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 /* FIXME: MTU in config. */
  28 #define MAX_PACKET_LEN (ETH_HLEN+ETH_DATA_LEN)
  29
  30 struct virtnet_info
  31 {
  32         struct virtio_device *vdev;
  33         struct virtqueue *rvq, *svq;
  34         struct net_device *dev;
  35         struct napi_struct napi;
  36
  37         /* Number of input buffers, and max we've ever had. */
  38         unsigned int num, max;
  39
  40         /* Receive & send queues. */
  41         struct sk_buff_head recv;
  42         struct sk_buff_head send;
  43 };
  44
  45 static inline struct virtio_net_hdr *skb_vnet_hdr(struct sk_buff *skb)
  46 {
  47         return (struct virtio_net_hdr *)skb->cb;
  48 }
  49
  50 static inline void vnet_hdr_to_sg(struct scatterlist *sg, struct sk_buff *skb)
  51 {
  52         sg_init_one(sg, skb_vnet_hdr(skb), sizeof(struct virtio_net_hdr));
  53 }
  54
  55 static bool skb_xmit_done(struct virtqueue *rvq)
  56 {
  57         struct virtnet_info *vi = rvq->vdev->priv;
  58
  59         /* In case we were waiting for output buffers. */
  60         netif_wake_queue(vi->dev);
  61         return true;
  62 }
  63
  64 static void receive_skb(struct net_device *dev, struct sk_buff *skb,
  65                         unsigned len)
  66 {
  67         struct virtio_net_hdr *hdr = skb_vnet_hdr(skb);
  68
  69         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
  70                 pr_debug("%s: short packet %i\n", dev->name, len);
  71                 dev->stats.rx_length_errors++;
  72                 goto drop;
  73         }
  74         len -= sizeof(struct virtio_net_hdr);
  75         BUG_ON(len > MAX_PACKET_LEN);
  76
  77         skb_trim(skb, len);
  78         skb->protocol = eth_type_trans(skb, dev);
  79         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
  80                  ntohs(skb->protocol), skb->len, skb->pkt_type);
  81         dev->stats.rx_bytes += skb->len;
  82         dev->stats.rx_packets++;
  83
  84         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
  85                 pr_debug("Needs csum!\n");
  86                 if (!skb_partial_csum_set(skb,hdr->csum_start,hdr->csum_offset))
  87                         goto frame_err;
  88         }
  89
  90         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
  91                 pr_debug("GSO!\n");
  92                 switch (hdr->gso_type) {
  93                 case VIRTIO_NET_HDR_GSO_TCPV4:
  94                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
  95                         break;
  96                 case VIRTIO_NET_HDR_GSO_TCPV4_ECN:
  97                         skb_shinfo(skb)->gso_type = SKB_GSO_TCP_ECN;
  98                         break;
  99                 case VIRTIO_NET_HDR_GSO_UDP:
 100                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 101                         break;
 102                 case VIRTIO_NET_HDR_GSO_TCPV6:
 103                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 104                         break;
 105                 default:
 106                         if (net_ratelimit())
 107                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
 108                                        dev->name, hdr->gso_type);
 109                         goto frame_err;
 110                 }
 111
 112                 skb_shinfo(skb)->gso_size = hdr->gso_size;
 113                 if (skb_shinfo(skb)->gso_size == 0) {
 114                         if (net_ratelimit())
 115                                 printk(KERN_WARNING "%s: zero gso size.\n",
 116                                        dev->name);
 117                         goto frame_err;
 118                 }
 119
 120                 /* Header must be checked, and gso_segs computed. */
 121                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 122                 skb_shinfo(skb)->gso_segs = 0;
 123         }
 124
 125         netif_receive_skb(skb);
 126         return;
 127
 128 frame_err:
 129         dev->stats.rx_frame_errors++;
 130 drop:
 131         dev_kfree_skb(skb);
 132 }
 133
 134 static void try_fill_recv(struct virtnet_info *vi)
 135 {
 136         struct sk_buff *skb;
 137         struct scatterlist sg[1+MAX_SKB_FRAGS];
 138         int num, err;
 139
 140         sg_init_table(sg, 1+MAX_SKB_FRAGS);
 141         for (;;) {
 142                 skb = netdev_alloc_skb(vi->dev, MAX_PACKET_LEN);
 143                 if (unlikely(!skb))
 144                         break;
 145
 146                 skb_put(skb, MAX_PACKET_LEN);
 147                 vnet_hdr_to_sg(sg, skb);
 148                 num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 149                 skb_queue_head(&vi->recv, skb);
 150
 151                 err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, num, skb);
 152                 if (err) {
 153                         skb_unlink(skb, &vi->recv);
 154                         kfree_skb(skb);
 155                         break;
 156                 }
 157                 vi->num++;
 158         }
 159         if (unlikely(vi->num > vi->max))
 160                 vi->max = vi->num;
 161         vi->rvq->vq_ops->kick(vi->rvq);
 162 }
 163
 164 static bool skb_recv_done(struct virtqueue *rvq)
 165 {
 166         struct virtnet_info *vi = rvq->vdev->priv;
 167         netif_rx_schedule(vi->dev, &vi->napi);
 168         /* Suppress further interrupts. */
 169         return false;
 170 }
 171
 172 static int virtnet_poll(struct napi_struct *napi, int budget)
 173 {
 174         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
 175         struct sk_buff *skb = NULL;
 176         unsigned int len, received = 0;
 177
 178 again:
 179         while (received < budget &&
 180                (skb = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
 181                 __skb_unlink(skb, &vi->recv);
 182                 receive_skb(vi->dev, skb, len);
 183                 vi->num--;
 184                 received++;
 185         }
 186
 187         /* FIXME: If we oom and completely run out of inbufs, we need
 188          * to start a timer trying to fill more. */
 189         if (vi->num < vi->max / 2)
 190                 try_fill_recv(vi);
 191
 192         /* Out of packets? */
 193         if (received < budget) {
 194                 netif_rx_complete(vi->dev, napi);
 195                 if (unlikely(!vi->rvq->vq_ops->restart(vi->rvq))
 196                     && netif_rx_reschedule(vi->dev, napi))
 197                         goto again;
 198         }
 199
 200         return received;
 201 }
 202
 203 static void free_old_xmit_skbs(struct virtnet_info *vi)
 204 {
 205         struct sk_buff *skb;
 206         unsigned int len;
 207
 208         while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
 209                 pr_debug("Sent skb %p\n", skb);
 210                 __skb_unlink(skb, &vi->send);
 211                 vi->dev->stats.tx_bytes += len;
 212                 vi->dev->stats.tx_packets++;
 213                 kfree_skb(skb);
 214         }
 215 }
 216
 217 static int start_xmit(struct sk_buff *skb, struct net_device *dev)
 218 {
 219         struct virtnet_info *vi = netdev_priv(dev);
 220         int num, err;
 221         struct scatterlist sg[1+MAX_SKB_FRAGS];
 222         struct virtio_net_hdr *hdr;
 223         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 224         DECLARE_MAC_BUF(mac);
 225
 226         sg_init_table(sg, 1+MAX_SKB_FRAGS);
 227
 228         pr_debug("%s: xmit %p %s\n", dev->name, skb, print_mac(mac, dest));
 229
 230         free_old_xmit_skbs(vi);
 231
 232         /* Encode metadata header at front. */
 233         hdr = skb_vnet_hdr(skb);
 234         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 235                 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 236                 hdr->csum_start = skb->csum_start - skb_headroom(skb);
 237                 hdr->csum_offset = skb->csum_offset;
 238         } else {
 239                 hdr->flags = 0;
 240                 hdr->csum_offset = hdr->csum_start = 0;
 241         }
 242
 243         if (skb_is_gso(skb)) {
 244                 hdr->gso_size = skb_shinfo(skb)->gso_size;
 245                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 246                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4_ECN;
 247                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 248                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 249                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 250                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 251                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 252                         hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 253                 else
 254                         BUG();
 255         } else {
 256                 hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
 257                 hdr->gso_size = 0;
 258         }
 259
 260         vnet_hdr_to_sg(sg, skb);
 261         num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 262         __skb_queue_head(&vi->send, skb);
 263         err = vi->svq->vq_ops->add_buf(vi->svq, sg, num, 0, skb);
 264         if (err) {
 265                 pr_debug("%s: virtio not prepared to send\n", dev->name);
 266                 skb_unlink(skb, &vi->send);
 267                 netif_stop_queue(dev);
 268                 return NETDEV_TX_BUSY;
 269         }
 270         vi->svq->vq_ops->kick(vi->svq);
 271
 272         return 0;
 273 }
 274
 275 static int virtnet_open(struct net_device *dev)
 276 {
 277         struct virtnet_info *vi = netdev_priv(dev);
 278
 279         try_fill_recv(vi);
 280
 281         /* If we didn't even get one input buffer, we're useless. */
 282         if (vi->num == 0)
 283                 return -ENOMEM;
 284
 285         napi_enable(&vi->napi);
 286         return 0;
 287 }
 288
 289 static int virtnet_close(struct net_device *dev)
 290 {
 291         struct virtnet_info *vi = netdev_priv(dev);
 292         struct sk_buff *skb;
 293
 294         napi_disable(&vi->napi);
 295
 296         /* networking core has neutered skb_xmit_done/skb_recv_done, so don't
 297          * worry about races vs. get(). */
 298         vi->rvq->vq_ops->shutdown(vi->rvq);
 299         while ((skb = __skb_dequeue(&vi->recv)) != NULL) {
 300                 kfree_skb(skb);
 301                 vi->num--;
 302         }
 303         vi->svq->vq_ops->shutdown(vi->svq);
 304         while ((skb = __skb_dequeue(&vi->send)) != NULL)
 305                 kfree_skb(skb);
 306
 307         BUG_ON(vi->num != 0);
 308         return 0;
 309 }
 310
 311 static int virtnet_probe(struct virtio_device *vdev)
 312 {
 313         int err;
 314         struct net_device *dev;
 315         struct virtnet_info *vi;
 316
 317         /* Allocate ourselves a network device with room for our info */
 318         dev = alloc_etherdev(sizeof(struct virtnet_info));
 319         if (!dev)
 320                 return -ENOMEM;
 321
 322         /* Set up network device as normal. */
 323         ether_setup(dev);
 324         dev->open = virtnet_open;
 325         dev->stop = virtnet_close;
 326         dev->hard_start_xmit = start_xmit;
 327         dev->features = NETIF_F_HIGHDMA;
 328         SET_NETDEV_DEV(dev, &vdev->dev);
 329
 330         /* Do we support "hardware" checksums? */
 331         if (vdev->config->feature(vdev, VIRTIO_NET_F_NO_CSUM)) {
 332                 /* This opens up the world of extra features. */
 333                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
 334                 if (vdev->config->feature(vdev, VIRTIO_NET_F_TSO4))
 335                         dev->features |= NETIF_F_TSO;
 336                 if (vdev->config->feature(vdev, VIRTIO_NET_F_UFO))
 337                         dev->features |= NETIF_F_UFO;
 338                 if (vdev->config->feature(vdev, VIRTIO_NET_F_TSO4_ECN))
 339                         dev->features |= NETIF_F_TSO_ECN;
 340                 if (vdev->config->feature(vdev, VIRTIO_NET_F_TSO6))
 341                         dev->features |= NETIF_F_TSO6;
 342         }
 343
 344         /* Configuration may specify what MAC to use.  Otherwise random. */
 345         if (vdev->config->feature(vdev, VIRTIO_NET_F_MAC)) {
 346                 vdev->config->get(vdev,
 347                                   offsetof(struct virtio_net_config, mac),
 348                                   dev->dev_addr, dev->addr_len);
 349         } else
 350                 random_ether_addr(dev->dev_addr);
 351
 352         /* Set up our device-specific information */
 353         vi = netdev_priv(dev);
 354         netif_napi_add(dev, &vi->napi, virtnet_poll, 16);
 355         vi->dev = dev;
 356         vi->vdev = vdev;
 357
 358         /* We expect two virtqueues, receive then send. */
 359         vi->rvq = vdev->config->find_vq(vdev, 0, skb_recv_done);
 360         if (IS_ERR(vi->rvq)) {
 361                 err = PTR_ERR(vi->rvq);
 362                 goto free;
 363         }
 364
 365         vi->svq = vdev->config->find_vq(vdev, 1, skb_xmit_done);
 366         if (IS_ERR(vi->svq)) {
 367                 err = PTR_ERR(vi->svq);
 368                 goto free_recv;
 369         }
 370
 371         /* Initialize our empty receive and send queues. */
 372         skb_queue_head_init(&vi->recv);
 373         skb_queue_head_init(&vi->send);
 374
 375         err = register_netdev(dev);
 376         if (err) {
 377                 pr_debug("virtio_net: registering device failed\n");
 378                 goto free_send;
 379         }
 380         pr_debug("virtnet: registered device %s\n", dev->name);
 381         vdev->priv = vi;
 382         return 0;
 383
 384 free_send:
 385         vdev->config->del_vq(vi->svq);
 386 free_recv:
 387         vdev->config->del_vq(vi->rvq);
 388 free:
 389         free_netdev(dev);
 390         return err;
 391 }
 392
 393 static void virtnet_remove(struct virtio_device *vdev)
 394 {
 395         struct virtnet_info *vi = vdev->priv;
 396
 397         vdev->config->del_vq(vi->svq);
 398         vdev->config->del_vq(vi->rvq);
 399         unregister_netdev(vi->dev);
 400         free_netdev(vi->dev);
 401 }
 402
 403 static struct virtio_device_id id_table[] = {
 404         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
 405         { 0 },
 406 };
 407
 408 static struct virtio_driver virtio_net = {
 409         .driver.name =  KBUILD_MODNAME,
 410         .driver.owner = THIS_MODULE,
 411         .id_table =     id_table,
 412         .probe =        virtnet_probe,
 413         .remove =       __devexit_p(virtnet_remove),
 414 };
 415
 416 static int __init init(void)
 417 {
 418         return register_virtio_driver(&virtio_net);
 419 }
 420
 421 static void __exit fini(void)
 422 {
 423         unregister_virtio_driver(&virtio_net);
 424 }
 425 module_init(init);
 426 module_exit(fini);
 427
 428 MODULE_DEVICE_TABLE(virtio, id_table);
 429 MODULE_DESCRIPTION("Virtio network driver");
 430 MODULE_LICENSE("GPL");