo->arp_tx);
737                 skb_queue_head_init(&npinfo->txq);
738                 INIT_DELAYED_WORK(&npinfo->tx_work, queue_process);
739
740                 atomic_set(&npinfo->refcnt, 1);
741
742                 ops = np->dev->netdev_ops;
743                 if (ops->ndo_netpoll_setup) {
744                         err = ops->ndo_netpoll_setup(ndev, npinfo);
745                         if (err)
746                                 goto free_npinfo;
747                 }
748         } else {
749                 npinfo = ndev->npinfo;
750                 atomic_inc(&npinfo->refcnt);
751         }
752
753         npinfo->netpoll = np;
754
755         if (np->rx_hook) {
756                 spin_lock_irqsave(&npinfo->rx_lock, flags);
757                 npinfo->rx_flags |= NETPOLL_RX_ENABLED;
758                 list_add_tail(&np->rx, &npinfo->rx_np);
759                 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
760         }
761
762         /* last thing to do is link it to the net device structure */
763         RCU_INIT_POINTER(ndev->npinfo, npinfo);
764
765         return 0;
766
767 free_npinfo:
768         kfree(npinfo);
769 out:
770         return err;
771 }
772 EXPORT_SYMBOL_GPL(__netpoll_setup);
773
774 int netpoll_setup(struct netpoll *np)
775 {
776         struct net_device *ndev = NULL;
777         struct in_device *in_dev;
778         int err;
779
780         if (np->dev_name)
781                 ndev = dev_get_by_name(&init_net, np->dev_name);
782         if (!ndev) {
783                 printk(KERN_ERR "%s: %s doesn't exist, aborting.\n",
784                        np->name, np->dev_name);
785                 return -ENODEV;
786         }
787
788         if (ndev->master) {
789                 printk(KERN_ERR "%s: %s is a slave device, aborting.\n",
790                        np->name, np->dev_name);
791                 err = -EBUSY;
792                 goto put;
793         }
794
795         if (!netif_running(ndev)) {
796                 unsigned long atmost, atleast;
797
798                 printk(KERN_INFO "%s: device %s not up yet, forcing it\n",
799                        np->name, np->dev_name);
800
801                 rtnl_lock();
802                 err = dev_open(ndev);
803                 rtnl_unlock();
804
805                 if (err) {
806                         printk(KERN_ERR "%s: failed to open %s\n",
807                                np->name, ndev->name);
808                         goto put;
809                 }
810
811                 atleast = jiffies + HZ/10;
812                 atmost = jiffies + carrier_timeout * HZ;
813                 while (!netif_carrier_ok(ndev)) {
814                         if (time_after(jiffies, atmost)) {
815                                 printk(KERN_NOTICE
816                                        "%s: timeout waiting for carrier\n",
817                                        np->name);
818                                 break;
819                         }
820                         msleep(1);
821                 }
822
823                 /* If carrier appears to come up instantly, we don't
824                  * trust it and pause so that we don't pump all our
825                  * queued console messages into the bitbucket.
826                  */
827
828                 if (time_before(jiffies, atleast)) {
829                         printk(KERN_NOTICE "%s: carrier detect appears"
830                                " untrustworthy, waiting 4 seconds\n",
831                                np->name);
832                         msleep(4000);
833                 }
834         }
835
836         if (!np->local_ip) {
837                 rcu_read_lock();
838                 in_dev = __in_dev_get_rcu(ndev);
839
840                 if (!in_dev || !in_dev->ifa_list) {
841                         rcu_read_unlock();
842                         printk(KERN_ERR "%s: no IP address for %s, aborting\n",
843                                np->name, np->dev_name);
844                         err = -EDESTADDRREQ;
845                         goto put;
846                 }
847
848                 np->local_ip = in_dev->ifa_list->ifa_local;
849                 rcu_read_unlock();
850                 printk(KERN_INFO "%s: local IP %pI4\n", np->name, &np->local_ip);
851         }
852
853         np->dev = ndev;
854
855         /* fill up the skb queue */
856         refill_skbs();
857
858         rtnl_lock();
859         err = __netpoll_setup(np);
860         rtnl_unlock();
861
862         if (err)
863                 goto put;
864
865         return 0;
866
867 put:
868         dev_put(ndev);
869         return err;
870 }
871 EXPORT_SYMBOL(netpoll_setup);
872
873 static int __init netpoll_init(void)
874 {
875         skb_queue_head_init(&skb_pool);
876         return 0;
877 }
878 core_initcall(netpoll_init);
879
880 void __netpoll_cleanup(struct netpoll *np)
881 {
882         struct netpoll_info *npinfo;
883         unsigned long flags;
884
885         npinfo = np->dev->npinfo;
886         if (!npinfo)
887                 return;
888
889         if (!list_empty(&npinfo->rx_np)) {
890                 spin_lock_irqsave(&npinfo->rx_lock, flags);
891                 list_del(&np->rx);
892                 if (list_empty(&npinfo->rx_np))
893                         npinfo->rx_flags &= ~NETPOLL_RX_ENABLED;
894                 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
895         }
896
897         if (atomic_dec_and_test(&npinfo->refcnt)) {
898                 const struct net_device_ops *ops;
899
900                 ops = np->dev->netdev_ops;
901                 if (ops->ndo_netpoll_cleanup)
902                         ops->ndo_netpoll_cleanup(np->dev);
903
904                 RCU_INIT_POINTER(np->dev->npinfo, NULL);
905
906                 /* avoid racing with NAPI reading npinfo */
907                 synchronize_rcu_bh();
908
909                 skb_queue_purge(&npinfo->arp_tx);
910                 skb_queue_purge(&npinfo->txq);
911                 cancel_delayed_work_sync(&npinfo->tx_work);
912
913                 /* clean after last, unfinished work */
914                 __skb_queue_purge(&npinfo->txq);
915                 kfree(npinfo);
916         }
917 }
918 EXPORT_SYMBOL_GPL(__netpoll_cleanup);
919
920 void netpoll_cleanup(struct netpoll *np)
921 {
922         if (!np->dev)
923                 return;
924
925         rtnl_lock();
926         __netpoll_cleanup(np);
927         rtnl_unlock();
928
929         dev_put(np->dev);
930         np->dev = NULL;
931 }
932 EXPORT_SYMBOL(netpoll_cleanup);
933
934 int netpoll_trap(void)
935 {
936         return atomic_read(&trapped);
937 }
938 EXPORT_SYMBOL(netpoll_trap);
939
940 void netpoll_set_trap(int trap)
941 {
942         if (trap)
943                 atomic_inc(&trapped);
944         else
945                 atomic_dec(&trapped);
946 }
947 EXPORT_SYMBOL(netpoll_set_trap);