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 <linux/export.h>
  27 #include <net/tcp.h>
  28 #include <net/udp.h>
  29 #include <asm/unaligned.h>
  30 #include <trace/events/napi.h>
  31
  32 /*
  33  * We maintain a small pool of fully-sized skbs, to make sure the
  34  * message gets out even in extreme OOM situations.
  35  */
  36
  37 #define MAX_UDP_CHUNK 1460
  38 #define MAX_SKBS 32
  39
  40 static struct sk_buff_head skb_pool;
  41
  42 static atomic_t trapped;
  43
  44 #define USEC_PER_POLL   50
  45 #define NETPOLL_RX_ENABLED  1
  46 #define NETPOLL_RX_DROP     2
  47
  48 #define MAX_SKB_SIZE \
  49                 (MAX_UDP_CHUNK + sizeof(struct udphdr) + \
  50                                 sizeof(struct iphdr) + sizeof(struct ethhdr))
  51
  52 static void zap_completion_queue(void);
  53 static void arp_reply(struct sk_buff *skb);
  54
  55 static unsigned int carrier_timeout = 4;
  56 module_param(carrier_timeout, uint, 0644);
  57
  58 static void queue_process(struct work_struct *work)
  59 {
  60         struct netpoll_info *npinfo =
  61                 container_of(work, struct netpoll_info, tx_work.work);
  62         struct sk_buff *skb;
  63         unsigned long flags;
  64
  65         while ((skb = skb_dequeue(&npinfo->txq))) {
  66                 struct net_device *dev = skb->dev;
  67                 const struct net_device_ops *ops = dev->netdev_ops;
  68                 struct netdev_queue *txq;
  69
  70                 if (!netif_device_present(dev) || !netif_running(dev)) {
  71                         __kfree_skb(skb);
  72                         continue;
  73                 }
  74
  75                 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
  76
  77                 local_irq_save(flags);
  78                 __netif_tx_lock(txq, smp_processor_id());
  79                 if (netif_xmit_frozen_or_stopped(txq) ||
  80                     ops->ndo_start_xmit(skb, dev) != NETDEV_TX_OK) {
  81                         skb_queue_head(&npinfo->txq, skb);
  82                         __netif_tx_unlock(txq);
  83                         local_irq_restore(flags);
  84
  85                         schedule_delayed_work(&npinfo->tx_work, HZ/10);
  86                         return;
  87                 }
  88                 __netif_tx_unlock(txq);
  89                 local_irq_restore(flags);
  90         }
  91 }
  92
  93 static __sum16 checksum_udp(struct sk_buff *skb, struct udphdr *uh,
  94                             unsigned short ulen, __be32 saddr, __be32 daddr)
  95 {
  96         __wsum psum;
  97
  98         if (uh->check == 0 || skb_csum_unnecessary(skb))
  99                 return 0;
 100
 101         psum = csum_tcpudp_nofold(saddr, daddr, ulen, IPPROTO_UDP, 0);
 102
 103         if (skb->ip_summed == CHECKSUM_COMPLETE &&
 104             !csum_fold(csum_add(psum, skb->csum)))
 105                 return 0;
 106
 107         skb->csum = psum;
 108
 109         return __skb_checksum_complete(skb);
 110 }
 111
 112 /*
 113  * Check whether delayed processing was scheduled for our NIC. If so,
 114  * we attempt to grab the poll lock and use ->poll() to pump the card.
 115  * If this fails, either we've recursed in ->poll() or it's already
 116  * running on another CPU.
 117  *
 118  * Note: we don't mask interrupts with this lock because we're using
 119  * trylock here and interrupts are already disabled in the softirq
 120  * case. Further, we test the poll_owner to avoid recursion on UP
 121  * systems where the lock doesn't exist.
 122  *
 123  * In cases where there is bi-directional communications, reading only
 124  * one message at a time can lead to packets being dropped by the
 125  * network adapter, forcing superfluous retries and possibly timeouts.
 126  * Thus, we set our budget to greater than 1.
 127  */
 128 static int poll_one_napi(struct netpoll_info *npinfo,
 129                          struct napi_struct *napi, int budget)
 130 {
 131         int work;
 132
 133         /* net_rx_action's ->poll() invocations and our's are
 134          * synchronized by this test which is only made while
 135          * holding the napi->poll_lock.
 136          */
 137         if (!test_bit(NAPI_STATE_SCHED, &napi->state))
 138                 return budget;
 139
 140         npinfo->rx_flags |= NETPOLL_RX_DROP;
 141         atomic_inc(&trapped);
 142         set_bit(NAPI_STATE_NPSVC, &napi->state);
 143
 144         work = napi->poll(napi, budget);
 145         trace_napi_poll(napi);
 146
 147         clear_bit(NAPI_STATE_NPSVC, &napi->state);
 148         atomic_dec(&trapped);
 149         npinfo->rx_flags &= ~NETPOLL_RX_DROP;
 150
 151         return budget - work;
 152 }
 153
 154 static void poll_napi(struct net_device *dev)
 155 {
 156         struct napi_struct *napi;
 157         int budget = 16;
 158
 159         list_for_each_entry(napi, &dev->napi_list, dev_list) {
 160                 if (napi->poll_owner != smp_processor_id() &&
 161                     spin_trylock(&napi->poll_lock)) {
 162                         budget = poll_one_napi(dev->npinfo, napi, budget);
 163                         spin_unlock(&napi->poll_lock);
 164
 165                         if (!budget)
 166                                 break;
 167                 }
 168         }
 169 }
 170
 171 static void service_arp_queue(struct netpoll_info *npi)
 172 {
 173         if (npi) {
 174                 struct sk_buff *skb;
 175
 176                 while ((skb = skb_dequeue(&npi->arp_tx)))
 177                         arp_reply(skb);
 178         }
 179 }
 180
 181 static void netpoll_poll_dev(struct net_device *dev)
 182 {
 183         const struct net_device_ops *ops;
 184
 185         if (!dev || !netif_running(dev))
 186                 return;
 187
 188         ops = dev->netdev_ops;
 189         if (!ops->ndo_poll_controller)
 190                 return;
 191
 192         /* Process pending work on NIC */
 193         ops->ndo_poll_controller(dev);
 194
 195         poll_napi(dev);
 196
 197         if (dev->priv_flags & IFF_SLAVE) {
 198                 if (dev->npinfo) {
 199                         struct net_device *bond_dev = dev->master;
 200                         struct sk_buff *skb;
 201                         while ((skb = skb_dequeue(&dev->npinfo->arp_tx))) {
 202                                 skb->dev = bond_dev;
 203                                 skb_queue_tail(&bond_dev->npinfo->arp_tx, skb);
 204                         }
 205                 }
 206         }
 207
 208         service_arp_queue(dev->npinfo);
 209
 210         zap_completion_queue();
 211 }
 212
 213 static void refill_skbs(void)
 214 {
 215         struct sk_buff *skb;
 216         unsigned long flags;
 217
 218         spin_lock_irqsave(&skb_pool.lock, flags);
 219         while (skb_pool.qlen < MAX_SKBS) {
 220                 skb = alloc_skb(MAX_SKB_SIZE, GFP_ATOMIC);
 221                 if (!skb)
 222                         break;
 223
 224                 __skb_queue_tail(&skb_pool, skb);
 225         }
 226         spin_unlock_irqrestore(&skb_pool.lock, flags);
 227 }
 228
 229 static void zap_completion_queue(void)
 230 {
 231         unsigned long flags;
 232         struct softnet_data *sd = &get_cpu_var(softnet_data);
 233
 234         if (sd->completion_queue) {
 235                 struct sk_buff *clist;
 236
 237                 local_irq_save(flags);
 238                 clist = sd->completion_queue;
 239                 sd->completion_queue = NULL;
 240                 local_irq_restore(flags);
 241
 242                 while (clist != NULL) {
 243                         struct sk_buff *skb = clist;
 244                         clist = clist->next;
 245                         if (skb->destructor) {
 246                                 atomic_inc(&skb->users);
 247                                 dev_kfree_skb_any(skb); /* put this one back */
 248                         } else {
 249                                 __kfree_skb(skb);
 250                         }
 251                 }
 252         }
 253
 254         put_cpu_var(softnet_data);
 255 }
 256
 257 static struct sk_buff *find_skb(struct netpoll *np, int len, int reserve)
 258 {
 259         int count = 0;
 260         struct sk_buff *skb;
 261
 262         zap_completion_queue();
 263         refill_skbs();
 264 repeat:
 265
 266         skb = alloc_skb(len, GFP_ATOMIC);
 267         if (!skb)
 268                 skb = skb_dequeue(&skb_pool);
 269
 270         if (!skb) {
 271                 if (++count < 10) {
 272                         netpoll_poll_dev(np->dev);
 273                         goto repeat;
 274                 }
 275                 return NULL;
 276         }
 277
 278         atomic_set(&skb->users, 1);
 279         skb_reserve(skb, reserve);
 280         return skb;
 281 }
 282
 283 static int netpoll_owner_active(struct net_device *dev)
 284 {
 285         struct napi_struct *napi;
 286
 287         list_for_each_entry(napi, &dev->napi_list, dev_list) {
 288                 if (napi->poll_owner == smp_processor_id())
 289                         return 1;
 290         }
 291         return 0;
 292 }
 293
 294 void netpoll_send_skb_on_dev(struct netpoll *np, struct sk_buff *skb,
 295                              struct net_device *dev)
 296 {
 297         int status = NETDEV_TX_BUSY;
 298         unsigned long tries;
 299         const struct net_device_ops *ops = dev->netdev_ops;
 300         /* It is up to the caller to keep npinfo alive. */
 301         struct netpoll_info *npinfo = np->dev->npinfo;
 302
 303         if (!npinfo || !netif_running(dev) || !netif_device_present(dev)) {
 304                 __kfree_skb(skb);
 305                 return;
 306         }
 307
 308         /* don't get messages out of order, and no recursion */
 309         if (skb_queue_len(&npinfo->txq) == 0 && !netpoll_owner_active(dev)) {
 310                 struct netdev_queue *txq;
 311                 unsigned long flags;
 312
 313                 txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 314
 315                 local_irq_save(flags);
 316                 /* try until next clock tick */
 317                 for (tries = jiffies_to_usecs(1)/USEC_PER_POLL;
 318                      tries > 0; --tries) {
 319                         if (__netif_tx_trylock(txq)) {
 320                                 if (!netif_xmit_stopped(txq)) {
 321                                         status = ops->ndo_start_xmit(skb, dev);
 322                                         if (status == NETDEV_TX_OK)
 323                                                 txq_trans_update(txq);
 324                                 }
 325                                 __netif_tx_unlock(txq);
 326
 327                                 if (status == NETDEV_TX_OK)
 328                                         break;
 329
 330                         }
 331
 332                         /* tickle device maybe there is some cleanup */
 333                         netpoll_poll_dev(np->dev);
 334
 335                         udelay(USEC_PER_POLL);
 336                 }
 337
 338                 WARN_ONCE(!irqs_disabled(),
 339                         "netpoll_send_skb(): %s enabled interrupts in poll (%pF)\n",
 340                         dev->name, ops->ndo_start_xmit);
 341
 342                 local_irq_restore(flags);
 343         }
 344
 345         if (status != NETDEV_TX_OK) {
 346                 skb_queue_tail(&npinfo->txq, skb);
 347                 schedule_delayed_work(&npinfo->tx_work,0);
 348         }
 349 }
 350 EXPORT_SYMBOL(netpoll_send_skb_on_dev);
 351
 352 void netpoll_send_udp(struct netpoll *np, const char *msg, int len)
 353 {
 354         int total_len, eth_len, ip_len, udp_len;
 355         struct sk_buff *skb;
 356         struct udphdr *udph;
 357         struct iphdr *iph;
 358         struct ethhdr *eth;
 359
 360         udp_len = len + sizeof(*udph);
 361         ip_len = eth_len = udp_len + sizeof(*iph);
 362         total_len = eth_len + ETH_HLEN + NET_IP_ALIGN;
 363
 364         skb = find_skb(np, total_len, total_len - len);
 365         if (!skb)
 366                 return;
 367
 368         skb_copy_to_linear_data(skb, msg, len);
 369         skb->len += len;
 370
 371         skb_push(skb, sizeof(*udph));
 372         skb_reset_transport_header(skb);
 373         udph = udp_hdr(skb);
 374         udph->source = htons(np->local_port);
 375         udph->dest = htons(np->remote_port);
 376         udph->len = htons(udp_len);
 377         udph->check = 0;
 378         udph->check = csum_tcpudp_magic(np->local_ip,
 379                                         np->remote_ip,
 380                                         udp_len, IPPROTO_UDP,
 381                                         csum_partial(udph, udp_len, 0));
 382         if (udph->check == 0)
 383                 udph->check = CSUM_MANGLED_0;
 384
 385         skb_push(skb, sizeof(*iph));
 386         skb_reset_network_header(skb);
 387         iph = ip_hdr(skb);
 388
 389         /* iph->version = 4; iph->ihl = 5; */
 390         put_unaligned(0x45, (unsigned char *)iph);
 391         iph->tos      = 0;
 392         put_unaligned(htons(ip_len), &(iph->tot_len));
 393         iph->id       = 0;
 394         iph->frag_off = 0;
 395         iph->ttl      = 64;
 396         iph->protocol = IPPROTO_UDP;
 397         iph->check    = 0;
 398         put_unaligned(np->local_ip, &(iph->saddr));
 399         put_unaligned(np->remote_ip, &(iph->daddr));
 400         iph->check    = ip_fast_csum((unsigned char *)iph, iph->ihl);
 401
 402         eth = (struct ethhdr *) skb_push(skb, ETH_HLEN);
 403         skb_reset_mac_header(skb);
 404         skb->protocol = eth->h_proto = htons(ETH_P_IP);
 405         memcpy(eth->h_source, np->dev->dev_addr, ETH_ALEN);
 406         memcpy(eth->h_dest, np->remote_mac, ETH_ALEN);
 407
 408         skb->dev = np->dev;
 409
 410         netpoll_send_skb(np, skb);
 411 }
 412 EXPORT_SYMBOL(netpoll_send_udp);
 413
 414 static void arp_reply(struct sk_buff *skb)
 415 {
 416         struct netpoll_info *npinfo = skb->dev->npinfo;
 417         struct arphdr *arp;
 418         unsigned char *arp_ptr;
 419         int size, type = ARPOP_REPLY, ptype = ETH_P_ARP;
 420         __be32 sip, tip;
 421         unsigned char *sha;
 422         struct sk_buff *send_skb;
 423         struct netpoll *np, *tmp;
 424         unsigned long flags;
 425         int hlen, tlen;
 426         int hits = 0;
 427
 428         if (list_empty(&npinfo->rx_np))
 429                 return;
 430
 431         /* Before checking the packet, we do some early
 432            inspection whether this is interesting at all */
 433         spin_lock_irqsave(&npinfo->rx_lock, flags);
 434         list_for_each_entry_safe(np, tmp, &npinfo->rx_np, rx) {
 435                 if (np->dev == skb->dev)
 436                         hits++;
 437         }
 438         spin_unlock_irqrestore(&npinfo->rx_lock, flags);
 439
 440         /* No netpoll struct is using this dev */
 441         if (!hits)
 442                 return;
 443
 444         /* No arp on this interface */
 445         if (skb->dev->flags & IFF_NOARP)
 446                 return;
 447
 448         if (!pskb_may_pull(skb, arp_hdr_len(skb->dev)))
 449                 return;
 450
 451         skb_reset_network_header(skb);
 452         skb_reset_transport_header(skb);
 453         arp = arp_hdr(skb);
 454
 455         if ((arp->ar_hrd != htons(ARPHRD_ETHER) &&
 456              arp->ar_hrd != htons(ARPHRD_IEEE802)) ||
 457             arp->ar_pro != htons(ETH_P_IP) ||
 458             arp->ar_op != htons(ARPOP_REQUEST))
 459                 return;
 460
 461         arp_ptr = (unsigned char *)(arp+1);
 462         /* save the location of the src hw addr */
 463         sha = arp_ptr;
 464         arp_ptr += skb->dev->addr_len;
 465         memcpy(&sip, arp_ptr, 4);
 466         arp_ptr += 4;
 467         /* If we actually cared about dst hw addr,
 468            it would get copied here */
 469         arp_ptr += skb->dev->addr_len;
 470         memcpy(&tip, arp_ptr, 4);
 471
 472         /* Should we ignore arp? */
 473         if (ipv4_is_loopback(tip) || ipv4_is_multicast(tip))
 474                 return;
 475
 476         size = arp_hdr_len(skb->dev);
 477
 478         spin_lock_irqsave(&npinfo->rx_lock, flags);
 479         list_for_each_entry_safe(np, tmp, &npinfo->rx_np, rx) {
 480                 if (tip != np->local_ip)
 481                         continue;
 482
 483                 hlen = LL_RESERVED_SPACE(np->dev);
 484                 tlen = np->dev->needed_tailroom;
 485                 send_skb = find_skb(np, size + hlen + tlen, hlen);
 486                 if (!send_skb)
 487                         continue;
 488
 489                 skb_reset_network_header(send_skb);
 490                 arp = (struct arphdr *) skb_put(send_skb, size);
 491                 send_skb->dev = skb->dev;
 492                 send_skb->protocol = htons(ETH_P_ARP);
 493
 494                 /* Fill the device header for the ARP frame */
 495                 if (dev_hard_header(send_skb, skb->dev, ptype,
 496                                     sha, np->dev->dev_addr,
 497                                     send_skb->len) < 0) {
 498                         kfree_skb(send_skb);
 499                         continue;
 500                 }
 501
 502                 /*
 503                  * Fill out the arp protocol part.
 504                  *
 505                  * we only support ethernet device type,
 506                  * which (according to RFC 1390) should
 507                  * always equal 1 (Ethernet).
 508                  */
 509
 510                 arp->ar_hrd = htons(np->dev->type);
 511                 arp->ar_pro = htons(ETH_P_IP);
 512                 arp->ar_hln = np->dev->addr_len;
 513                 arp->ar_pln = 4;
 514                 arp->ar_op = htons(type);
 515
 516                 arp_ptr = (unsigned char *)(arp + 1);
 517                 memcpy(arp_ptr, np->dev->dev_addr, np->dev->addr_len);
 518                 arp_ptr += np->dev->addr_len;
 519                 memcpy(arp_ptr, &tip, 4);
 520                 arp_ptr += 4;
 521                 memcpy(arp_ptr, sha, np->dev->addr_len);
 522                 arp_ptr += np->dev->addr_len;
 523                 memcpy(arp_ptr, &sip, 4);
 524
 525                 netpoll_send_skb(np, send_skb);
 526
 527                 /* If there are several rx_hooks for the same address,
 528                    we're fine by sending a single reply */
 529                 break;
 530         }
 531         spin_unlock_irqrestore(&npinfo->rx_lock, flags);
 532 }
 533
 534 int __netpoll_rx(struct sk_buff *skb)
 535 {
 536         int proto, len, ulen;
 537         int hits = 0;
 538         const struct iphdr *iph;
 539         struct udphdr *uh;
 540         struct netpoll_info *npinfo = skb->dev->npinfo;
 541         struct netpoll *np, *tmp;
 542
 543         if (list_empty(&npinfo->rx_np))
 544                 goto out;
 545
 546         if (skb->dev->type != ARPHRD_ETHER)
 547                 goto out;
 548
 549         /* check if netpoll clients need ARP */
 550         if (skb->protocol == htons(ETH_P_ARP) &&
 551             atomic_read(&trapped)) {
 552                 skb_queue_tail(&npinfo->arp_tx, skb);
 553                 return 1;
 554         }
 555
 556         proto = ntohs(eth_hdr(skb)->h_proto);
 557         if (proto != ETH_P_IP)
 558                 goto out;
 559         if (skb->pkt_type == PACKET_OTHERHOST)
 560                 goto out;
 561         if (skb_shared(skb))
 562                 goto out;
 563
 564         if (!pskb_may_pull(skb, sizeof(struct iphdr)))
 565                 goto out;
 566         iph = (struct iphdr *)skb->data;
 567         if (iph->ihl < 5 || iph->version != 4)
 568                 goto out;
 569         if (!pskb_may_pull(skb, iph->ihl*4))
 570                 goto out;
 571         iph = (struct iphdr *)skb->data;
 572         if (ip_fast_csum((u8 *)iph, iph->ihl) != 0)
 573                 goto out;
 574
 575         len = ntohs(iph->tot_len);
 576         if (skb->len < len || len < iph->ihl*4)
 577                 goto out;
 578
 579         /*
 580          * Our transport medium may have padded the buffer out.
 581          * Now We trim to the true length of the frame.
 582          */
 583         if (pskb_trim_rcsum(skb, len))
 584                 goto out;
 585
 586         iph = (struct iphdr *)skb->data;
 587         if (iph->protocol != IPPROTO_UDP)
 588                 goto out;
 589
 590         len -= iph->ihl*4;
 591         uh = (struct udphdr *)(((char *)iph) + iph->ihl*4);
 592         ulen = ntohs(uh->len);
 593
 594         if (ulen != len)
 595                 goto out;
 596         if (checksum_udp(skb, uh, ulen, iph->saddr, iph->daddr))
 597                 goto out;
 598
 599         list_for_each_entry_safe(np, tmp, &npinfo->rx_np, rx) {
 600                 if (np->local_ip && np->local_ip != iph->daddr)
 601                         continue;
 602                 if (np->remote_ip && np->remote_ip != iph->saddr)
 603                         continue;
 604                 if (np->local_port && np->local_port != ntohs(uh->dest))
 605                         continue;
 606
 607                 np->rx_hook(np, ntohs(uh->source),
 608                                (char *)(uh+1),
 609                                ulen - sizeof(struct udphdr));
 610                 hits++;
 611         }
 612
 613         if (!hits)
 614                 goto out;
 615
 616         kfree_skb(skb);
 617         return 1;
 618
 619 out:
 620         if (atomic_read(&trapped)) {
 621                 kfree_skb(skb);
 622                 return 1;
 623         }
 624
 625         return 0;
 626 }
 627
 628 void netpoll_print_options(struct netpoll *np)
 629 {
 630         printk(KERN_INFO "%s: local port %d\n",
 631                          np->name, np->local_port);
 632         printk(KERN_INFO "%s: local IP %pI4\n",
 633                          np->name, &np->local_ip);
 634         printk(KERN_INFO "%s: interface '%s'\n",
 635                          np->name, np->dev_name);
 636         printk(KERN_INFO "%s: remote port %d\n",
 637                          np->name, np->remote_port);
 638         printk(KERN_INFO "%s: remote IP %pI4\n",
 639                          np->name, &np->remote_ip);
 640         printk(KERN_INFO "%s: remote ethernet address %pM\n",
 641                          np->name, np->remote_mac);
 642 }
 643 EXPORT_SYMBOL(netpoll_print_options);
 644
 645 int netpoll_parse_options(struct netpoll *np, char *opt)
 646 {
 647         char *cur=opt, *delim;
 648
 649         if (*cur != '@') {
 650                 if ((delim = strchr(cur, '@')) == NULL)
 651                         goto parse_failed;
 652                 *delim = 0;
 653                 np->local_port = simple_strtol(cur, NULL, 10);
 654                 cur = delim;
 655         }
 656         cur++;
 657
 658         if (*cur != '/') {
 659                 if ((delim = strchr(cur, '/')) == NULL)
 660                         goto parse_failed;
 661                 *delim = 0;
 662                 np->local_ip = in_aton(cur);
 663                 cur = delim;
 664         }
 665         cur++;
 666
 667         if (*cur != ',') {
 668                 /* parse out dev name */
 669                 if ((delim = strchr(cur, ',')) == NULL)
 670                         goto parse_failed;
 671                 *delim = 0;
 672                 strlcpy(np->dev_name, cur, sizeof(np->dev_name));
 673                 cur = delim;
 674         }
 675         cur++;
 676
 677         if (*cur != '@') {
 678                 /* dst port */
 679                 if ((delim = strchr(cur, '@')) == NULL)
 680                         goto parse_failed;
 681                 *delim = 0;
 682                 if (*cur == ' ' || *cur == '\t')
 683                         printk(KERN_INFO "%s: warning: whitespace"
 684                                         "is not allowed\n", np->name);
 685                 np->remote_port = simple_strtol(cur, NULL, 10);
 686                 cur = delim;
 687         }
 688         cur++;
 689
 690         /* dst ip */
 691         if ((delim = strchr(cur, '/')) == NULL)
 692                 goto parse_failed;
 693         *delim = 0;
 694         np->remote_ip = in_aton(cur);
 695         cur = delim + 1;
 696
 697         if (*cur != 0) {
 698                 /* MAC address */
 699                 if (!mac_pton(cur, np->remote_mac))
 700                         goto parse_failed;
 701         }
 702
 703         netpoll_print_options(np);
 704
 705         return 0;
 706
 707  parse_failed:
 708         printk(KERN_INFO "%s: couldn't parse config at '%s'!\n",
 709                np->name, cur);
 710         return -1;
 711 }
 712 EXPORT_SYMBOL(netpoll_parse_options);
 713
 714 int __netpoll_setup(struct netpoll *np)
 715 {
 716         struct net_device *ndev = np->dev;
 717         struct netpoll_info *npinfo;
 718         const struct net_device_ops *ops;
 719         unsigned long flags;
 720         int err;
 721
 722         if ((ndev->priv_flags & IFF_DISABLE_NETPOLL) ||
 723             !ndev->netdev_ops->ndo_poll_controller) {
 724                 printk(KERN_ERR "%s: %s doesn't support polling, aborting.\n",
 725                        np->name, np->dev_name);
 726                 err = -ENOTSUPP;
 727                 goto out;
 728         }
 729
 730         if (!ndev->npinfo) {
 731                 npinfo = kmalloc(sizeof(*npinfo), GFP_KERNEL);
 732                 if (!npinfo) {
 733                         err = -ENOMEM;
 734                         goto out;
 735                 }
 736
 737                 npinfo->rx_flags = 0;
 738                 INIT_LIST_HEAD(&npinfo->rx_np);
 739
 740                 spin_lock_init(&npinfo->rx_lock);
 741                 skb_queue_head_init(&npinfo->arp_tx);
 742                 skb_queue_head_init(&npinfo->txq);
 743                 INIT_DELAYED_WORK(&npinfo->tx_work, queue_process);
 744
 745                 atomic_set(&npinfo->refcnt, 1);
 746
 747                 ops = np->dev->netdev_ops;
 748                 if (ops->ndo_netpoll_setup) {
 749                         err = ops->ndo_netpoll_setup(ndev, npinfo);
 750                         if (err)
 751                                 goto free_npinfo;
 752                 }
 753         } else {
 754                 npinfo = ndev->npinfo;
 755                 atomic_inc(&npinfo->refcnt);
 756         }
 757
 758         npinfo->netpoll = np;
 759
 760         if (np->rx_hook) {
 761                 spin_lock_irqsave(&npinfo->rx_lock, flags);
 762                 npinfo->rx_flags |= NETPOLL_RX_ENABLED;
 763                 list_add_tail(&np->rx, &npinfo->rx_np);
 764                 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
 765         }
 766
 767         /* last thing to do is link it to the net device structure */
 768         rcu_assign_pointer(ndev->npinfo, npinfo);
 769
 770         return 0;
 771
 772 free_npinfo:
 773         kfree(npinfo);
 774 out:
 775         return err;
 776 }
 777 EXPORT_SYMBOL_GPL(__netpoll_setup);
 778
 779 int netpoll_setup(struct netpoll *np)
 780 {
 781         struct net_device *ndev = NULL;
 782         struct in_device *in_dev;
 783         int err;
 784
 785         if (np->dev_name)
 786                 ndev = dev_get_by_name(&init_net, np->dev_name);
 787         if (!ndev) {
 788                 printk(KERN_ERR "%s: %s doesn't exist, aborting.\n",
 789                        np->name, np->dev_name);
 790                 return -ENODEV;
 791         }
 792
 793         if (ndev->master) {
 794                 printk(KERN_ERR "%s: %s is a slave device, aborting.\n",
 795                        np->name, np->dev_name);
 796                 err = -EBUSY;
 797                 goto put;
 798         }
 799
 800         if (!netif_running(ndev)) {
 801                 unsigned long atmost, atleast;
 802
 803                 printk(KERN_INFO "%s: device %s not up yet, forcing it\n",
 804                        np->name, np->dev_name);
 805
 806                 rtnl_lock();
 807                 err = dev_open(ndev);
 808                 rtnl_unlock();
 809
 810                 if (err) {
 811                         printk(KERN_ERR "%s: failed to open %s\n",
 812                                np->name, ndev->name);
 813                         goto put;
 814                 }
 815
 816                 atleast = jiffies + HZ/10;
 817                 atmost = jiffies + carrier_timeout * HZ;
 818                 while (!netif_carrier_ok(ndev)) {
 819                         if (time_after(jiffies, atmost)) {
 820                                 printk(KERN_NOTICE
 821                                        "%s: timeout waiting for carrier\n",
 822                                        np->name);
 823                                 break;
 824                         }
 825                         msleep(1);
 826                 }
 827
 828                 /* If carrier appears to come up instantly, we don't
 829                  * trust it and pause so that we don't pump all our
 830                  * queued console messages into the bitbucket.
 831                  */
 832
 833                 if (time_before(jiffies, atleast)) {
 834                         printk(KERN_NOTICE "%s: carrier detect appears"
 835                                " untrustworthy, waiting 4 seconds\n",
 836                                np->name);
 837                         msleep(4000);
 838                 }
 839         }
 840
 841         if (!np->local_ip) {
 842                 rcu_read_lock();
 843                 in_dev = __in_dev_get_rcu(ndev);
 844
 845                 if (!in_dev || !in_dev->ifa_list) {
 846                         rcu_read_unlock();
 847                         printk(KERN_ERR "%s: no IP address for %s, aborting\n",
 848                                np->name, np->dev_name);
 849                         err = -EDESTADDRREQ;
 850                         goto put;
 851                 }
 852
 853                 np->local_ip = in_dev->ifa_list->ifa_local;
 854                 rcu_read_unlock();
 855                 printk(KERN_INFO "%s: local IP %pI4\n", np->name, &np->local_ip);
 856         }
 857
 858         np->dev = ndev;
 859
 860         /* fill up the skb queue */
 861         refill_skbs();
 862
 863         rtnl_lock();
 864         err = __netpoll_setup(np);
 865         rtnl_unlock();
 866
 867         if (err)
 868                 goto put;
 869
 870         return 0;
 871
 872 put:
 873         dev_put(ndev);
 874         return err;
 875 }
 876 EXPORT_SYMBOL(netpoll_setup);
 877
 878 static int __init netpoll_init(void)
 879 {
 880         skb_queue_head_init(&skb_pool);
 881         return 0;
 882 }
 883 core_initcall(netpoll_init);
 884
 885 void __netpoll_cleanup(struct netpoll *np)
 886 {
 887         struct netpoll_info *npinfo;
 888         unsigned long flags;
 889
 890         npinfo = np->dev->npinfo;
 891         if (!npinfo)
 892                 return;
 893
 894         if (!list_empty(&npinfo->rx_np)) {
 895                 spin_lock_irqsave(&npinfo->rx_lock, flags);
 896                 list_del(&np->rx);
 897                 if (list_empty(&npinfo->rx_np))
 898                         npinfo->rx_flags &= ~NETPOLL_RX_ENABLED;
 899                 spin_unlock_irqrestore(&npinfo->rx_lock, flags);
 900         }
 901
 902         if (atomic_dec_and_test(&npinfo->refcnt)) {
 903                 const struct net_device_ops *ops;
 904
 905                 ops = np->dev->netdev_ops;
 906                 if (ops->ndo_netpoll_cleanup)
 907                         ops->ndo_netpoll_cleanup(np->dev);
 908
 909                 RCU_INIT_POINTER(np->dev->npinfo, NULL);
 910
 911                 /* avoid racing with NAPI reading npinfo */
 912                 synchronize_rcu_bh();
 913
 914                 skb_queue_purge(&npinfo->arp_tx);
 915                 skb_queue_purge(&npinfo->txq);
 916                 cancel_delayed_work_sync(&npinfo->tx_work);
 917
 918                 /* clean after last, unfinished work */
 919                 __skb_queue_purge(&npinfo->txq);
 920                 kfree(npinfo);
 921         }
 922 }
 923 EXPORT_SYMBOL_GPL(__netpoll_cleanup);
 924
 925 void netpoll_cleanup(struct netpoll *np)
 926 {
 927         if (!np->dev)
 928                 return;
 929
 930         rtnl_lock();
 931         __netpoll_cleanup(np);
 932         rtnl_unlock();
 933
 934         dev_put(np->dev);
 935         np->dev = NULL;
 936 }
 937 EXPORT_SYMBOL(netpoll_cleanup);
 938
 939 int netpoll_trap(void)
 940 {
 941         return atomic_read(&trapped);
 942 }
 943 EXPORT_SYMBOL(netpoll_trap);
 944
 945 void netpoll_set_trap(int trap)
 946 {
 947         if (trap)
 948                 atomic_inc(&trapped);
 949         else
 950                 atomic_dec(&trapped);
 951 }
 952 EXPORT_SYMBOL(netpoll_set_trap);