git.oblomov.eu Git - linux-2.6/blob - 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
  62                 if (!netif_device_present(dev) || !netif_running(dev)) {
  63                         __kfree_skb(skb);
  64                         continue;
  65                 }
  66
  67                 local_irq_save(flags);
  68                 netif_tx_lock(dev);
  69                 if ((netif_queue_stopped(dev) ||
  70                      netif_subqueue_stopped(dev, skb)) ||
  71                      dev->hard_start_xmit(skb, dev) != NETDEV_TX_OK) {
  72                         skb_queue_head(&npinfo->txq, skb);
  73                         netif_tx_unlock(dev);
  74                         local_irq_restore(flags);
  75
  76                         schedule_delayed_work(&npinfo->tx_work, HZ/10);
  77                         return;
  78                 }
  79                 netif_tx_unlock(dev);
  80                 local_irq_restore(flags);
  81         }
  82 }
  83
  84 static __sum16 checksum_udp(struct sk_buff *skb, struct udphdr *uh,
  85                             unsigned short ulen, __be32 saddr, __be32 daddr)
  86 {
  87         __wsum psum;
  88
  89         if (uh->check == 0 || skb_csum_unnecessary(skb))
  90                 return 0;
  91
  92         psum = csum_tcpudp_nofold(saddr, daddr, ulen, IPPROTO_UDP, 0);
  93
  94         if (skb->ip_summed == CHECKSUM_COMPLETE &&
  95             !csum_fold(csum_add(psum, skb->csum)))
  96                 return 0;
  97
  98         skb->csum = psum;
  99
 100         return __skb_checksum_complete(skb);
 101 }
 102
 103 /*
 104  * Check whether delayed processing was scheduled for our NIC. If so,
 105  * we attempt to grab the poll lock and use ->poll() to pump the card.
 106  * If this fails, either we've recursed in ->poll() or it's already
 107  * running on another CPU.
 108  *
 109  * Note: we don't mask interrupts with this lock because we're using
 110  * trylock here and interrupts are already disabled in the softirq
 111  * case. Further, we test the poll_owner to avoid recursion on UP
 112  * systems where the lock doesn't exist.
 113  *
 114  * In cases where there is bi-directional communications, reading only
 115  * one message at a time can lead to packets being dropped by the
 116  * network adapter, forcing superfluous retries and possibly timeouts.
 117  * Thus, we set our budget to greater than 1.
 118  */
 119 static int poll_one_napi(struct netpoll_info *npinfo,
 120                          struct napi_struct *napi, int budget)
 121 {
 122         int work;
 123
 124         /* net_rx_action's ->poll() invocations and our's are
 125          * synchronized by this test which is only made while
 126          * holding the napi->poll_lock.
 127          */
 128         if (!test_bit(NAPI_STATE_SCHED, &napi->state))
 129                 return budget;
 130
 131         npinfo->rx_flags |= NETPOLL_RX_DROP;
 132         atomic_inc(&trapped);
 133
 134         work = napi->poll(napi, budget);
 135
 136         atomic_dec(&trapped);
 137         npinfo->rx_flags &= ~NETPOLL_RX_DROP;
 138
 139         return budget - work;
 140 }
 141
 142 static void poll_napi(struct netpoll *np)
 143 {
 144         struct netpoll_info *npinfo = np->dev->npinfo;
 145         struct napi_struct *napi;
 146         int budget = 16;
 147
 148         list_for_each_entry(napi, &np->dev->napi_list, dev_list) {
 149                 if (napi->poll_owner != smp_processor_id() &&
 150                     spin_trylock(&napi->poll_lock)) {
 151                         budget = poll_one_napi(npinfo, napi, budget);
 152                         spin_unlock(&napi->poll_lock);
 153
 154                         if (!budget)
 155                                 break;
 156                 }
 157         }
 158 }
 159
 160 static void service_arp_queue(struct netpoll_info *npi)
 161 {
 162         struct sk_buff *skb;
 163
 164         if (unlikely(!npi))
 165                 return;
 166
 167         skb = skb_dequeue(&npi->arp_tx);
 168
 169         while (skb != NULL) {
 170                 arp_reply(skb);
 171                 skb = skb_dequeue(&npi->arp_tx);
 172         }
 173 }
 174
 175 void netpoll_poll(struct netpoll *np)
 176 {
 177         if (!np->dev || !netif_running(np->dev) || !np->dev->poll_controller)
 178                 return;
 179
 180         /* Process pending work on NIC */
 181         np->dev->poll_controller(np->dev);
 182         if (!list_empty(&np->dev->napi_list))
 183                 poll_napi(np);
 184
 185         service_arp_queue(np->dev->npinfo);
 186
 187         zap_completion_queue();
 188 }
 189
 190 static void refill_skbs(void)
 191 {
 192         struct sk_buff *skb;
 193         unsigned long flags;
 194
 195         spin_lock_irqsave(&skb_pool.lock, flags);
 196         while (skb_pool.qlen < MAX_SKBS) {
 197                 skb = alloc_skb(MAX_SKB_SIZE, GFP_ATOMIC);
 198                 if (!skb)
 199                         break;
 200
 201                 __skb_queue_tail(&skb_pool, skb);
 202         }
 203         spin_unlock_irqrestore(&skb_pool.lock, flags);
 204 }
 205
 206 static void zap_completion_queue(void)
 207 {
 208         unsigned long flags;
 209         struct softnet_data *sd = &get_cpu_var(softnet_data);
 210
 211         if (sd->completion_queue) {
 212                 struct sk_buff *clist;
 213
 214                 local_irq_save(flags);
 215                 clist = sd->completion_queue;
 216                 sd->completion_queue = NULL;
 217                 local_irq_restore(flags);
 218
 219                 while (clist != NULL) {
 220                         struct sk_buff *skb = clist;
 221                         clist = clist->next;
 222                         if (skb->destructor)
 223                                 dev_kfree_skb_any(skb); /* put this one back */
 224                         else
 225                                 __kfree_skb(skb);
 226                 }
 227         }
 228
 229         put_cpu_var(softnet_data);
 230 }
 231
 232 static struct sk_buff *find_skb(struct netpoll *np, int len, int reserve)
 233 {
 234         int count = 0;
 235         struct sk_buff *skb;
 236
 237         zap_completion_queue();
 238         refill_skbs();
 239 repeat:
 240
 241         skb = alloc_skb(len, GFP_ATOMIC);
 242         if (!skb)
 243                 skb = skb_dequeue(&skb_pool);
 244
 245         if (!skb) {
 246                 if (++count < 10) {
 247                         netpoll_poll(np);
 248                         goto repeat;
 249                 }
 250                 return NULL;
 251         }
 252
 253         atomic_set(&skb->users, 1);
 254         skb_reserve(skb, reserve);
 255         return skb;
 256 }
 257
 258 static int netpoll_owner_active(struct net_device *dev)
 259 {
 260         struct napi_struct *napi;
 261
 262         list_for_each_entry(napi, &dev->napi_list, dev_list) {
 263                 if (napi->poll_owner == smp_processor_id())
 264                         return 1;
 265         }
 266         return 0;
 267 }
 268
 269 static void netpoll_send_skb(struct netpoll *np, struct sk_buff *skb)
 270 {
 271         int status = NETDEV_TX_BUSY;
 272         unsigned long tries;
 273         struct net_device *dev = np->dev;
 274         struct netpoll_info *npinfo = np->dev->npinfo;
 275
 276         if (!npinfo || !netif_running(dev) || !netif_device_present(dev)) {
 277                 __kfree_skb(skb);
 278                 return;
 279         }
 280
 281         /* don't get messages out of order, and no recursion */
 282         if (skb_queue_len(&npinfo->txq) == 0 && !netpoll_owner_active(dev)) {
 283                 unsigned long flags;
 284
 285                 local_irq_save(flags);
 286                 /* try until next clock tick */
 287                 for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
 288                      tries > 0; --tries) {
 289                         if (netif_tx_trylock(dev)) {
 290                                 if (!netif_queue_stopped(dev) &&
 291                                     !netif_subqueue_stopped(dev, skb))
 292                                         status = dev->hard_start_xmit(skb, dev);
 293                                 netif_tx_unlock(dev);
 294
 295                                 if (status == NETDEV_TX_OK)
 296                                         break;
 297
 298                         }
 299
 300                         /* tickle device maybe there is some cleanup */
 301                         netpoll_poll(np);
 302
 303                         udelay(USEC_PER_POLL);
 304                 }
 305                 local_irq_restore(flags);
 306         }
 307
 308         if (status != NETDEV_TX_OK) {
 309                 skb_queue_tail(&npinfo->txq, skb);
 310                 schedule_delayed_work(&npinfo->tx_work,0);
 311         }
 312 }
 313
 314 void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
 315 {
 316         int total_len, eth_len, ip_len, udp_len;
 317         struct sk_buff *skb;
 318         struct udphdr *udph;
 319         struct iphdr *iph;
 320         struct ethhdr *eth;
 321
 322         udp_len = len + sizeof(*udph);
 323         ip_len = eth_len = udp_len + sizeof(*iph);
 324         total_len = eth_len + ETH_HLEN + NET_IP_ALIGN;
 325
 326         skb = find_skb(np, total_len, total_len - len);
 327         if (!skb)
 328                 return;
 329
 330         skb_copy_to_linear_data(skb, msg, len);
 331         skb->len += len;
 332
 333         skb_push(skb, sizeof(*udph));
 334         skb_reset_transport_header(skb);
 335         udph = udp_hdr(skb);
 336         udph->source = htons(np->local_port);
 337         udph->dest = htons(np->remote_port);
 338         udph->len = htons(udp_len);
 339         udph->check = 0;
 340         udph->check = csum_tcpudp_magic(htonl(np->local_ip),
 341                                         htonl(np->remote_ip),
 342                                         udp_len, IPPROTO_UDP,
 343                                         csum_partial((unsigned char *)udph, udp_len, 0));
 344         if (udph->check == 0)
 345                 udph->check = CSUM_MANGLED_0;
 346
 347         skb_push(skb, sizeof(*iph));
 348         skb_reset_network_header(skb);
 349         iph = ip_hdr(skb);
 350
 351         /* iph->version = 4; iph->ihl = 5; */
 352         put_unaligned(0x45, (unsigned char *)iph);
 353         iph->tos      = 0;
 354         put_unaligned(htons(ip_len), &(iph->tot_len));
 355         iph->id       = 0;
 356         iph->frag_off = 0;
 357         iph->ttl      = 64;
 358         iph->protocol = IPPROTO_UDP;
 359         iph->check    = 0;
 360         put_unaligned(htonl(np->local_ip), &(iph->saddr));
 361         put_unaligned(htonl(np->remote_ip), &(iph->daddr));
 362         iph->check    = ip_fast_csum((unsigned char *)iph, iph->ihl);
 363
 364         eth = (struct ethhdr *) skb_push(skb, ETH_HLEN);
 365         skb_reset_mac_header(skb);
 366         skb->protocol = eth->h_proto = htons(ETH_P_IP);
 367         memcpy(eth->h_source, np->local_mac, 6);
 368         memcpy(eth->h_dest, np->remote_mac, 6);
 369
 370         skb->dev = np->dev;
 371
 372         netpoll_send_skb(np, skb);
 373 }
 374
 375 static void arp_reply(struct sk_buff *skb)
 376 {
 377         struct netpoll_info *npinfo = skb->dev->npinfo;
 378         struct arphdr *arp;
 379         unsigned char *arp_ptr;
 380         int size, type = ARPOP_REPLY, ptype = ETH_P_ARP;
 381         __be32 sip, tip;
 382         unsigned char *sha;
 383         struct sk_buff *send_skb;
 384         struct netpoll *np = NULL;
 385
 386         if (npinfo->rx_np && npinfo->rx_np->dev == skb->dev)
 387                 np = npinfo->rx_np;
 388         if (!np)
 389                 return;
 390
 391         /* No arp on this interface */
 392         if (skb->dev->flags & IFF_NOARP)
 393                 return;
 394
 395         if (!pskb_may_pull(skb, (sizeof(struct arphdr) +
 396                                  (2 * skb->dev->addr_len) +
 397                                  (2 * sizeof(u32)))))
 398                 return;
 399
 400         skb_reset_network_header(skb);
 401         skb_reset_transport_header(skb);
 402         arp = arp_hdr(skb);
 403
 404         if ((arp->ar_hrd != htons(ARPHRD_ETHER) &&
 405              arp->ar_hrd != htons(ARPHRD_IEEE802)) ||
 406             arp->ar_pro != htons(ETH_P_IP) ||
 407             arp->ar_op != htons(ARPOP_REQUEST))
 408                 return;
 409
 410         arp_ptr = (unsigned char *)(arp+1);
 411         /* save the location of the src hw addr */
 412         sha = arp_ptr;
 413         arp_ptr += skb->dev->addr_len;
 414         memcpy(&sip, arp_ptr, 4);
 415         arp_ptr += 4;
 416         /* if we actually cared about dst hw addr, it would get copied here */
 417         arp_ptr += skb->dev->addr_len;
 418         memcpy(&tip, arp_ptr, 4);
 419
 420         /* Should we ignore arp? */
 421         if (tip != htonl(np->local_ip) || LOOPBACK(tip) || MULTICAST(tip))
 422                 return;
 423
 424         size = sizeof(struct arphdr) + 2 * (skb->dev->addr_len + 4);
 425         send_skb = find_skb(np, size + LL_RESERVED_SPACE(np->dev),
 426                             LL_RESERVED_SPACE(np->dev));
 427
 428         if (!send_skb)
 429                 return;
 430
 431         skb_reset_network_header(send_skb);
 432         arp = (struct arphdr *) skb_put(send_skb, size);
 433         send_skb->dev = skb->dev;
 434         send_skb->protocol = htons(ETH_P_ARP);
 435
 436         /* Fill the device header for the ARP frame */
 437         if (dev_hard_header(send_skb, skb->dev, ptype,
 438                             sha, np->local_mac,
 439                             send_skb->len) < 0) {
 440                 kfree_skb(send_skb);
 441                 return;
 442         }
 443
 444         /*
 445          * Fill out the arp protocol part.
 446          *
 447          * we only support ethernet device type,
 448          * which (according to RFC 1390) should always equal 1 (Ethernet).
 449          */
 450
 451         arp->ar_hrd = htons(np->dev->type);
 452         arp->ar_pro = htons(ETH_P_IP);
 453         arp->ar_hln = np->dev->addr_len;
 454         arp->ar_pln = 4;
 455         arp->ar_op = htons(type);
 456
 457         arp_ptr=(unsigned char *)(arp + 1);
 458         memcpy(arp_ptr, np->dev->dev_addr, np->dev->addr_len);
 459         arp_ptr += np->dev->addr_len;
 460         memcpy(arp_ptr, &tip, 4);
 461         arp_ptr += 4;
 462         memcpy(arp_ptr, sha, np->dev->addr_len);
 463         arp_ptr += np->dev->addr_len;
 464         memcpy(arp_ptr, &sip, 4);
 465
 466         netpoll_send_skb(np, send_skb);
 467 }
 468
 469 int __netpoll_rx(struct sk_buff *skb)
 470 {
 471         int proto, len, ulen;
 472         struct iphdr *iph;
 473         struct udphdr *uh;
 474         struct netpoll_info *npi = skb->dev->npinfo;
 475         struct netpoll *np = npi->rx_np;
 476
 477         if (!np)
 478                 goto out;
 479         if (skb->dev->type != ARPHRD_ETHER)
 480                 goto out;
 481
 482         /* check if netpoll clients need ARP */
 483         if (skb->protocol == htons(ETH_P_ARP) &&
 484             atomic_read(&trapped)) {
 485                 skb_queue_tail(&npi->arp_tx, skb);
 486                 return 1;
 487         }
 488
 489         proto = ntohs(eth_hdr(skb)->h_proto);
 490         if (proto != ETH_P_IP)
 491                 goto out;
 492         if (skb->pkt_type == PACKET_OTHERHOST)
 493                 goto out;
 494         if (skb_shared(skb))
 495                 goto out;
 496
 497         iph = (struct iphdr *)skb->data;
 498         if (!pskb_may_pull(skb, sizeof(struct iphdr)))
 499                 goto out;
 500         if (iph->ihl < 5 || iph->version != 4)
 501                 goto out;
 502         if (!pskb_may_pull(skb, iph->ihl*4))
 503                 goto out;
 504         if (ip_fast_csum((u8 *)iph, iph->ihl) != 0)
 505                 goto out;
 506
 507         len = ntohs(iph->tot_len);
 508         if (skb->len < len || len < iph->ihl*4)
 509                 goto out;
 510
 511         /*
 512          * Our transport medium may have padded the buffer out.
 513          * Now We trim to the true length of the frame.
 514          */
 515         if (pskb_trim_rcsum(skb, len))
 516                 goto out;
 517
 518         if (iph->protocol != IPPROTO_UDP)
 519                 goto out;
 520
 521         len -= iph->ihl*4;
 522         uh = (struct udphdr *)(((char *)iph) + iph->ihl*4);
 523         ulen = ntohs(uh->len);
 524
 525         if (ulen != len)
 526                 goto out;
 527         if (checksum_udp(skb, uh, ulen, iph->saddr, iph->daddr))
 528                 goto out;
 529         if (np->local_ip && np->local_ip != ntohl(iph->daddr))
 530                 goto out;
 531         if (np->remote_ip && np->remote_ip != ntohl(iph->saddr))
 532                 goto out;
 533         if (np->local_port && np->local_port != ntohs(uh->dest))
 534                 goto out;
 535
 536         np->rx_hook(np, ntohs(uh->source),
 537                     (char *)(uh+1),
 538                     ulen - sizeof(struct udphdr));
 539
 540         kfree_skb(skb);
 541         return 1;
 542
 543 out:
 544         if (atomic_read(&trapped)) {
 545                 kfree_skb(skb);
 546                 return 1;
 547         }
 548
 549         return 0;
 550 }
 551
 552 void netpoll_print_options(struct netpoll *np)
 553 {
 554         DECLARE_MAC_BUF(mac);
 555         printk(KERN_INFO "%s: local port %d\n",
 556                          np->name, np->local_port);
 557         printk(KERN_INFO "%s: local IP %d.%d.%d.%d\n",
 558                          np->name, HIPQUAD(np->local_ip));
 559         printk(KERN_INFO "%s: interface %s\n",
 560                          np->name, np->dev_name);
 561         printk(KERN_INFO "%s: remote port %d\n",
 562                          np->name, np->remote_port);
 563         printk(KERN_INFO "%s: remote IP %d.%d.%d.%d\n",
 564                          np->name, HIPQUAD(np->remote_ip));
 565         printk(KERN_INFO "%s: remote ethernet address %s\n",
 566                          np->name, print_mac(mac, np->remote_mac));
 567 }
 568
 569 int netpoll_parse_options(struct netpoll *np, char *opt)
 570 {
 571         char *cur=opt, *delim;
 572
 573         if (*cur != '@') {
 574                 if ((delim = strchr(cur, '@')) == NULL)
 575                         goto parse_failed;
 576                 *delim = 0;
 577                 np->local_port = simple_strtol(cur, NULL, 10);
 578                 cur = delim;
 579         }
 580         cur++;
 581
 582         if (*cur != '/') {
 583                 if ((delim = strchr(cur, '/')) == NULL)
 584                         goto parse_failed;
 585                 *delim = 0;
 586                 np->local_ip = ntohl(in_aton(cur));
 587                 cur = delim;
 588         }
 589         cur++;
 590
 591         if (*cur != ',') {
 592                 /* parse out dev name */
 593                 if ((delim = strchr(cur, ',')) == NULL)
 594                         goto parse_failed;
 595                 *delim = 0;
 596                 strlcpy(np->dev_name, cur, sizeof(np->dev_name));
 597                 cur = delim;
 598         }
 599         cur++;
 600
 601         if (*cur != '@') {
 602                 /* dst port */
 603                 if ((delim = strchr(cur, '@')) == NULL)
 604                         goto parse_failed;
 605                 *delim = 0;
 606                 np->remote_port = simple_strtol(cur, NULL, 10);
 607                 cur = delim;
 608         }
 609         cur++;
 610
 611         /* dst ip */
 612         if ((delim = strchr(cur, '/')) == NULL)
 613                 goto parse_failed;
 614         *delim = 0;
 615         np->remote_ip = ntohl(in_aton(cur));
 616         cur = delim + 1;
 617
 618         if (*cur != 0) {
 619                 /* MAC address */
 620                 if ((delim = strchr(cur, ':')) == NULL)
 621                         goto parse_failed;
 622                 *delim = 0;
 623                 np->remote_mac[0] = simple_strtol(cur, NULL, 16);
 624                 cur = delim + 1;
 625                 if ((delim = strchr(cur, ':')) == NULL)
 626                         goto parse_failed;
 627                 *delim = 0;
 628                 np->remote_mac[1] = simple_strtol(cur, NULL, 16);
 629                 cur = delim + 1;
 630                 if ((delim = strchr(cur, ':')) == NULL)
 631                         goto parse_failed;
 632                 *delim = 0;
 633                 np->remote_mac[2] = simple_strtol(cur, NULL, 16);
 634                 cur = delim + 1;
 635                 if ((delim = strchr(cur, ':')) == NULL)
 636                         goto parse_failed;
 637                 *delim = 0;
 638                 np->remote_mac[3] = simple_strtol(cur, NULL, 16);
 639                 cur = delim + 1;
 640                 if ((delim = strchr(cur, ':')) == NULL)
 641                         goto parse_failed;
 642                 *delim = 0;
 643                 np->remote_mac[4] = simple_strtol(cur, NULL, 16);
 644                 cur = delim + 1;
 645                 np->remote_mac[5] = simple_strtol(cur, NULL, 16);
 646         }
 647
 648         netpoll_print_options(np);
 649
 650         return 0;
 651
 652  parse_failed:
 653         printk(KERN_INFO "%s: couldn't parse config at %s!\n",
 654                np->name, cur);
 655         return -1;
 656 }
 657
 658 int netpoll_setup(struct netpoll *np)
 659 {
 660         struct net_device *ndev = NULL;
 661         struct in_device *in_dev;
 662         struct netpoll_info *npinfo;
 663         unsigned long flags;
 664         int err;
 665
 666         if (np->dev_name)
 667                 ndev = dev_get_by_name(&init_net, np->dev_name);
 668         if (!ndev) {
 669                 printk(KERN_ERR "%s: %s doesn't exist, aborting.\n",
 670                        np->name, np->dev_name);
 671                 return -ENODEV;
 672         }
 673
 674         np->dev = ndev;
 675         if (!ndev->npinfo) {
 676                 npinfo = kmalloc(sizeof(*npinfo), GFP_KERNEL);
 677                 if (!npinfo) {
 678                         err = -ENOMEM;
 679                         goto release;
 680                 }
 681
 682                 npinfo->rx_flags = 0;
 683                 npinfo->rx_np = NULL;
 684
 685                 spin_lock_init(&npinfo->rx_lock);
 686                 skb_queue_head_init(&npinfo->arp_tx);
 687                 skb_queue_head_init(&npinfo->txq);
 688                 INIT_DELAYED_WORK(&npinfo->tx_work, queue_process);
 689
 690                 atomic_set(&npinfo->refcnt, 1);
 691         } else {
 692                 npinfo = ndev->npinfo;
 693                 atomic_inc(&npinfo->refcnt);
 694         }
 695
 696         if (!ndev->poll_controller) {
 697                 printk(KERN_ERR "%s: %s doesn't support polling, aborting.\n",
 698                        np->name, np->dev_name);
 699                 err = -ENOTSUPP;
 700                 goto release;
 701         }
 702
 703         if (!netif_running(ndev)) {
 704                 unsigned long atmost, atleast;
 705
 706                 printk(KERN_INFO "%s: device %s not up yet, forcing it\n",
 707                        np->name, np->dev_name);
 708
 709                 rtnl_lock();
 710                 err = dev_open(ndev);
 711                 rtnl_unlock();
 712
 713                 if (err) {
 714                         printk(KERN_ERR "%s: failed to open %s\n",
 715                                np->name, ndev->name);
 716                         goto release;
 717                 }
 718
 719                 atleast = jiffies + HZ/10;
 720                 atmost = jiffies + 4*HZ;
 721                 while (!netif_carrier_ok(ndev)) {
 722                         if (time_after(jiffies, atmost)) {
 723                                 printk(KERN_NOTICE
 724                                        "%s: timeout waiting for carrier\n",
 725                                        np->name);
 726                                 break;
 727                         }
 728                         cond_resched();
 729                 }
 730
 731                 /* If carrier appears to come up instantly, we don't
 732                  * trust it and pause so that we don't pump all our
 733                  * queued console messages into the bitbucket.
 734                  */
 735
 736                 if (time_before(jiffies, atleast)) {
 737                         printk(KERN_NOTICE "%s: carrier detect appears"
 738                                " untrustworthy, waiting 4 seconds\n",
 739                                np->name);
 740                         msleep(4000);
 741                 }
 742         }
 743
 744         if (is_zero_ether_addr(np->local_mac) && ndev->dev_addr)
 745                 memcpy(np->local_mac, ndev->dev_addr, 6);
 746
 747         if (!np->local_ip) {
 748                 rcu_read_lock();
 749                 in_dev = __in_dev_get_rcu(ndev);
 750
 751                 if (!in_dev || !in_dev->ifa_list) {
 752                         rcu_read_unlock();
 753                         printk(KERN_ERR "%s: no IP address for %s, aborting\n",
 754                                np->name, np->dev_name);
 755                         err = -EDESTADDRREQ;
 756                         goto release;
 757                 }
 758
 759                 np->local_ip = ntohl(in_dev->ifa_list->ifa_local);
 760                 rcu_read_unlock();
 761                 printk(KERN_INFO "%s: local IP %d.%d.%d.%d\n",
 762                        np->name, HIPQUAD(np->local_ip));
 763         }
 764
 765         if (np->rx_hook) {
 766                 spin_lock_irqsave(&npinfo->rx_lock, flags);
 767                 npinfo->rx_flags |= NETPOLL_RX_ENABLED;
 768                 npinfo->rx_np = np;
 769                 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
 770         }
 771
 772         /* fill up the skb queue */
 773         refill_skbs();
 774
 775         /* last thing to do is link it to the net device structure */
 776         ndev->npinfo = npinfo;
 777
 778         /* avoid racing with NAPI reading npinfo */
 779         synchronize_rcu();
 780
 781         return 0;
 782
 783  release:
 784         if (!ndev->npinfo)
 785                 kfree(npinfo);
 786         np->dev = NULL;
 787         dev_put(ndev);
 788         return err;
 789 }
 790
 791 static int __init netpoll_init(void)
 792 {
 793         skb_queue_head_init(&skb_pool);
 794         return 0;
 795 }
 796 core_initcall(netpoll_init);
 797
 798 void netpoll_cleanup(struct netpoll *np)
 799 {
 800         struct netpoll_info *npinfo;
 801         unsigned long flags;
 802
 803         if (np->dev) {
 804                 npinfo = np->dev->npinfo;
 805                 if (npinfo) {
 806                         if (npinfo->rx_np == np) {
 807                                 spin_lock_irqsave(&npinfo->rx_lock, flags);
 808                                 npinfo->rx_np = NULL;
 809                                 npinfo->rx_flags &= ~NETPOLL_RX_ENABLED;
 810                                 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
 811                         }
 812
 813                         if (atomic_dec_and_test(&npinfo->refcnt)) {
 814                                 skb_queue_purge(&npinfo->arp_tx);
 815                                 skb_queue_purge(&npinfo->txq);
 816                                 cancel_rearming_delayed_work(&npinfo->tx_work);
 817
 818                                 /* clean after last, unfinished work */
 819                                 if (!skb_queue_empty(&npinfo->txq)) {
 820                                         struct sk_buff *skb;
 821                                         skb = __skb_dequeue(&npinfo->txq);
 822                                         kfree_skb(skb);
 823                                 }
 824                                 kfree(npinfo);
 825                                 np->dev->npinfo = NULL;
 826                         }
 827                 }
 828
 829                 dev_put(np->dev);
 830         }
 831
 832         np->dev = NULL;
 833 }
 834
 835 int netpoll_trap(void)
 836 {
 837         return atomic_read(&trapped);
 838 }
 839
 840 void netpoll_set_trap(int trap)
 841 {
 842         if (trap)
 843                 atomic_inc(&trapped);
 844         else
 845                 atomic_dec(&trapped);
 846 }
 847
 848 EXPORT_SYMBOL(netpoll_set_trap);
 849 EXPORT_SYMBOL(netpoll_trap);
 850 EXPORT_SYMBOL(netpoll_print_options);
 851 EXPORT_SYMBOL(netpoll_parse_options);
 852 EXPORT_SYMBOL(netpoll_setup);
 853 EXPORT_SYMBOL(netpoll_cleanup);
 854 EXPORT_SYMBOL(netpoll_send_udp);
 855 EXPORT_SYMBOL(netpoll_poll);