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