drivers/net/virtio_net.c

   1 /* A simple network driver using virtio.
   2  *
   3  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation; either version 2 of the License, or
   8  * (at your option) any later version.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  18  */
  19 //#define DEBUG
  20 #include <linux/netdevice.h>
  21 #include <linux/etherdevice.h>
  22 #include <linux/module.h>
  23 #include <linux/virtio.h>
  24 #include <linux/virtio_net.h>
  25 #include <linux/scatterlist.h>
  26
  27 static int napi_weight = 128;
  28 module_param(napi_weight, int, 0444);
  29
  30 static int csum = 1, gso = 1;
  31 module_param(csum, bool, 0444);
  32 module_param(gso, bool, 0444);
  33
  34 /* FIXME: MTU in config. */
  35 #define MAX_PACKET_LEN (ETH_HLEN+ETH_DATA_LEN)
  36
  37 struct virtnet_info
  38 {
  39         struct virtio_device *vdev;
  40         struct virtqueue *rvq, *svq;
  41         struct net_device *dev;
  42         struct napi_struct napi;
  43
  44         /* The skb we couldn't send because buffers were full. */
  45         struct sk_buff *last_xmit_skb;
  46
  47         struct timer_list xmit_free_timer;
  48
  49         /* Number of input buffers, and max we've ever had. */
  50         unsigned int num, max;
  51
  52         /* For cleaning up after transmission. */
  53         struct tasklet_struct tasklet;
  54
  55         /* Receive & send queues. */
  56         struct sk_buff_head recv;
  57         struct sk_buff_head send;
  58 };
  59
  60 static inline struct virtio_net_hdr *skb_vnet_hdr(struct sk_buff *skb)
  61 {
  62         return (struct virtio_net_hdr *)skb->cb;
  63 }
  64
  65 static inline void vnet_hdr_to_sg(struct scatterlist *sg, struct sk_buff *skb)
  66 {
  67         sg_init_one(sg, skb_vnet_hdr(skb), sizeof(struct virtio_net_hdr));
  68 }
  69
  70 static void skb_xmit_done(struct virtqueue *svq)
  71 {
  72         struct virtnet_info *vi = svq->vdev->priv;
  73
  74         /* Suppress further interrupts. */
  75         svq->vq_ops->disable_cb(svq);
  76
  77         /* We were waiting for more output buffers. */
  78         netif_wake_queue(vi->dev);
  79
  80         /* Make sure we re-xmit last_xmit_skb: if there are no more packets
  81          * queued, start_xmit won't be called. */
  82         tasklet_schedule(&vi->tasklet);
  83 }
  84
  85 static void receive_skb(struct net_device *dev, struct sk_buff *skb,
  86                         unsigned len)
  87 {
  88         struct virtio_net_hdr *hdr = skb_vnet_hdr(skb);
  89
  90         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
  91                 pr_debug("%s: short packet %i\n", dev->name, len);
  92                 dev->stats.rx_length_errors++;
  93                 goto drop;
  94         }
  95         len -= sizeof(struct virtio_net_hdr);
  96         BUG_ON(len > MAX_PACKET_LEN);
  97
  98         skb_trim(skb, len);
  99
 100         dev->stats.rx_bytes += skb->len;
 101         dev->stats.rx_packets++;
 102
 103         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 104                 pr_debug("Needs csum!\n");
 105                 if (!skb_partial_csum_set(skb,hdr->csum_start,hdr->csum_offset))
 106                         goto frame_err;
 107         }
 108
 109         skb->protocol = eth_type_trans(skb, dev);
 110         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 111                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 112
 113         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 114                 pr_debug("GSO!\n");
 115                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 116                 case VIRTIO_NET_HDR_GSO_TCPV4:
 117                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 118                         break;
 119                 case VIRTIO_NET_HDR_GSO_UDP:
 120                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 121                         break;
 122                 case VIRTIO_NET_HDR_GSO_TCPV6:
 123                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 124                         break;
 125                 default:
 126                         if (net_ratelimit())
 127                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
 128                                        dev->name, hdr->gso_type);
 129                         goto frame_err;
 130                 }
 131
 132                 if (hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN)
 133                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 134
 135                 skb_shinfo(skb)->gso_size = hdr->gso_size;
 136                 if (skb_shinfo(skb)->gso_size == 0) {
 137                         if (net_ratelimit())
 138                                 printk(KERN_WARNING "%s: zero gso size.\n",
 139                                        dev->name);
 140                         goto frame_err;
 141                 }
 142
 143                 /* Header must be checked, and gso_segs computed. */
 144                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 145                 skb_shinfo(skb)->gso_segs = 0;
 146         }
 147
 148         netif_receive_skb(skb);
 149         return;
 150
 151 frame_err:
 152         dev->stats.rx_frame_errors++;
 153 drop:
 154         dev_kfree_skb(skb);
 155 }
 156
 157 static void try_fill_recv(struct virtnet_info *vi)
 158 {
 159         struct sk_buff *skb;
 160         struct scatterlist sg[2+MAX_SKB_FRAGS];
 161         int num, err;
 162
 163         sg_init_table(sg, 2+MAX_SKB_FRAGS);
 164         for (;;) {
 165                 skb = netdev_alloc_skb(vi->dev, MAX_PACKET_LEN);
 166                 if (unlikely(!skb))
 167                         break;
 168
 169                 skb_put(skb, MAX_PACKET_LEN);
 170                 vnet_hdr_to_sg(sg, skb);
 171                 num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 172                 skb_queue_head(&vi->recv, skb);
 173
 174                 err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, num, skb);
 175                 if (err) {
 176                         skb_unlink(skb, &vi->recv);
 177                         kfree_skb(skb);
 178                         break;
 179                 }
 180                 vi->num++;
 181         }
 182         if (unlikely(vi->num > vi->max))
 183                 vi->max = vi->num;
 184         vi->rvq->vq_ops->kick(vi->rvq);
 185 }
 186
 187 static void skb_recv_done(struct virtqueue *rvq)
 188 {
 189         struct virtnet_info *vi = rvq->vdev->priv;
 190         /* Schedule NAPI, Suppress further interrupts if successful. */
 191         if (netif_rx_schedule_prep(vi->dev, &vi->napi)) {
 192                 rvq->vq_ops->disable_cb(rvq);
 193                 __netif_rx_schedule(vi->dev, &vi->napi);
 194         }
 195 }
 196
 197 static int virtnet_poll(struct napi_struct *napi, int budget)
 198 {
 199         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
 200         struct sk_buff *skb = NULL;
 201         unsigned int len, received = 0;
 202
 203 again:
 204         while (received < budget &&
 205                (skb = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
 206                 __skb_unlink(skb, &vi->recv);
 207                 receive_skb(vi->dev, skb, len);
 208                 vi->num--;
 209                 received++;
 210         }
 211
 212         /* FIXME: If we oom and completely run out of inbufs, we need
 213          * to start a timer trying to fill more. */
 214         if (vi->num < vi->max / 2)
 215                 try_fill_recv(vi);
 216
 217         /* Out of packets? */
 218         if (received < budget) {
 219                 netif_rx_complete(vi->dev, napi);
 220                 if (unlikely(!vi->rvq->vq_ops->enable_cb(vi->rvq))
 221                     && napi_schedule_prep(napi)) {
 222                         vi->rvq->vq_ops->disable_cb(vi->rvq);
 223                         __netif_rx_schedule(vi->dev, napi);
 224                         goto again;
 225                 }
 226         }
 227
 228         return received;
 229 }
 230
 231 static void free_old_xmit_skbs(struct virtnet_info *vi)
 232 {
 233         struct sk_buff *skb;
 234         unsigned int len;
 235
 236         while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
 237                 pr_debug("Sent skb %p\n", skb);
 238                 __skb_unlink(skb, &vi->send);
 239                 vi->dev->stats.tx_bytes += skb->len;
 240                 vi->dev->stats.tx_packets++;
 241                 kfree_skb(skb);
 242         }
 243 }
 244
 245 static void xmit_free(unsigned long data)
 246 {
 247         struct virtnet_info *vi = (void *)data;
 248
 249         netif_tx_lock(vi->dev);
 250
 251         free_old_xmit_skbs(vi);
 252
 253         if (!skb_queue_empty(&vi->send))
 254                 mod_timer(&vi->xmit_free_timer, jiffies + (HZ/10));
 255
 256         netif_tx_unlock(vi->dev);
 257 }
 258
 259 static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 260 {
 261         int num, err;
 262         struct scatterlist sg[2+MAX_SKB_FRAGS];
 263         struct virtio_net_hdr *hdr;
 264         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 265
 266         sg_init_table(sg, 2+MAX_SKB_FRAGS);
 267
 268         pr_debug("%s: xmit %p " MAC_FMT "\n", vi->dev->name, skb,
 269                  dest[0], dest[1], dest[2],
 270                  dest[3], dest[4], dest[5]);
 271
 272         /* Encode metadata header at front. */
 273         hdr = skb_vnet_hdr(skb);
 274         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 275                 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 276                 hdr->csum_start = skb->csum_start - skb_headroom(skb);
 277                 hdr->csum_offset = skb->csum_offset;
 278         } else {
 279                 hdr->flags = 0;
 280                 hdr->csum_offset = hdr->csum_start = 0;
 281         }
 282
 283         if (skb_is_gso(skb)) {
 284                 hdr->hdr_len = skb_transport_header(skb) - skb->data;
 285                 hdr->gso_size = skb_shinfo(skb)->gso_size;
 286                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 287                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 288                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 289                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 290                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 291                         hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 292                 else
 293                         BUG();
 294                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 295                         hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 296         } else {
 297                 hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
 298                 hdr->gso_size = hdr->hdr_len = 0;
 299         }
 300
 301         vnet_hdr_to_sg(sg, skb);
 302         num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 303
 304         err = vi->svq->vq_ops->add_buf(vi->svq, sg, num, 0, skb);
 305         if (!err)
 306                 mod_timer(&vi->xmit_free_timer, jiffies + (HZ/10));
 307
 308         return err;
 309 }
 310
 311 static void xmit_tasklet(unsigned long data)
 312 {
 313         struct virtnet_info *vi = (void *)data;
 314
 315         netif_tx_lock_bh(vi->dev);
 316         if (vi->last_xmit_skb && xmit_skb(vi, vi->last_xmit_skb) == 0) {
 317                 vi->svq->vq_ops->kick(vi->svq);
 318                 vi->last_xmit_skb = NULL;
 319         }
 320         netif_tx_unlock_bh(vi->dev);
 321 }
 322
 323 static int start_xmit(struct sk_buff *skb, struct net_device *dev)
 324 {
 325         struct virtnet_info *vi = netdev_priv(dev);
 326
 327 again:
 328         /* Free up any pending old buffers before queueing new ones. */
 329         free_old_xmit_skbs(vi);
 330
 331         /* If we has a buffer left over from last time, send it now. */
 332         if (unlikely(vi->last_xmit_skb)) {
 333                 if (xmit_skb(vi, vi->last_xmit_skb) != 0) {
 334                         /* Drop this skb: we only queue one. */
 335                         vi->dev->stats.tx_dropped++;
 336                         kfree_skb(skb);
 337                         skb = NULL;
 338                         goto stop_queue;
 339                 }
 340                 vi->last_xmit_skb = NULL;
 341         }
 342
 343         /* Put new one in send queue and do transmit */
 344         if (likely(skb)) {
 345                 __skb_queue_head(&vi->send, skb);
 346                 if (xmit_skb(vi, skb) != 0) {
 347                         vi->last_xmit_skb = skb;
 348                         skb = NULL;
 349                         goto stop_queue;
 350                 }
 351         }
 352 done:
 353         vi->svq->vq_ops->kick(vi->svq);
 354         return NETDEV_TX_OK;
 355
 356 stop_queue:
 357         pr_debug("%s: virtio not prepared to send\n", dev->name);
 358         netif_stop_queue(dev);
 359
 360         /* Activate callback for using skbs: if this returns false it
 361          * means some were used in the meantime. */
 362         if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
 363                 vi->svq->vq_ops->disable_cb(vi->svq);
 364                 netif_start_queue(dev);
 365                 goto again;
 366         }
 367         goto done;
 368 }
 369
 370 #ifdef CONFIG_NET_POLL_CONTROLLER
 371 static void virtnet_netpoll(struct net_device *dev)
 372 {
 373         struct virtnet_info *vi = netdev_priv(dev);
 374
 375         napi_schedule(&vi->napi);
 376 }
 377 #endif
 378
 379 static int virtnet_open(struct net_device *dev)
 380 {
 381         struct virtnet_info *vi = netdev_priv(dev);
 382
 383         napi_enable(&vi->napi);
 384
 385         /* If all buffers were filled by other side before we napi_enabled, we
 386          * won't get another interrupt, so process any outstanding packets
 387          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 388          * We synchronize against interrupts via NAPI_STATE_SCHED */
 389         if (netif_rx_schedule_prep(dev, &vi->napi)) {
 390                 vi->rvq->vq_ops->disable_cb(vi->rvq);
 391                 __netif_rx_schedule(dev, &vi->napi);
 392         }
 393         return 0;
 394 }
 395
 396 static int virtnet_close(struct net_device *dev)
 397 {
 398         struct virtnet_info *vi = netdev_priv(dev);
 399
 400         napi_disable(&vi->napi);
 401
 402         return 0;
 403 }
 404
 405 static int virtnet_probe(struct virtio_device *vdev)
 406 {
 407         int err;
 408         struct net_device *dev;
 409         struct virtnet_info *vi;
 410
 411         /* Allocate ourselves a network device with room for our info */
 412         dev = alloc_etherdev(sizeof(struct virtnet_info));
 413         if (!dev)
 414                 return -ENOMEM;
 415
 416         /* Set up network device as normal. */
 417         dev->open = virtnet_open;
 418         dev->stop = virtnet_close;
 419         dev->hard_start_xmit = start_xmit;
 420         dev->features = NETIF_F_HIGHDMA;
 421 #ifdef CONFIG_NET_POLL_CONTROLLER
 422         dev->poll_controller = virtnet_netpoll;
 423 #endif
 424         SET_NETDEV_DEV(dev, &vdev->dev);
 425
 426         /* Do we support "hardware" checksums? */
 427         if (csum && virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
 428                 /* This opens up the world of extra features. */
 429                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
 430                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
 431                         dev->features |= NETIF_F_TSO | NETIF_F_UFO
 432                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
 433                 }
 434                 /* Individual feature bits: what can host handle? */
 435                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
 436                         dev->features |= NETIF_F_TSO;
 437                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
 438                         dev->features |= NETIF_F_TSO6;
 439                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
 440                         dev->features |= NETIF_F_TSO_ECN;
 441                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
 442                         dev->features |= NETIF_F_UFO;
 443         }
 444
 445         /* Configuration may specify what MAC to use.  Otherwise random. */
 446         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 447                 vdev->config->get(vdev,
 448                                   offsetof(struct virtio_net_config, mac),
 449                                   dev->dev_addr, dev->addr_len);
 450         } else
 451                 random_ether_addr(dev->dev_addr);
 452
 453         /* Set up our device-specific information */
 454         vi = netdev_priv(dev);
 455         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
 456         vi->dev = dev;
 457         vi->vdev = vdev;
 458         vdev->priv = vi;
 459
 460         /* We expect two virtqueues, receive then send. */
 461         vi->rvq = vdev->config->find_vq(vdev, 0, skb_recv_done);
 462         if (IS_ERR(vi->rvq)) {
 463                 err = PTR_ERR(vi->rvq);
 464                 goto free;
 465         }
 466
 467         vi->svq = vdev->config->find_vq(vdev, 1, skb_xmit_done);
 468         if (IS_ERR(vi->svq)) {
 469                 err = PTR_ERR(vi->svq);
 470                 goto free_recv;
 471         }
 472
 473         /* Initialize our empty receive and send queues. */
 474         skb_queue_head_init(&vi->recv);
 475         skb_queue_head_init(&vi->send);
 476
 477         tasklet_init(&vi->tasklet, xmit_tasklet, (unsigned long)vi);
 478
 479         setup_timer(&vi->xmit_free_timer, xmit_free, (unsigned long)vi);
 480
 481         err = register_netdev(dev);
 482         if (err) {
 483                 pr_debug("virtio_net: registering device failed\n");
 484                 goto free_send;
 485         }
 486
 487         /* Last of all, set up some receive buffers. */
 488         try_fill_recv(vi);
 489
 490         /* If we didn't even get one input buffer, we're useless. */
 491         if (vi->num == 0) {
 492                 err = -ENOMEM;
 493                 goto unregister;
 494         }
 495
 496         pr_debug("virtnet: registered device %s\n", dev->name);
 497         return 0;
 498
 499 unregister:
 500         unregister_netdev(dev);
 501 free_send:
 502         vdev->config->del_vq(vi->svq);
 503 free_recv:
 504         vdev->config->del_vq(vi->rvq);
 505 free:
 506         free_netdev(dev);
 507         return err;
 508 }
 509
 510 static void virtnet_remove(struct virtio_device *vdev)
 511 {
 512         struct virtnet_info *vi = vdev->priv;
 513         struct sk_buff *skb;
 514
 515         /* Stop all the virtqueues. */
 516         vdev->config->reset(vdev);
 517
 518         del_timer_sync(&vi->xmit_free_timer);
 519
 520         /* Free our skbs in send and recv queues, if any. */
 521         while ((skb = __skb_dequeue(&vi->recv)) != NULL) {
 522                 kfree_skb(skb);
 523                 vi->num--;
 524         }
 525         __skb_queue_purge(&vi->send);
 526
 527         BUG_ON(vi->num != 0);
 528
 529         vdev->config->del_vq(vi->svq);
 530         vdev->config->del_vq(vi->rvq);
 531         unregister_netdev(vi->dev);
 532         free_netdev(vi->dev);
 533 }
 534
 535 static struct virtio_device_id id_table[] = {
 536         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
 537         { 0 },
 538 };
 539
 540 static unsigned int features[] = {
 541         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
 542         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
 543         VIRTIO_NET_F_HOST_ECN,
 544 };
 545
 546 static struct virtio_driver virtio_net = {
 547         .feature_table = features,
 548         .feature_table_size = ARRAY_SIZE(features),
 549         .driver.name =  KBUILD_MODNAME,
 550         .driver.owner = THIS_MODULE,
 551         .id_table =     id_table,
 552         .probe =        virtnet_probe,
 553         .remove =       __devexit_p(virtnet_remove),
 554 };
 555
 556 static int __init init(void)
 557 {
 558         return register_virtio_driver(&virtio_net);
 559 }
 560
 561 static void __exit fini(void)
 562 {
 563         unregister_virtio_driver(&virtio_net);
 564 }
 565 module_init(init);
 566 module_exit(fini);
 567
 568 MODULE_DEVICE_TABLE(virtio, id_table);
 569 MODULE_DESCRIPTION("Virtio network driver");
 570 MODULE_LICENSE("GPL");