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 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         service_arp_queue(dev->npinfo);
 197
 198         zap_completion_queue();
 199 }
 200 EXPORT_SYMBOL(netpoll_poll_dev);
 201
 202 void netpoll_poll(struct netpoll *np)
 203 {
 204         netpoll_poll_dev(np->dev);
 205 }
 206 EXPORT_SYMBOL(netpoll_poll);
 207
 208 static void refill_skbs(void)
 209 {
 210         struct sk_buff *skb;
 211         unsigned long flags;
 212
 213         spin_lock_irqsave(&skb_pool.lock, flags);
 214         while (skb_pool.qlen < MAX_SKBS) {
 215                 skb = alloc_skb(MAX_SKB_SIZE, GFP_ATOMIC);
 216                 if (!skb)
 217                         break;
 218
 219                 __skb_queue_tail(&skb_pool, skb);
 220         }
 221         spin_unlock_irqrestore(&skb_pool.lock, flags);
 222 }
 223
 224 static void zap_completion_queue(void)
 225 {
 226         unsigned long flags;
 227         struct softnet_data *sd = &get_cpu_var(softnet_data);
 228
 229         if (sd->completion_queue) {
 230                 struct sk_buff *clist;
 231
 232                 local_irq_save(flags);
 233                 clist = sd->completion_queue;
 234                 sd->completion_queue = NULL;
 235                 local_irq_restore(flags);
 236
 237                 while (clist != NULL) {
 238                         struct sk_buff *skb = clist;
 239                         clist = clist->next;
 240                         if (skb->destructor) {
 241                                 atomic_inc(&skb->users);
 242                                 dev_kfree_skb_any(skb); /* put this one back */
 243                         } else {
 244                                 __kfree_skb(skb);
 245                         }
 246                 }
 247         }
 248
 249         put_cpu_var(softnet_data);
 250 }
 251
 252 static struct sk_buff *find_skb(struct netpoll *np, int len, int reserve)
 253 {
 254         int count = 0;
 255         struct sk_buff *skb;
 256
 257         zap_completion_queue();
 258         refill_skbs();
 259 repeat:
 260
 261         skb = alloc_skb(len, GFP_ATOMIC);
 262         if (!skb)
 263                 skb = skb_dequeue(&skb_pool);
 264
 265         if (!skb) {
 266                 if (++count < 10) {
 267                         netpoll_poll(np);
 268                         goto repeat;
 269                 }
 270                 return NULL;
 271         }
 272
 273         atomic_set(&skb->users, 1);
 274         skb_reserve(skb, reserve);
 275         return skb;
 276 }
 277
 278 static int netpoll_owner_active(struct net_device *dev)
 279 {
 280         struct napi_struct *napi;
 281
 282         list_for_each_entry(napi, &dev->napi_list, dev_list) {
 283                 if (napi->poll_owner == smp_processor_id())
 284                         return 1;
 285         }
 286         return 0;
 287 }
 288
 289 void netpoll_send_skb_on_dev(struct netpoll *np, struct sk_buff *skb,
 290                              struct net_device *dev)
 291 {
 292         int status = NETDEV_TX_BUSY;
 293         unsigned long tries;
 294         const struct net_device_ops *ops = dev->netdev_ops;
 295         /* It is up to the caller to keep npinfo alive. */
 296         struct netpoll_info *npinfo = np->dev->npinfo;
 297
 298         if (!npinfo || !netif_running(dev) || !netif_device_present(dev)) {
 299                 __kfree_skb(skb);
 300                 return;
 301         }
 302
 303         /* don't get messages out of order, and no recursion */
 304         if (skb_queue_len(&npinfo->txq) == 0 && !netpoll_owner_active(dev)) {
 305                 struct netdev_queue *txq;
 306                 unsigned long flags;
 307
 308                 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 309
 310                 local_irq_save(flags);
 311                 /* try until next clock tick */
 312                 for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
 313                      tries > 0; --tries) {
 314                         if (__netif_tx_trylock(txq)) {
 315                                 if (!netif_tx_queue_stopped(txq)) {
 316                                         dev->priv_flags |= IFF_IN_NETPOLL;
 317                                         status = ops->ndo_start_xmit(skb, dev);
 318                                         dev->priv_flags &= ~IFF_IN_NETPOLL;
 319                                         if (status == NETDEV_TX_OK)
 320                                                 txq_trans_update(txq);
 321                                 }
 322                                 __netif_tx_unlock(txq);
 323
 324                                 if (status == NETDEV_TX_OK)
 325                                         break;
 326
 327                         }
 328
 329                         /* tickle device maybe there is some cleanup */
 330                         netpoll_poll(np);
 331
 332                         udelay(USEC_PER_POLL);
 333                 }
 334
 335                 WARN_ONCE(!irqs_disabled(),
 336                         "netpoll_send_skb(): %s enabled interrupts in poll (%pF)\n",
 337                         dev->name, ops->ndo_start_xmit);
 338
 339                 local_irq_restore(flags);
 340         }
 341
 342         if (status != NETDEV_TX_OK) {
 343                 skb_queue_tail(&npinfo->txq, skb);
 344                 schedule_delayed_work(&npinfo->tx_work,0);
 345         }
 346 }
 347 EXPORT_SYMBOL(netpoll_send_skb_on_dev);
 348
 349 void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
 350 {
 351         int total_len, eth_len, ip_len, udp_len;
 352         struct sk_buff *skb;
 353         struct udphdr *udph;
 354         struct iphdr *iph;
 355         struct ethhdr *eth;
 356
 357         udp_len = len + sizeof(*udph);
 358         ip_len = eth_len = udp_len + sizeof(*iph);
 359         total_len = eth_len + ETH_HLEN + NET_IP_ALIGN;
 360
 361         skb = find_skb(np, total_len, total_len - len);
 362         if (!skb)
 363                 return;
 364
 365         skb_copy_to_linear_data(skb, msg, len);
 366         skb->len += len;
 367
 368         skb_push(skb, sizeof(*udph));
 369         skb_reset_transport_header(skb);
 370         udph = udp_hdr(skb);
 371         udph->source = htons(np->local_port);
 372         udph->dest = htons(np->remote_port);
 373         udph->len = htons(udp_len);
 374         udph->check = 0;
 375         udph->check = csum_tcpudp_magic(np->local_ip,
 376                                         np->remote_ip,
 377                                         udp_len, IPPROTO_UDP,
 378                                         csum_partial(udph, udp_len, 0));
 379         if (udph->check == 0)
 380                 udph->check = CSUM_MANGLED_0;
 381
 382         skb_push(skb, sizeof(*iph));
 383         skb_reset_network_header(skb);
 384         iph = ip_hdr(skb);
 385
 386         /* iph->version = 4; iph->ihl = 5; */
 387         put_unaligned(0x45, (unsigned char *)iph);
 388         iph->tos      = 0;
 389         put_unaligned(htons(ip_len), &(iph->tot_len));
 390         iph->id       = 0;
 391         iph->frag_off = 0;
 392         iph->ttl      = 64;
 393         iph->protocol = IPPROTO_UDP;
 394         iph->check    = 0;
 395         put_unaligned(np->local_ip, &(iph->saddr));
 396         put_unaligned(np->remote_ip, &(iph->daddr));
 397         iph->check    = ip_fast_csum((unsigned char *)iph, iph->ihl);
 398
 399         eth = (struct ethhdr *) skb_push(skb, ETH_HLEN);
 400         skb_reset_mac_header(skb);
 401         skb->protocol = eth->h_proto = htons(ETH_P_IP);
 402         memcpy(eth->h_source, np->dev->dev_addr, ETH_ALEN);
 403         memcpy(eth->h_dest, np->remote_mac, ETH_ALEN);
 404
 405         skb->dev = np->dev;
 406
 407         netpoll_send_skb(np, skb);
 408 }
 409 EXPORT_SYMBOL(netpoll_send_udp);
 410
 411 static void arp_reply(struct sk_buff *skb)
 412 {
 413         struct netpoll_info *npinfo = skb->dev->npinfo;
 414         struct arphdr *arp;
 415         unsigned char *arp_ptr;
 416         int size, type = ARPOP_REPLY, ptype = ETH_P_ARP;
 417         __be32 sip, tip;
 418         unsigned char *sha;
 419         struct sk_buff *send_skb;
 420         struct netpoll *np, *tmp;
 421         unsigned long flags;
 422         int hits = 0;
 423
 424         if (list_empty(&npinfo->rx_np))
 425                 return;
 426
 427         /* Before checking the packet, we do some early
 428            inspection whether this is interesting at all */
 429         spin_lock_irqsave(&npinfo->rx_lock, flags);
 430         list_for_each_entry_safe(np, tmp, &npinfo->rx_np, rx) {
 431                 if (np->dev == skb->dev)
 432                         hits++;
 433         }
 434         spin_unlock_irqrestore(&npinfo->rx_lock, flags);
 435
 436         /* No netpoll struct is using this dev */
 437         if (!hits)
 438                 return;
 439
 440         /* No arp on this interface */
 441         if (skb->dev->flags & IFF_NOARP)
 442                 return;
 443
 444         if (!pskb_may_pull(skb, arp_hdr_len(skb->dev)))
 445                 return;
 446
 447         skb_reset_network_header(skb);
 448         skb_reset_transport_header(skb);
 449         arp = arp_hdr(skb);
 450
 451         if ((arp->ar_hrd != htons(ARPHRD_ETHER) &&
 452              arp->ar_hrd != htons(ARPHRD_IEEE802)) ||
 453             arp->ar_pro != htons(ETH_P_IP) ||
 454             arp->ar_op != htons(ARPOP_REQUEST))
 455                 return;
 456
 457         arp_ptr = (unsigned char *)(arp+1);
 458         /* save the location of the src hw addr */
 459         sha = arp_ptr;
 460         arp_ptr += skb->dev->addr_len;
 461         memcpy(&sip, arp_ptr, 4);
 462         arp_ptr += 4;
 463         /* If we actually cared about dst hw addr,
 464            it would get copied here */
 465         arp_ptr += skb->dev->addr_len;
 466         memcpy(&tip, arp_ptr, 4);
 467
 468         /* Should we ignore arp? */
 469         if (ipv4_is_loopback(tip) || ipv4_is_multicast(tip))
 470                 return;
 471
 472         size = arp_hdr_len(skb->dev);
 473
 474         spin_lock_irqsave(&npinfo->rx_lock, flags);
 475         list_for_each_entry_safe(np, tmp, &npinfo->rx_np, rx) {
 476                 if (tip != np->local_ip)
 477                         continue;
 478
 479                 send_skb = find_skb(np, size + LL_ALLOCATED_SPACE(np->dev),
 480                                     LL_RESERVED_SPACE(np->dev));
 481                 if (!send_skb)
 482                         continue;
 483
 484                 skb_reset_network_header(send_skb);
 485                 arp = (struct arphdr *) skb_put(send_skb, size);
 486                 send_skb->dev = skb->dev;
 487                 send_skb->protocol = htons(ETH_P_ARP);
 488
 489                 /* Fill the device header for the ARP frame */
 490                 if (dev_hard_header(send_skb, skb->dev, ptype,
 491                                     sha, np->dev->dev_addr,
 492                                     send_skb->len) < 0) {
 493                         kfree_skb(send_skb);
 494                         continue;
 495                 }
 496
 497                 /*
 498                  * Fill out the arp protocol part.
 499                  *
 500                  * we only support ethernet device type,
 501                  * which (according to RFC 1390) should
 502                  * always equal 1 (Ethernet).
 503                  */
 504
 505                 arp->ar_hrd = htons(np->dev->type);
 506                 arp->ar_pro = htons(ETH_P_IP);
 507                 arp->ar_hln = np->dev->addr_len;
 508                 arp->ar_pln = 4;
 509                 arp->ar_op = htons(type);
 510
 511                 arp_ptr = (unsigned char *)(arp + 1);
 512                 memcpy(arp_ptr, np->dev->dev_addr, np->dev->addr_len);
 513                 arp_ptr += np->dev->addr_len;
 514                 memcpy(arp_ptr, &tip, 4);
 515                 arp_ptr += 4;
 516                 memcpy(arp_ptr, sha, np->dev->addr_len);
 517                 arp_ptr += np->dev->addr_len;
 518                 memcpy(arp_ptr, &sip, 4);
 519
 520                 netpoll_send_skb(np, send_skb);
 521
 522                 /* If there are several rx_hooks for the same address,
 523                    we're fine by sending a single reply */
 524                 break;
 525         }
 526         spin_unlock_irqrestore(&npinfo->rx_lock, flags);
 527 }
 528
 529 int __netpoll_rx(struct sk_buff *skb)
 530 {
 531         int proto, len, ulen;
 532         int hits = 0;
 533         struct iphdr *iph;
 534         struct udphdr *uh;
 535         struct netpoll_info *npinfo = skb->dev->npinfo;
 536         struct netpoll *np, *tmp;
 537
 538         if (list_empty(&npinfo->rx_np))
 539                 goto out;
 540
 541         if (skb->dev->type != ARPHRD_ETHER)
 542                 goto out;
 543
 544         /* check if netpoll clients need ARP */
 545         if (skb->protocol == htons(ETH_P_ARP) &&
 546             atomic_read(&trapped)) {
 547                 skb_queue_tail(&npinfo->arp_tx, skb);
 548                 return 1;
 549         }
 550
 551         proto = ntohs(eth_hdr(skb)->h_proto);
 552         if (proto != ETH_P_IP)
 553                 goto out;
 554         if (skb->pkt_type == PACKET_OTHERHOST)
 555                 goto out;
 556         if (skb_shared(skb))
 557                 goto out;
 558
 559         iph = (struct iphdr *)skb->data;
 560         if (!pskb_may_pull(skb, sizeof(struct iphdr)))
 561                 goto out;
 562         if (iph->ihl < 5 || iph->version != 4)
 563                 goto out;
 564         if (!pskb_may_pull(skb, iph->ihl*4))
 565                 goto out;
 566         if (ip_fast_csum((u8 *)iph, iph->ihl) != 0)
 567                 goto out;
 568
 569         len = ntohs(iph->tot_len);
 570         if (skb->len < len || len < iph->ihl*4)
 571                 goto out;
 572
 573         /*
 574          * Our transport medium may have padded the buffer out.
 575          * Now We trim to the true length of the frame.
 576          */
 577         if (pskb_trim_rcsum(skb, len))
 578                 goto out;
 579
 580         if (iph->protocol != IPPROTO_UDP)
 581                 goto out;
 582
 583         len -= iph->ihl*4;
 584         uh = (struct udphdr *)(((char *)iph) + iph->ihl*4);
 585         ulen = ntohs(uh->len);
 586
 587         if (ulen != len)
 588                 goto out;
 589         if (checksum_udp(skb, uh, ulen, iph->saddr, iph->daddr))
 590                 goto out;
 591
 592         list_for_each_entry_safe(np, tmp, &npinfo->rx_np, rx) {
 593                 if (np->local_ip && np->local_ip != iph->daddr)
 594                         continue;
 595                 if (np->remote_ip && np->remote_ip != iph->saddr)
 596                         continue;
 597                 if (np->local_port && np->local_port != ntohs(uh->dest))
 598                         continue;
 599
 600                 np->rx_hook(np, ntohs(uh->source),
 601                                (char *)(uh+1),
 602                                ulen - sizeof(struct udphdr));
 603                 hits++;
 604         }
 605
 606         if (!hits)
 607                 goto out;
 608
 609         kfree_skb(skb);
 610         return 1;
 611
 612 out:
 613         if (atomic_read(&trapped)) {
 614                 kfree_skb(skb);
 615                 return 1;
 616         }
 617
 618         return 0;
 619 }
 620
 621 void netpoll_print_options(struct netpoll *np)
 622 {
 623         printk(KERN_INFO "%s: local port %d\n",
 624                          np->name, np->local_port);
 625         printk(KERN_INFO "%s: local IP %pI4\n",
 626                          np->name, &np->local_ip);
 627         printk(KERN_INFO "%s: interface '%s'\n",
 628                          np->name, np->dev_name);
 629         printk(KERN_INFO "%s: remote port %d\n",
 630                          np->name, np->remote_port);
 631         printk(KERN_INFO "%s: remote IP %pI4\n",
 632                          np->name, &np->remote_ip);
 633         printk(KERN_INFO "%s: remote ethernet address %pM\n",
 634                          np->name, np->remote_mac);
 635 }
 636 EXPORT_SYMBOL(netpoll_print_options);
 637
 638 int netpoll_parse_options(struct netpoll *np, char *opt)
 639 {
 640         char *cur=opt, *delim;
 641
 642         if (*cur != '@') {
 643                 if ((delim = strchr(cur, '@')) == NULL)
 644                         goto parse_failed;
 645                 *delim = 0;
 646                 np->local_port = simple_strtol(cur, NULL, 10);
 647                 cur = delim;
 648         }
 649         cur++;
 650
 651         if (*cur != '/') {
 652                 if ((delim = strchr(cur, '/')) == NULL)
 653                         goto parse_failed;
 654                 *delim = 0;
 655                 np->local_ip = in_aton(cur);
 656                 cur = delim;
 657         }
 658         cur++;
 659
 660         if (*cur != ',') {
 661                 /* parse out dev name */
 662                 if ((delim = strchr(cur, ',')) == NULL)
 663                         goto parse_failed;
 664                 *delim = 0;
 665                 strlcpy(np->dev_name, cur, sizeof(np->dev_name));
 666                 cur = delim;
 667         }
 668         cur++;
 669
 670         if (*cur != '@') {
 671                 /* dst port */
 672                 if ((delim = strchr(cur, '@')) == NULL)
 673                         goto parse_failed;
 674                 *delim = 0;
 675                 if (*cur == ' ' || *cur == '\t')
 676                         printk(KERN_INFO "%s: warning: whitespace"
 677                                         "is not allowed\n", np->name);
 678                 np->remote_port = simple_strtol(cur, NULL, 10);
 679                 cur = delim;
 680         }
 681         cur++;
 682
 683         /* dst ip */
 684         if ((delim = strchr(cur, '/')) == NULL)
 685                 goto parse_failed;
 686         *delim = 0;
 687         np->remote_ip = in_aton(cur);
 688         cur = delim + 1;
 689
 690         if (*cur != 0) {
 691                 /* MAC address */
 692                 if ((delim = strchr(cur, ':')) == NULL)
 693                         goto parse_failed;
 694                 *delim = 0;
 695                 np->remote_mac[0] = simple_strtol(cur, NULL, 16);
 696                 cur = delim + 1;
 697                 if ((delim = strchr(cur, ':')) == NULL)
 698                         goto parse_failed;
 699                 *delim = 0;
 700                 np->remote_mac[1] = simple_strtol(cur, NULL, 16);
 701                 cur = delim + 1;
 702                 if ((delim = strchr(cur, ':')) == NULL)
 703                         goto parse_failed;
 704                 *delim = 0;
 705                 np->remote_mac[2] = simple_strtol(cur, NULL, 16);
 706                 cur = delim + 1;
 707                 if ((delim = strchr(cur, ':')) == NULL)
 708                         goto parse_failed;
 709                 *delim = 0;
 710                 np->remote_mac[3] = simple_strtol(cur, NULL, 16);
 711                 cur = delim + 1;
 712                 if ((delim = strchr(cur, ':')) == NULL)
 713                         goto parse_failed;
 714                 *delim = 0;
 715                 np->remote_mac[4] = simple_strtol(cur, NULL, 16);
 716                 cur = delim + 1;
 717                 np->remote_mac[5] = simple_strtol(cur, NULL, 16);
 718         }
 719
 720         netpoll_print_options(np);
 721
 722         return 0;
 723
 724  parse_failed:
 725         printk(KERN_INFO "%s: couldn't parse config at '%s'!\n",
 726                np->name, cur);
 727         return -1;
 728 }
 729 EXPORT_SYMBOL(netpoll_parse_options);
 730
 731 int __netpoll_setup(struct netpoll *np)
 732 {
 733         struct net_device *ndev = np->dev;
 734         struct netpoll_info *npinfo;
 735         const struct net_device_ops *ops;
 736         unsigned long flags;
 737         int err;
 738
 739         if ((ndev->priv_flags & IFF_DISABLE_NETPOLL) ||
 740             !ndev->netdev_ops->ndo_poll_controller) {
 741                 printk(KERN_ERR "%s: %s doesn't support polling, aborting.\n",
 742                        np->name, np->dev_name);
 743                 err = -ENOTSUPP;
 744                 goto out;
 745         }
 746
 747         if (!ndev->npinfo) {
 748                 npinfo = kmalloc(sizeof(*npinfo), GFP_KERNEL);
 749                 if (!npinfo) {
 750                         err = -ENOMEM;
 751                         goto out;
 752                 }
 753
 754                 npinfo->rx_flags = 0;
 755                 INIT_LIST_HEAD(&npinfo->rx_np);
 756
 757                 spin_lock_init(&npinfo->rx_lock);
 758                 skb_queue_head_init(&npinfo->arp_tx);
 759                 skb_queue_head_init(&npinfo->txq);
 760                 INIT_DELAYED_WORK(&npinfo->tx_work, queue_process);
 761
 762                 atomic_set(&npinfo->refcnt, 1);
 763
 764                 ops = np->dev->netdev_ops;
 765                 if (ops->ndo_netpoll_setup) {
 766                         err = ops->ndo_netpoll_setup(ndev, npinfo);
 767                         if (err)
 768                                 goto free_npinfo;
 769                 }
 770         } else {
 771                 npinfo = ndev->npinfo;
 772                 atomic_inc(&npinfo->refcnt);
 773         }
 774
 775         npinfo->netpoll = np;
 776
 777         if (np->rx_hook) {
 778                 spin_lock_irqsave(&npinfo->rx_lock, flags);
 779                 npinfo->rx_flags |= NETPOLL_RX_ENABLED;
 780                 list_add_tail(&np->rx, &npinfo->rx_np);
 781                 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
 782         }
 783
 784         /* last thing to do is link it to the net device structure */
 785         rcu_assign_pointer(ndev->npinfo, npinfo);
 786
 787         return 0;
 788
 789 free_npinfo:
 790         kfree(npinfo);
 791 out:
 792         return err;
 793 }
 794 EXPORT_SYMBOL_GPL(__netpoll_setup);
 795
 796 int netpoll_setup(struct netpoll *np)
 797 {
 798         struct net_device *ndev = NULL;
 799         struct in_device *in_dev;
 800         int err;
 801
 802         if (np->dev_name)
 803                 ndev = dev_get_by_name(&init_net, np->dev_name);
 804         if (!ndev) {
 805                 printk(KERN_ERR "%s: %s doesn't exist, aborting.\n",
 806                        np->name, np->dev_name);
 807                 return -ENODEV;
 808         }
 809
 810         if (!netif_running(ndev)) {
 811                 unsigned long atmost, atleast;
 812
 813                 printk(KERN_INFO "%s: device %s not up yet, forcing it\n",
 814                        np->name, np->dev_name);
 815
 816                 rtnl_lock();
 817                 err = dev_open(ndev);
 818                 rtnl_unlock();
 819
 820                 if (err) {
 821                         printk(KERN_ERR "%s: failed to open %s\n",
 822                                np->name, ndev->name);
 823                         goto put;
 824                 }
 825
 826                 atleast = jiffies + HZ/10;
 827                 atmost = jiffies + carrier_timeout * HZ;
 828                 while (!netif_carrier_ok(ndev)) {
 829                         if (time_after(jiffies, atmost)) {
 830                                 printk(KERN_NOTICE
 831                                        "%s: timeout waiting for carrier\n",
 832                                        np->name);
 833                                 break;
 834                         }
 835                         msleep(1);
 836                 }
 837
 838                 /* If carrier appears to come up instantly, we don't
 839                  * trust it and pause so that we don't pump all our
 840                  * queued console messages into the bitbucket.
 841                  */
 842
 843                 if (time_before(jiffies, atleast)) {
 844                         printk(KERN_NOTICE "%s: carrier detect appears"
 845                                " untrustworthy, waiting 4 seconds\n",
 846                                np->name);
 847                         msleep(4000);
 848                 }
 849         }
 850
 851         if (!np->local_ip) {
 852                 rcu_read_lock();
 853                 in_dev = __in_dev_get_rcu(ndev);
 854
 855                 if (!in_dev || !in_dev->ifa_list) {
 856                         rcu_read_unlock();
 857                         printk(KERN_ERR "%s: no IP address for %s, aborting\n",
 858                                np->name, np->dev_name);
 859                         err = -EDESTADDRREQ;
 860                         goto put;
 861                 }
 862
 863                 np->local_ip = in_dev->ifa_list->ifa_local;
 864                 rcu_read_unlock();
 865                 printk(KERN_INFO "%s: local IP %pI4\n", np->name, &np->local_ip);
 866         }
 867
 868         np->dev = ndev;
 869
 870         /* fill up the skb queue */
 871         refill_skbs();
 872
 873         rtnl_lock();
 874         err = __netpoll_setup(np);
 875         rtnl_unlock();
 876
 877         if (err)
 878                 goto put;
 879
 880         return 0;
 881
 882 put:
 883         dev_put(ndev);
 884         return err;
 885 }
 886 EXPORT_SYMBOL(netpoll_setup);
 887
 888 static int __init netpoll_init(void)
 889 {
 890         skb_queue_head_init(&skb_pool);
 891         return 0;
 892 }
 893 core_initcall(netpoll_init);
 894
 895 void __netpoll_cleanup(struct netpoll *np)
 896 {
 897         struct netpoll_info *npinfo;
 898         unsigned long flags;
 899
 900         npinfo = np->dev->npinfo;
 901         if (!npinfo)
 902                 return;
 903
 904         if (!list_empty(&npinfo->rx_np)) {
 905                 spin_lock_irqsave(&npinfo->rx_lock, flags);
 906                 list_del(&np->rx);
 907                 if (list_empty(&npinfo->rx_np))
 908                         npinfo->rx_flags &= ~NETPOLL_RX_ENABLED;
 909                 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
 910         }
 911
 912         if (atomic_dec_and_test(&npinfo->refcnt)) {
 913                 const struct net_device_ops *ops;
 914
 915                 ops = np->dev->netdev_ops;
 916                 if (ops->ndo_netpoll_cleanup)
 917                         ops->ndo_netpoll_cleanup(np->dev);
 918
 919                 rcu_assign_pointer(np->dev->npinfo, NULL);
 920
 921                 /* avoid racing with NAPI reading npinfo */
 922                 synchronize_rcu_bh();
 923
 924                 skb_queue_purge(&npinfo->arp_tx);
 925                 skb_queue_purge(&npinfo->txq);
 926                 cancel_delayed_work_sync(&npinfo->tx_work);
 927
 928                 /* clean after last, unfinished work */
 929                 __skb_queue_purge(&npinfo->txq);
 930                 kfree(npinfo);
 931         }
 932 }
 933 EXPORT_SYMBOL_GPL(__netpoll_cleanup);
 934
 935 void netpoll_cleanup(struct netpoll *np)
 936 {
 937         if (!np->dev)
 938                 return;
 939
 940         rtnl_lock();
 941         __netpoll_cleanup(np);
 942         rtnl_unlock();
 943
 944         dev_put(np->dev);
 945         np->dev = NULL;
 946 }
 947 EXPORT_SYMBOL(netpoll_cleanup);
 948
 949 int netpoll_trap(void)
 950 {
 951         return atomic_read(&trapped);
 952 }
 953 EXPORT_SYMBOL(netpoll_trap);
 954
 955 void netpoll_set_trap(int trap)
 956 {
 957         if (trap)
 958                 atomic_inc(&trapped);
 959         else
 960                 atomic_dec(&trapped);
 961 }
 962 EXPORT_SYMBOL(netpoll_set_trap);