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