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