2 * Common framework for low-level network console, dump, and debugger code
4 * Sep 8 2003 Matt Mackall <mpm@selenic.com>
6 * based on the netconsole code from:
8 * Copyright (C) 2001 Ingo Molnar <mingo@redhat.com>
9 * Copyright (C) 2002 Red Hat, Inc.
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>
28 #include <asm/unaligned.h>
29 #include <trace/events/napi.h>
32 * We maintain a small pool of fully-sized skbs, to make sure the
33 * message gets out even in extreme OOM situations.
36 #define MAX_UDP_CHUNK 1460
38 #define MAX_QUEUE_DEPTH (MAX_SKBS / 2)
40 static struct sk_buff_head skb_pool;
42 static atomic_t trapped;
44 #define USEC_PER_POLL 50
45 #define NETPOLL_RX_ENABLED 1
46 #define NETPOLL_RX_DROP 2
48 #define MAX_SKB_SIZE \
49 (MAX_UDP_CHUNK + sizeof(struct udphdr) + \
50 sizeof(struct iphdr) + sizeof(struct ethhdr))
52 static void zap_completion_queue(void);
53 static void arp_reply(struct sk_buff *skb);
55 static unsigned int carrier_timeout = 4;
56 module_param(carrier_timeout, uint, 0644);
58 static void queue_process(struct work_struct *work)
60 struct netpoll_info *npinfo =
61 container_of(work, struct netpoll_info, tx_work.work);
65 while ((skb = skb_dequeue(&npinfo->txq))) {
66 struct net_device *dev = skb->dev;
67 const struct net_device_ops *ops = dev->netdev_ops;
68 struct netdev_queue *txq;
70 if (!netif_device_present(dev) || !netif_running(dev)) {
75 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
77 local_irq_save(flags);
78 __netif_tx_lock(txq, smp_processor_id());
79 if (netif_tx_queue_stopped(txq) ||
80 netif_tx_queue_frozen(txq) ||
81 ops->ndo_start_xmit(skb, dev) != NETDEV_TX_OK) {
82 skb_queue_head(&npinfo->txq, skb);
83 __netif_tx_unlock(txq);
84 local_irq_restore(flags);
86 schedule_delayed_work(&npinfo->tx_work, HZ/10);
89 __netif_tx_unlock(txq);
90 local_irq_restore(flags);
94 static __sum16 checksum_udp(struct sk_buff *skb, struct udphdr *uh,
95 unsigned short ulen, __be32 saddr, __be32 daddr)
99 if (uh->check == 0 || skb_csum_unnecessary(skb))
102 psum = csum_tcpudp_nofold(saddr, daddr, ulen, IPPROTO_UDP, 0);
104 if (skb->ip_summed == CHECKSUM_COMPLETE &&
105 !csum_fold(csum_add(psum, skb->csum)))
110 return __skb_checksum_complete(skb);
114 * Check whether delayed processing was scheduled for our NIC. If so,
115 * we attempt to grab the poll lock and use ->poll() to pump the card.
116 * If this fails, either we've recursed in ->poll() or it's already
117 * running on another CPU.
119 * Note: we don't mask interrupts with this lock because we're using
120 * trylock here and interrupts are already disabled in the softirq
121 * case. Further, we test the poll_owner to avoid recursion on UP
122 * systems where the lock doesn't exist.
124 * In cases where there is bi-directional communications, reading only
125 * one message at a time can lead to packets being dropped by the
126 * network adapter, forcing superfluous retries and possibly timeouts.
127 * Thus, we set our budget to greater than 1.
129 static int poll_one_napi(struct netpoll_info *npinfo,
130 struct napi_struct *napi, int budget)
134 /* net_rx_action's ->poll() invocations and our's are
135 * synchronized by this test which is only made while
136 * holding the napi->poll_lock.
138 if (!test_bit(NAPI_STATE_SCHED, &napi->state))
141 npinfo->rx_flags |= NETPOLL_RX_DROP;
142 atomic_inc(&trapped);
143 set_bit(NAPI_STATE_NPSVC, &napi->state);
145 work = napi->poll(napi, budget);
146 trace_napi_poll(napi);
148 clear_bit(NAPI_STATE_NPSVC, &napi->state);
149 atomic_dec(&trapped);
150 npinfo->rx_flags &= ~NETPOLL_RX_DROP;
152 return budget - work;
155 static void poll_napi(struct net_device *dev)
157 struct napi_struct *napi;
160 list_for_each_entry(napi, &dev->napi_list, dev_list) {
161 if (napi->poll_owner != smp_processor_id() &&
162 spin_trylock(&napi->poll_lock)) {
163 budget = poll_one_napi(dev->npinfo, napi, budget);
164 spin_unlock(&napi->poll_lock);
172 static void service_arp_queue(struct netpoll_info *npi)
177 while ((skb = skb_dequeue(&npi->arp_tx)))
182 void netpoll_poll_dev(struct net_device *dev)
184 const struct net_device_ops *ops;
186 if (!dev || !netif_running(dev))
189 ops = dev->netdev_ops;
190 if (!ops->ndo_poll_controller)
193 /* Process pending work on NIC */
194 ops->ndo_poll_controller(dev);
198 service_arp_queue(dev->npinfo);
200 zap_completion_queue();
202 EXPORT_SYMBOL(netpoll_poll_dev);
204 void netpoll_poll(struct netpoll *np)
206 netpoll_poll_dev(np->dev);
208 EXPORT_SYMBOL(netpoll_poll);
210 static void refill_skbs(void)
215 spin_lock_irqsave(&skb_pool.lock, flags);
216 while (skb_pool.qlen < MAX_SKBS) {
217 skb = alloc_skb(MAX_SKB_SIZE, GFP_ATOMIC);
221 __skb_queue_tail(&skb_pool, skb);
223 spin_unlock_irqrestore(&skb_pool.lock, flags);
226 static void zap_completion_queue(void)
229 struct softnet_data *sd = &get_cpu_var(softnet_data);
231 if (sd->completion_queue) {
232 struct sk_buff *clist;
234 local_irq_save(flags);
235 clist = sd->completion_queue;
236 sd->completion_queue = NULL;
237 local_irq_restore(flags);
239 while (clist != NULL) {
240 struct sk_buff *skb = clist;
242 if (skb->destructor) {
243 atomic_inc(&skb->users);
244 dev_kfree_skb_any(skb); /* put this one back */
251 put_cpu_var(softnet_data);
254 static struct sk_buff *find_skb(struct netpoll *np, int len, int reserve)
259 zap_completion_queue();
263 skb = alloc_skb(len, GFP_ATOMIC);
265 skb = skb_dequeue(&skb_pool);
275 atomic_set(&skb->users, 1);
276 skb_reserve(skb, reserve);
280 static int netpoll_owner_active(struct net_device *dev)
282 struct napi_struct *napi;
284 list_for_each_entry(napi, &dev->napi_list, dev_list) {
285 if (napi->poll_owner == smp_processor_id())
291 void netpoll_send_skb_on_dev(struct netpoll *np, struct sk_buff *skb,
292 struct net_device *dev)
294 int status = NETDEV_TX_BUSY;
296 const struct net_device_ops *ops = dev->netdev_ops;
297 /* It is up to the caller to keep npinfo alive. */
298 struct netpoll_info *npinfo = np->dev->npinfo;
300 if (!npinfo || !netif_running(dev) || !netif_device_present(dev)) {
305 /* don't get messages out of order, and no recursion */
306 if (skb_queue_len(&npinfo->txq) == 0 && !netpoll_owner_active(dev)) {
307 struct netdev_queue *txq;
310 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
312 local_irq_save(flags);
313 /* try until next clock tick */
314 for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
315 tries > 0; --tries) {
316 if (__netif_tx_trylock(txq)) {
317 if (!netif_tx_queue_stopped(txq)) {
318 dev->priv_flags |= IFF_IN_NETPOLL;
319 status = ops->ndo_start_xmit(skb, dev);
320 dev->priv_flags &= ~IFF_IN_NETPOLL;
321 if (status == NETDEV_TX_OK)
322 txq_trans_update(txq);
324 __netif_tx_unlock(txq);
326 if (status == NETDEV_TX_OK)
331 /* tickle device maybe there is some cleanup */
334 udelay(USEC_PER_POLL);
337 WARN_ONCE(!irqs_disabled(),
338 "netpoll_send_skb(): %s enabled interrupts in poll (%pF)\n",
339 dev->name, ops->ndo_start_xmit);
341 local_irq_restore(flags);
344 if (status != NETDEV_TX_OK) {
345 skb_queue_tail(&npinfo->txq, skb);
346 schedule_delayed_work(&npinfo->tx_work,0);
349 EXPORT_SYMBOL(netpoll_send_skb_on_dev);
351 void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
353 int total_len, eth_len, ip_len, udp_len;
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;
363 skb = find_skb(np, total_len, total_len - len);
367 skb_copy_to_linear_data(skb, msg, len);
370 skb_push(skb, sizeof(*udph));
371 skb_reset_transport_header(skb);
373 udph->source = htons(np->local_port);
374 udph->dest = htons(np->remote_port);
375 udph->len = htons(udp_len);
377 udph->check = csum_tcpudp_magic(np->local_ip,
379 udp_len, IPPROTO_UDP,
380 csum_partial(udph, udp_len, 0));
381 if (udph->check == 0)
382 udph->check = CSUM_MANGLED_0;
384 skb_push(skb, sizeof(*iph));
385 skb_reset_network_header(skb);
388 /* iph->version = 4; iph->ihl = 5; */
389 put_unaligned(0x45, (unsigned char *)iph);
391 put_unaligned(htons(ip_len), &(iph->tot_len));
395 iph->protocol = IPPROTO_UDP;
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);
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);
409 netpoll_send_skb(np, skb);
411 EXPORT_SYMBOL(netpoll_send_udp);
413 static void arp_reply(struct sk_buff *skb)
415 struct netpoll_info *npinfo = skb->dev->npinfo;
417 unsigned char *arp_ptr;
418 int size, type = ARPOP_REPLY, ptype = ETH_P_ARP;
421 struct sk_buff *send_skb;
422 struct netpoll *np, *tmp;
426 if (list_empty(&npinfo->rx_np))
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)
436 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
438 /* No netpoll struct is using this dev */
442 /* No arp on this interface */
443 if (skb->dev->flags & IFF_NOARP)
446 if (!pskb_may_pull(skb, arp_hdr_len(skb->dev)))
449 skb_reset_network_header(skb);
450 skb_reset_transport_header(skb);
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))
459 arp_ptr = (unsigned char *)(arp+1);
460 /* save the location of the src hw addr */
462 arp_ptr += skb->dev->addr_len;
463 memcpy(&sip, 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);
470 /* Should we ignore arp? */
471 if (ipv4_is_loopback(tip) || ipv4_is_multicast(tip))
474 size = arp_hdr_len(skb->dev);
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)
481 send_skb = find_skb(np, size + LL_ALLOCATED_SPACE(np->dev),
482 LL_RESERVED_SPACE(np->dev));
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);
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) {
500 * Fill out the arp protocol part.
502 * we only support ethernet device type,
503 * which (according to RFC 1390) should
504 * always equal 1 (Ethernet).
507 arp->ar_hrd = htons(np->dev->type);
508 arp->ar_pro = htons(ETH_P_IP);
509 arp->ar_hln = np->dev->addr_len;
511 arp->ar_op = htons(type);
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);
518 memcpy(arp_ptr, sha, np->dev->addr_len);
519 arp_ptr += np->dev->addr_len;
520 memcpy(arp_ptr, &sip, 4);
522 netpoll_send_skb(np, send_skb);
524 /* If there are several rx_hooks for the same address,
525 we're fine by sending a single reply */
528 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
531 int __netpoll_rx(struct sk_buff *skb)
533 int proto, len, ulen;
537 struct netpoll_info *npinfo = skb->dev->npinfo;
538 struct netpoll *np, *tmp;
540 if (list_empty(&npinfo->rx_np))
543 if (skb->dev->type != ARPHRD_ETHER)
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);
553 proto = ntohs(eth_hdr(skb)->h_proto);
554 if (proto != ETH_P_IP)
556 if (skb->pkt_type == PACKET_OTHERHOST)
561 iph = (struct iphdr *)skb->data;
562 if (!pskb_may_pull(skb, sizeof(struct iphdr)))
564 if (iph->ihl < 5 || iph->version != 4)
566 if (!pskb_may_pull(skb, iph->ihl*4))
568 if (ip_fast_csum((u8 *)iph, iph->ihl) != 0)
571 len = ntohs(iph->tot_len);
572 if (skb->len < len || len < iph->ihl*4)
576 * Our transport medium may have padded the buffer out.
577 * Now We trim to the true length of the frame.
579 if (pskb_trim_rcsum(skb, len))
582 if (iph->protocol != IPPROTO_UDP)
586 uh = (struct udphdr *)(((char *)iph) + iph->ihl*4);
587 ulen = ntohs(uh->len);
591 if (checksum_udp(skb, uh, ulen, iph->saddr, iph->daddr))
594 list_for_each_entry_safe(np, tmp, &npinfo->rx_np, rx) {
595 if (np->local_ip && np->local_ip != iph->daddr)
597 if (np->remote_ip && np->remote_ip != iph->saddr)
599 if (np->local_port && np->local_port != ntohs(uh->dest))
602 np->rx_hook(np, ntohs(uh->source),
604 ulen - sizeof(struct udphdr));
615 if (atomic_read(&trapped)) {
623 void netpoll_print_options(struct netpoll *np)
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);
638 EXPORT_SYMBOL(netpoll_print_options);
640 int netpoll_parse_options(struct netpoll *np, char *opt)
642 char *cur=opt, *delim;
645 if ((delim = strchr(cur, '@')) == NULL)
648 np->local_port = simple_strtol(cur, NULL, 10);
654 if ((delim = strchr(cur, '/')) == NULL)
657 np->local_ip = in_aton(cur);
663 /* parse out dev name */
664 if ((delim = strchr(cur, ',')) == NULL)
667 strlcpy(np->dev_name, cur, sizeof(np->dev_name));
674 if ((delim = strchr(cur, '@')) == NULL)
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);
686 if ((delim = strchr(cur, '/')) == NULL)
689 np->remote_ip = in_aton(cur);
694 if ((delim = strchr(cur, ':')) == NULL)
697 np->remote_mac[0] = simple_strtol(cur, NULL, 16);
699 if ((delim = strchr(cur, ':')) == NULL)
702 np->remote_mac[1] = simple_strtol(cur, NULL, 16);
704 if ((delim = strchr(cur, ':')) == NULL)
707 np->remote_mac[2] = simple_strtol(cur, NULL, 16);
709 if ((delim = strchr(cur, ':')) == NULL)
712 np->remote_mac[3] = simple_strtol(cur, NULL, 16);
714 if ((delim = strchr(cur, ':')) == NULL)
717 np->remote_mac[4] = simple_strtol(cur, NULL, 16);
719 np->remote_mac[5] = simple_strtol(cur, NULL, 16);
722 netpoll_print_options(np);
727 printk(KERN_INFO "%s: couldn't parse config at '%s'!\n",
731 EXPORT_SYMBOL(netpoll_parse_options);
733 int __netpoll_setup(struct netpoll *np)
735 struct net_device *ndev = np->dev;
736 struct netpoll_info *npinfo;
737 const struct net_device_ops *ops;
741 if ((ndev->priv_flags & IFF_DISABLE_NETPOLL) ||
742 !ndev->netdev_ops->ndo_poll_controller) {
743 printk(KERN_ERR "%s: %s doesn't support polling, aborting.\n",
744 np->name, np->dev_name);
750 npinfo = kmalloc(sizeof(*npinfo), GFP_KERNEL);
756 npinfo->rx_flags = 0;
757 INIT_LIST_HEAD(&npinfo->rx_np);
759 spin_lock_init(&npinfo->rx_lock);
760 skb_queue_head_init(&npinfo->arp_tx);
761 skb_queue_head_init(&npinfo->txq);
762 INIT_DELAYED_WORK(&npinfo->tx_work, queue_process);
764 atomic_set(&npinfo->refcnt, 1);
766 ops = np->dev->netdev_ops;
767 if (ops->ndo_netpoll_setup) {
768 err = ops->ndo_netpoll_setup(ndev, npinfo);
773 npinfo = ndev->npinfo;
774 atomic_inc(&npinfo->refcnt);
777 npinfo->netpoll = np;
780 spin_lock_irqsave(&npinfo->rx_lock, flags);
781 npinfo->rx_flags |= NETPOLL_RX_ENABLED;
782 list_add_tail(&np->rx, &npinfo->rx_np);
783 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
786 /* last thing to do is link it to the net device structure */
787 rcu_assign_pointer(ndev->npinfo, npinfo);
796 EXPORT_SYMBOL_GPL(__netpoll_setup);
798 int netpoll_setup(struct netpoll *np)
800 struct net_device *ndev = NULL;
801 struct in_device *in_dev;
805 ndev = dev_get_by_name(&init_net, np->dev_name);
807 printk(KERN_ERR "%s: %s doesn't exist, aborting.\n",
808 np->name, np->dev_name);
812 if (!netif_running(ndev)) {
813 unsigned long atmost, atleast;
815 printk(KERN_INFO "%s: device %s not up yet, forcing it\n",
816 np->name, np->dev_name);
819 err = dev_open(ndev);
823 printk(KERN_ERR "%s: failed to open %s\n",
824 np->name, ndev->name);
828 atleast = jiffies + HZ/10;
829 atmost = jiffies + carrier_timeout * HZ;
830 while (!netif_carrier_ok(ndev)) {
831 if (time_after(jiffies, atmost)) {
833 "%s: timeout waiting for carrier\n",
840 /* If carrier appears to come up instantly, we don't
841 * trust it and pause so that we don't pump all our
842 * queued console messages into the bitbucket.
845 if (time_before(jiffies, atleast)) {
846 printk(KERN_NOTICE "%s: carrier detect appears"
847 " untrustworthy, waiting 4 seconds\n",
855 in_dev = __in_dev_get_rcu(ndev);
857 if (!in_dev || !in_dev->ifa_list) {
859 printk(KERN_ERR "%s: no IP address for %s, aborting\n",
860 np->name, np->dev_name);
865 np->local_ip = in_dev->ifa_list->ifa_local;
867 printk(KERN_INFO "%s: local IP %pI4\n", np->name, &np->local_ip);
872 /* fill up the skb queue */
876 err = __netpoll_setup(np);
888 EXPORT_SYMBOL(netpoll_setup);
890 static int __init netpoll_init(void)
892 skb_queue_head_init(&skb_pool);
895 core_initcall(netpoll_init);
897 void __netpoll_cleanup(struct netpoll *np)
899 struct netpoll_info *npinfo;
902 npinfo = np->dev->npinfo;
906 if (!list_empty(&npinfo->rx_np)) {
907 spin_lock_irqsave(&npinfo->rx_lock, flags);
909 if (list_empty(&npinfo->rx_np))
910 npinfo->rx_flags &= ~NETPOLL_RX_ENABLED;
911 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
914 if (atomic_dec_and_test(&npinfo->refcnt)) {
915 const struct net_device_ops *ops;
917 ops = np->dev->netdev_ops;
918 if (ops->ndo_netpoll_cleanup)
919 ops->ndo_netpoll_cleanup(np->dev);
921 rcu_assign_pointer(np->dev->npinfo, NULL);
923 /* avoid racing with NAPI reading npinfo */
924 synchronize_rcu_bh();
926 skb_queue_purge(&npinfo->arp_tx);
927 skb_queue_purge(&npinfo->txq);
928 cancel_rearming_delayed_work(&npinfo->tx_work);
930 /* clean after last, unfinished work */
931 __skb_queue_purge(&npinfo->txq);
935 EXPORT_SYMBOL_GPL(__netpoll_cleanup);
937 void netpoll_cleanup(struct netpoll *np)
943 __netpoll_cleanup(np);
949 EXPORT_SYMBOL(netpoll_cleanup);
951 int netpoll_trap(void)
953 return atomic_read(&trapped);
955 EXPORT_SYMBOL(netpoll_trap);
957 void netpoll_set_trap(int trap)
960 atomic_inc(&trapped);
962 atomic_dec(&trapped);
964 EXPORT_SYMBOL(netpoll_set_trap);