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/ethtool.h>
  23 #include <linux/module.h>
  24 #include <linux/virtio.h>
  25 #include <linux/virtio_net.h>
  26 #include <linux/scatterlist.h>
  27 #include <linux/if_vlan.h>
  28
  29 static int napi_weight = 128;
  30 module_param(napi_weight, int, 0444);
  31
  32 static int csum = 1, gso = 1;
  33 module_param(csum, bool, 0444);
  34 module_param(gso, bool, 0444);
  35
  36 /* FIXME: MTU in config. */
  37 #define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  38 #define GOOD_COPY_LEN   128
  39
  40 #define VIRTNET_SEND_COMMAND_SG_MAX    2
  41
  42 struct virtnet_info
  43 {
  44         struct virtio_device *vdev;
  45         struct virtqueue *rvq, *svq, *cvq;
  46         struct net_device *dev;
  47         struct napi_struct napi;
  48         unsigned int status;
  49
  50         /* The skb we couldn't send because buffers were full. */
  51         struct sk_buff *last_xmit_skb;
  52
  53         /* If we need to free in a timer, this is it. */
  54         struct timer_list xmit_free_timer;
  55
  56         /* Number of input buffers, and max we've ever had. */
  57         unsigned int num, max;
  58
  59         /* For cleaning up after transmission. */
  60         struct tasklet_struct tasklet;
  61         bool free_in_tasklet;
  62
  63         /* I like... big packets and I cannot lie! */
  64         bool big_packets;
  65
  66         /* Host will merge rx buffers for big packets (shake it! shake it!) */
  67         bool mergeable_rx_bufs;
  68
  69         /* Receive & send queues. */
  70         struct sk_buff_head recv;
  71         struct sk_buff_head send;
  72
  73         /* Work struct for refilling if we run low on memory. */
  74         struct delayed_work refill;
  75
  76         /* Chain pages by the private ptr. */
  77         struct page *pages;
  78 };
  79
  80 static inline void *skb_vnet_hdr(struct sk_buff *skb)
  81 {
  82         return (struct virtio_net_hdr *)skb->cb;
  83 }
  84
  85 static void give_a_page(struct virtnet_info *vi, struct page *page)
  86 {
  87         page->private = (unsigned long)vi->pages;
  88         vi->pages = page;
  89 }
  90
  91 static void trim_pages(struct virtnet_info *vi, struct sk_buff *skb)
  92 {
  93         unsigned int i;
  94
  95         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
  96                 give_a_page(vi, skb_shinfo(skb)->frags[i].page);
  97         skb_shinfo(skb)->nr_frags = 0;
  98         skb->data_len = 0;
  99 }
 100
 101 static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
 102 {
 103         struct page *p = vi->pages;
 104
 105         if (p)
 106                 vi->pages = (struct page *)p->private;
 107         else
 108                 p = alloc_page(gfp_mask);
 109         return p;
 110 }
 111
 112 static void skb_xmit_done(struct virtqueue *svq)
 113 {
 114         struct virtnet_info *vi = svq->vdev->priv;
 115
 116         /* Suppress further interrupts. */
 117         svq->vq_ops->disable_cb(svq);
 118
 119         /* We were probably waiting for more output buffers. */
 120         netif_wake_queue(vi->dev);
 121
 122         /* Make sure we re-xmit last_xmit_skb: if there are no more packets
 123          * queued, start_xmit won't be called. */
 124         tasklet_schedule(&vi->tasklet);
 125 }
 126
 127 static void receive_skb(struct net_device *dev, struct sk_buff *skb,
 128                         unsigned len)
 129 {
 130         struct virtnet_info *vi = netdev_priv(dev);
 131         struct virtio_net_hdr *hdr = skb_vnet_hdr(skb);
 132         int err;
 133         int i;
 134
 135         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 136                 pr_debug("%s: short packet %i\n", dev->name, len);
 137                 dev->stats.rx_length_errors++;
 138                 goto drop;
 139         }
 140
 141         if (vi->mergeable_rx_bufs) {
 142                 struct virtio_net_hdr_mrg_rxbuf *mhdr = skb_vnet_hdr(skb);
 143                 unsigned int copy;
 144                 char *p = page_address(skb_shinfo(skb)->frags[0].page);
 145
 146                 if (len > PAGE_SIZE)
 147                         len = PAGE_SIZE;
 148                 len -= sizeof(struct virtio_net_hdr_mrg_rxbuf);
 149
 150                 memcpy(hdr, p, sizeof(*mhdr));
 151                 p += sizeof(*mhdr);
 152
 153                 copy = len;
 154                 if (copy > skb_tailroom(skb))
 155                         copy = skb_tailroom(skb);
 156
 157                 memcpy(skb_put(skb, copy), p, copy);
 158
 159                 len -= copy;
 160
 161                 if (!len) {
 162                         give_a_page(vi, skb_shinfo(skb)->frags[0].page);
 163                         skb_shinfo(skb)->nr_frags--;
 164                 } else {
 165                         skb_shinfo(skb)->frags[0].page_offset +=
 166                                 sizeof(*mhdr) + copy;
 167                         skb_shinfo(skb)->frags[0].size = len;
 168                         skb->data_len += len;
 169                         skb->len += len;
 170                 }
 171
 172                 while (--mhdr->num_buffers) {
 173                         struct sk_buff *nskb;
 174
 175                         i = skb_shinfo(skb)->nr_frags;
 176                         if (i >= MAX_SKB_FRAGS) {
 177                                 pr_debug("%s: packet too long %d\n", dev->name,
 178                                          len);
 179                                 dev->stats.rx_length_errors++;
 180                                 goto drop;
 181                         }
 182
 183                         nskb = vi->rvq->vq_ops->get_buf(vi->rvq, &len);
 184                         if (!nskb) {
 185                                 pr_debug("%s: rx error: %d buffers missing\n",
 186                                          dev->name, mhdr->num_buffers);
 187                                 dev->stats.rx_length_errors++;
 188                                 goto drop;
 189                         }
 190
 191                         __skb_unlink(nskb, &vi->recv);
 192                         vi->num--;
 193
 194                         skb_shinfo(skb)->frags[i] = skb_shinfo(nskb)->frags[0];
 195                         skb_shinfo(nskb)->nr_frags = 0;
 196                         kfree_skb(nskb);
 197
 198                         if (len > PAGE_SIZE)
 199                                 len = PAGE_SIZE;
 200
 201                         skb_shinfo(skb)->frags[i].size = len;
 202                         skb_shinfo(skb)->nr_frags++;
 203                         skb->data_len += len;
 204                         skb->len += len;
 205                 }
 206         } else {
 207                 len -= sizeof(struct virtio_net_hdr);
 208
 209                 if (len <= MAX_PACKET_LEN)
 210                         trim_pages(vi, skb);
 211
 212                 err = pskb_trim(skb, len);
 213                 if (err) {
 214                         pr_debug("%s: pskb_trim failed %i %d\n", dev->name,
 215                                  len, err);
 216                         dev->stats.rx_dropped++;
 217                         goto drop;
 218                 }
 219         }
 220
 221         skb->truesize += skb->data_len;
 222         dev->stats.rx_bytes += skb->len;
 223         dev->stats.rx_packets++;
 224
 225         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 226                 pr_debug("Needs csum!\n");
 227                 if (!skb_partial_csum_set(skb,hdr->csum_start,hdr->csum_offset))
 228                         goto frame_err;
 229         }
 230
 231         skb->protocol = eth_type_trans(skb, dev);
 232         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 233                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 234
 235         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 236                 pr_debug("GSO!\n");
 237                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 238                 case VIRTIO_NET_HDR_GSO_TCPV4:
 239                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 240                         break;
 241                 case VIRTIO_NET_HDR_GSO_UDP:
 242                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 243                         break;
 244                 case VIRTIO_NET_HDR_GSO_TCPV6:
 245                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 246                         break;
 247                 default:
 248                         if (net_ratelimit())
 249                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
 250                                        dev->name, hdr->gso_type);
 251                         goto frame_err;
 252                 }
 253
 254                 if (hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN)
 255                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 256
 257                 skb_shinfo(skb)->gso_size = hdr->gso_size;
 258                 if (skb_shinfo(skb)->gso_size == 0) {
 259                         if (net_ratelimit())
 260                                 printk(KERN_WARNING "%s: zero gso size.\n",
 261                                        dev->name);
 262                         goto frame_err;
 263                 }
 264
 265                 /* Header must be checked, and gso_segs computed. */
 266                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 267                 skb_shinfo(skb)->gso_segs = 0;
 268         }
 269
 270         netif_receive_skb(skb);
 271         return;
 272
 273 frame_err:
 274         dev->stats.rx_frame_errors++;
 275 drop:
 276         dev_kfree_skb(skb);
 277 }
 278
 279 static bool try_fill_recv_maxbufs(struct virtnet_info *vi, gfp_t gfp)
 280 {
 281         struct sk_buff *skb;
 282         struct scatterlist sg[2+MAX_SKB_FRAGS];
 283         int num, err, i;
 284         bool oom = false;
 285
 286         sg_init_table(sg, 2+MAX_SKB_FRAGS);
 287         for (;;) {
 288                 struct virtio_net_hdr *hdr;
 289
 290                 skb = netdev_alloc_skb(vi->dev, MAX_PACKET_LEN + NET_IP_ALIGN);
 291                 if (unlikely(!skb)) {
 292                         oom = true;
 293                         break;
 294                 }
 295
 296                 skb_reserve(skb, NET_IP_ALIGN);
 297                 skb_put(skb, MAX_PACKET_LEN);
 298
 299                 hdr = skb_vnet_hdr(skb);
 300                 sg_set_buf(sg, hdr, sizeof(*hdr));
 301
 302                 if (vi->big_packets) {
 303                         for (i = 0; i < MAX_SKB_FRAGS; i++) {
 304                                 skb_frag_t *f = &skb_shinfo(skb)->frags[i];
 305                                 f->page = get_a_page(vi, gfp);
 306                                 if (!f->page)
 307                                         break;
 308
 309                                 f->page_offset = 0;
 310                                 f->size = PAGE_SIZE;
 311
 312                                 skb->data_len += PAGE_SIZE;
 313                                 skb->len += PAGE_SIZE;
 314
 315                                 skb_shinfo(skb)->nr_frags++;
 316                         }
 317                 }
 318
 319                 num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 320                 skb_queue_head(&vi->recv, skb);
 321
 322                 err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, num, skb);
 323                 if (err) {
 324                         skb_unlink(skb, &vi->recv);
 325                         trim_pages(vi, skb);
 326                         kfree_skb(skb);
 327                         break;
 328                 }
 329                 vi->num++;
 330         }
 331         if (unlikely(vi->num > vi->max))
 332                 vi->max = vi->num;
 333         vi->rvq->vq_ops->kick(vi->rvq);
 334         return !oom;
 335 }
 336
 337 /* Returns false if we couldn't fill entirely (OOM). */
 338 static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
 339 {
 340         struct sk_buff *skb;
 341         struct scatterlist sg[1];
 342         int err;
 343         bool oom = false;
 344
 345         if (!vi->mergeable_rx_bufs)
 346                 return try_fill_recv_maxbufs(vi, gfp);
 347
 348         for (;;) {
 349                 skb_frag_t *f;
 350
 351                 skb = netdev_alloc_skb(vi->dev, GOOD_COPY_LEN + NET_IP_ALIGN);
 352                 if (unlikely(!skb)) {
 353                         oom = true;
 354                         break;
 355                 }
 356
 357                 skb_reserve(skb, NET_IP_ALIGN);
 358
 359                 f = &skb_shinfo(skb)->frags[0];
 360                 f->page = get_a_page(vi, gfp);
 361                 if (!f->page) {
 362                         oom = true;
 363                         kfree_skb(skb);
 364                         break;
 365                 }
 366
 367                 f->page_offset = 0;
 368                 f->size = PAGE_SIZE;
 369
 370                 skb_shinfo(skb)->nr_frags++;
 371
 372                 sg_init_one(sg, page_address(f->page), PAGE_SIZE);
 373                 skb_queue_head(&vi->recv, skb);
 374
 375                 err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, 1, skb);
 376                 if (err) {
 377                         skb_unlink(skb, &vi->recv);
 378                         kfree_skb(skb);
 379                         break;
 380                 }
 381                 vi->num++;
 382         }
 383         if (unlikely(vi->num > vi->max))
 384                 vi->max = vi->num;
 385         vi->rvq->vq_ops->kick(vi->rvq);
 386         return !oom;
 387 }
 388
 389 static void skb_recv_done(struct virtqueue *rvq)
 390 {
 391         struct virtnet_info *vi = rvq->vdev->priv;
 392         /* Schedule NAPI, Suppress further interrupts if successful. */
 393         if (napi_schedule_prep(&vi->napi)) {
 394                 rvq->vq_ops->disable_cb(rvq);
 395                 __napi_schedule(&vi->napi);
 396         }
 397 }
 398
 399 static void refill_work(struct work_struct *work)
 400 {
 401         struct virtnet_info *vi;
 402         bool still_empty;
 403
 404         vi = container_of(work, struct virtnet_info, refill.work);
 405         napi_disable(&vi->napi);
 406         try_fill_recv(vi, GFP_KERNEL);
 407         still_empty = (vi->num == 0);
 408         napi_enable(&vi->napi);
 409
 410         /* In theory, this can happen: if we don't get any buffers in
 411          * we will *never* try to fill again. */
 412         if (still_empty)
 413                 schedule_delayed_work(&vi->refill, HZ/2);
 414 }
 415
 416 static int virtnet_poll(struct napi_struct *napi, int budget)
 417 {
 418         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
 419         struct sk_buff *skb = NULL;
 420         unsigned int len, received = 0;
 421
 422 again:
 423         while (received < budget &&
 424                (skb = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
 425                 __skb_unlink(skb, &vi->recv);
 426                 receive_skb(vi->dev, skb, len);
 427                 vi->num--;
 428                 received++;
 429         }
 430
 431         if (vi->num < vi->max / 2) {
 432                 if (!try_fill_recv(vi, GFP_ATOMIC))
 433                         schedule_delayed_work(&vi->refill, 0);
 434         }
 435
 436         /* Out of packets? */
 437         if (received < budget) {
 438                 napi_complete(napi);
 439                 if (unlikely(!vi->rvq->vq_ops->enable_cb(vi->rvq))
 440                     && napi_schedule_prep(napi)) {
 441                         vi->rvq->vq_ops->disable_cb(vi->rvq);
 442                         __napi_schedule(napi);
 443                         goto again;
 444                 }
 445         }
 446
 447         return received;
 448 }
 449
 450 static void free_old_xmit_skbs(struct virtnet_info *vi)
 451 {
 452         struct sk_buff *skb;
 453         unsigned int len;
 454
 455         while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
 456                 pr_debug("Sent skb %p\n", skb);
 457                 __skb_unlink(skb, &vi->send);
 458                 vi->dev->stats.tx_bytes += skb->len;
 459                 vi->dev->stats.tx_packets++;
 460                 kfree_skb(skb);
 461         }
 462 }
 463
 464 /* If the virtio transport doesn't always notify us when all in-flight packets
 465  * are consumed, we fall back to using this function on a timer to free them. */
 466 static void xmit_free(unsigned long data)
 467 {
 468         struct virtnet_info *vi = (void *)data;
 469
 470         netif_tx_lock(vi->dev);
 471
 472         free_old_xmit_skbs(vi);
 473
 474         if (!skb_queue_empty(&vi->send))
 475                 mod_timer(&vi->xmit_free_timer, jiffies + (HZ/10));
 476
 477         netif_tx_unlock(vi->dev);
 478 }
 479
 480 static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 481 {
 482         int num, err;
 483         struct scatterlist sg[2+MAX_SKB_FRAGS];
 484         struct virtio_net_hdr_mrg_rxbuf *mhdr = skb_vnet_hdr(skb);
 485         struct virtio_net_hdr *hdr = skb_vnet_hdr(skb);
 486         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 487
 488         sg_init_table(sg, 2+MAX_SKB_FRAGS);
 489
 490         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 491
 492         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 493                 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 494                 hdr->csum_start = skb->csum_start - skb_headroom(skb);
 495                 hdr->csum_offset = skb->csum_offset;
 496         } else {
 497                 hdr->flags = 0;
 498                 hdr->csum_offset = hdr->csum_start = 0;
 499         }
 500
 501         if (skb_is_gso(skb)) {
 502                 hdr->hdr_len = skb_headlen(skb);
 503                 hdr->gso_size = skb_shinfo(skb)->gso_size;
 504                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 505                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 506                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 507                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 508                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 509                         hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 510                 else
 511                         BUG();
 512                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 513                         hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 514         } else {
 515                 hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
 516                 hdr->gso_size = hdr->hdr_len = 0;
 517         }
 518
 519         mhdr->num_buffers = 0;
 520
 521         /* Encode metadata header at front. */
 522         if (vi->mergeable_rx_bufs)
 523                 sg_set_buf(sg, mhdr, sizeof(*mhdr));
 524         else
 525                 sg_set_buf(sg, hdr, sizeof(*hdr));
 526
 527         num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 528
 529         err = vi->svq->vq_ops->add_buf(vi->svq, sg, num, 0, skb);
 530         if (!err && !vi->free_in_tasklet)
 531                 mod_timer(&vi->xmit_free_timer, jiffies + (HZ/10));
 532
 533         return err;
 534 }
 535
 536 static void xmit_tasklet(unsigned long data)
 537 {
 538         struct virtnet_info *vi = (void *)data;
 539
 540         netif_tx_lock_bh(vi->dev);
 541         if (vi->last_xmit_skb && xmit_skb(vi, vi->last_xmit_skb) == 0) {
 542                 vi->svq->vq_ops->kick(vi->svq);
 543                 vi->last_xmit_skb = NULL;
 544         }
 545         if (vi->free_in_tasklet)
 546                 free_old_xmit_skbs(vi);
 547         netif_tx_unlock_bh(vi->dev);
 548 }
 549
 550 static int start_xmit(struct sk_buff *skb, struct net_device *dev)
 551 {
 552         struct virtnet_info *vi = netdev_priv(dev);
 553
 554 again:
 555         /* Free up any pending old buffers before queueing new ones. */
 556         free_old_xmit_skbs(vi);
 557
 558         /* If we has a buffer left over from last time, send it now. */
 559         if (unlikely(vi->last_xmit_skb) &&
 560             xmit_skb(vi, vi->last_xmit_skb) != 0)
 561                 goto stop_queue;
 562
 563         vi->last_xmit_skb = NULL;
 564
 565         /* Put new one in send queue and do transmit */
 566         if (likely(skb)) {
 567                 __skb_queue_head(&vi->send, skb);
 568                 if (xmit_skb(vi, skb) != 0) {
 569                         vi->last_xmit_skb = skb;
 570                         skb = NULL;
 571                         goto stop_queue;
 572                 }
 573         }
 574 done:
 575         vi->svq->vq_ops->kick(vi->svq);
 576         return NETDEV_TX_OK;
 577
 578 stop_queue:
 579         pr_debug("%s: virtio not prepared to send\n", dev->name);
 580         netif_stop_queue(dev);
 581
 582         /* Activate callback for using skbs: if this returns false it
 583          * means some were used in the meantime. */
 584         if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
 585                 vi->svq->vq_ops->disable_cb(vi->svq);
 586                 netif_start_queue(dev);
 587                 goto again;
 588         }
 589         if (skb) {
 590                 /* Drop this skb: we only queue one. */
 591                 vi->dev->stats.tx_dropped++;
 592                 kfree_skb(skb);
 593         }
 594         goto done;
 595 }
 596
 597 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 598 {
 599         struct virtnet_info *vi = netdev_priv(dev);
 600         struct virtio_device *vdev = vi->vdev;
 601         int ret;
 602
 603         ret = eth_mac_addr(dev, p);
 604         if (ret)
 605                 return ret;
 606
 607         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
 608                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 609                                   dev->dev_addr, dev->addr_len);
 610
 611         return 0;
 612 }
 613
 614 #ifdef CONFIG_NET_POLL_CONTROLLER
 615 static void virtnet_netpoll(struct net_device *dev)
 616 {
 617         struct virtnet_info *vi = netdev_priv(dev);
 618
 619         napi_schedule(&vi->napi);
 620 }
 621 #endif
 622
 623 static int virtnet_open(struct net_device *dev)
 624 {
 625         struct virtnet_info *vi = netdev_priv(dev);
 626
 627         napi_enable(&vi->napi);
 628
 629         /* If all buffers were filled by other side before we napi_enabled, we
 630          * won't get another interrupt, so process any outstanding packets
 631          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 632          * We synchronize against interrupts via NAPI_STATE_SCHED */
 633         if (napi_schedule_prep(&vi->napi)) {
 634                 vi->rvq->vq_ops->disable_cb(vi->rvq);
 635                 __napi_schedule(&vi->napi);
 636         }
 637         return 0;
 638 }
 639
 640 /*
 641  * Send command via the control virtqueue and check status.  Commands
 642  * supported by the hypervisor, as indicated by feature bits, should
 643  * never fail unless improperly formated.
 644  */
 645 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 646                                  struct scatterlist *data, int out, int in)
 647 {
 648         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
 649         struct virtio_net_ctrl_hdr ctrl;
 650         virtio_net_ctrl_ack status = ~0;
 651         unsigned int tmp;
 652         int i;
 653
 654         /* Caller should know better */
 655         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
 656                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
 657
 658         out++; /* Add header */
 659         in++; /* Add return status */
 660
 661         ctrl.class = class;
 662         ctrl.cmd = cmd;
 663
 664         sg_init_table(sg, out + in);
 665
 666         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
 667         for_each_sg(data, s, out + in - 2, i)
 668                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
 669         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
 670
 671         BUG_ON(vi->cvq->vq_ops->add_buf(vi->cvq, sg, out, in, vi));
 672
 673         vi->cvq->vq_ops->kick(vi->cvq);
 674
 675         /*
 676          * Spin for a response, the kick causes an ioport write, trapping
 677          * into the hypervisor, so the request should be handled immediately.
 678          */
 679         while (!vi->cvq->vq_ops->get_buf(vi->cvq, &tmp))
 680                 cpu_relax();
 681
 682         return status == VIRTIO_NET_OK;
 683 }
 684
 685 static int virtnet_close(struct net_device *dev)
 686 {
 687         struct virtnet_info *vi = netdev_priv(dev);
 688
 689         napi_disable(&vi->napi);
 690
 691         return 0;
 692 }
 693
 694 static int virtnet_set_tx_csum(struct net_device *dev, u32 data)
 695 {
 696         struct virtnet_info *vi = netdev_priv(dev);
 697         struct virtio_device *vdev = vi->vdev;
 698
 699         if (data && !virtio_has_feature(vdev, VIRTIO_NET_F_CSUM))
 700                 return -ENOSYS;
 701
 702         return ethtool_op_set_tx_hw_csum(dev, data);
 703 }
 704
 705 static void virtnet_set_rx_mode(struct net_device *dev)
 706 {
 707         struct virtnet_info *vi = netdev_priv(dev);
 708         struct scatterlist sg[2];
 709         u8 promisc, allmulti;
 710         struct virtio_net_ctrl_mac *mac_data;
 711         struct dev_addr_list *addr;
 712         struct netdev_hw_addr *ha;
 713         void *buf;
 714         int i;
 715
 716         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 717         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 718                 return;
 719
 720         promisc = ((dev->flags & IFF_PROMISC) != 0);
 721         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 722
 723         sg_init_one(sg, &promisc, sizeof(promisc));
 724
 725         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 726                                   VIRTIO_NET_CTRL_RX_PROMISC,
 727                                   sg, 1, 0))
 728                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 729                          promisc ? "en" : "dis");
 730
 731         sg_init_one(sg, &allmulti, sizeof(allmulti));
 732
 733         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 734                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 735                                   sg, 1, 0))
 736                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
 737                          allmulti ? "en" : "dis");
 738
 739         /* MAC filter - use one buffer for both lists */
 740         mac_data = buf = kzalloc(((dev->uc.count + dev->mc_count) * ETH_ALEN) +
 741                                  (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
 742         if (!buf) {
 743                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
 744                 return;
 745         }
 746
 747         sg_init_table(sg, 2);
 748
 749         /* Store the unicast list and count in the front of the buffer */
 750         mac_data->entries = dev->uc.count;
 751         i = 0;
 752         list_for_each_entry(ha, &dev->uc.list, list)
 753                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 754
 755         sg_set_buf(&sg[0], mac_data,
 756                    sizeof(mac_data->entries) + (dev->uc.count * ETH_ALEN));
 757
 758         /* multicast list and count fill the end */
 759         mac_data = (void *)&mac_data->macs[dev->uc.count][0];
 760
 761         mac_data->entries = dev->mc_count;
 762         addr = dev->mc_list;
 763         for (i = 0; i < dev->mc_count; i++, addr = addr->next)
 764                 memcpy(&mac_data->macs[i][0], addr->da_addr, ETH_ALEN);
 765
 766         sg_set_buf(&sg[1], mac_data,
 767                    sizeof(mac_data->entries) + (dev->mc_count * ETH_ALEN));
 768
 769         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 770                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
 771                                   sg, 2, 0))
 772                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
 773
 774         kfree(buf);
 775 }
 776
 777 static void virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
 778 {
 779         struct virtnet_info *vi = netdev_priv(dev);
 780         struct scatterlist sg;
 781
 782         sg_init_one(&sg, &vid, sizeof(vid));
 783
 784         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 785                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
 786                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
 787 }
 788
 789 static void virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
 790 {
 791         struct virtnet_info *vi = netdev_priv(dev);
 792         struct scatterlist sg;
 793
 794         sg_init_one(&sg, &vid, sizeof(vid));
 795
 796         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 797                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
 798                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
 799 }
 800
 801 static struct ethtool_ops virtnet_ethtool_ops = {
 802         .set_tx_csum = virtnet_set_tx_csum,
 803         .set_sg = ethtool_op_set_sg,
 804         .set_tso = ethtool_op_set_tso,
 805         .get_link = ethtool_op_get_link,
 806 };
 807
 808 #define MIN_MTU 68
 809 #define MAX_MTU 65535
 810
 811 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
 812 {
 813         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
 814                 return -EINVAL;
 815         dev->mtu = new_mtu;
 816         return 0;
 817 }
 818
 819 static const struct net_device_ops virtnet_netdev = {
 820         .ndo_open            = virtnet_open,
 821         .ndo_stop            = virtnet_close,
 822         .ndo_start_xmit      = start_xmit,
 823         .ndo_validate_addr   = eth_validate_addr,
 824         .ndo_set_mac_address = virtnet_set_mac_address,
 825         .ndo_set_rx_mode     = virtnet_set_rx_mode,
 826         .ndo_change_mtu      = virtnet_change_mtu,
 827         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
 828         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
 829 #ifdef CONFIG_NET_POLL_CONTROLLER
 830         .ndo_poll_controller = virtnet_netpoll,
 831 #endif
 832 };
 833
 834 static void virtnet_update_status(struct virtnet_info *vi)
 835 {
 836         u16 v;
 837
 838         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS))
 839                 return;
 840
 841         vi->vdev->config->get(vi->vdev,
 842                               offsetof(struct virtio_net_config, status),
 843                               &v, sizeof(v));
 844
 845         /* Ignore unknown (future) status bits */
 846         v &= VIRTIO_NET_S_LINK_UP;
 847
 848         if (vi->status == v)
 849                 return;
 850
 851         vi->status = v;
 852
 853         if (vi->status & VIRTIO_NET_S_LINK_UP) {
 854                 netif_carrier_on(vi->dev);
 855                 netif_wake_queue(vi->dev);
 856         } else {
 857                 netif_carrier_off(vi->dev);
 858                 netif_stop_queue(vi->dev);
 859         }
 860 }
 861
 862 static void virtnet_config_changed(struct virtio_device *vdev)
 863 {
 864         struct virtnet_info *vi = vdev->priv;
 865
 866         virtnet_update_status(vi);
 867 }
 868
 869 static int virtnet_probe(struct virtio_device *vdev)
 870 {
 871         int err;
 872         struct net_device *dev;
 873         struct virtnet_info *vi;
 874         struct virtqueue *vqs[3];
 875         vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
 876         const char *names[] = { "input", "output", "control" };
 877         int nvqs;
 878
 879         /* Allocate ourselves a network device with room for our info */
 880         dev = alloc_etherdev(sizeof(struct virtnet_info));
 881         if (!dev)
 882                 return -ENOMEM;
 883
 884         /* Set up network device as normal. */
 885         dev->netdev_ops = &virtnet_netdev;
 886         dev->features = NETIF_F_HIGHDMA;
 887         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
 888         SET_NETDEV_DEV(dev, &vdev->dev);
 889
 890         /* Do we support "hardware" checksums? */
 891         if (csum && virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
 892                 /* This opens up the world of extra features. */
 893                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
 894                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
 895                         dev->features |= NETIF_F_TSO | NETIF_F_UFO
 896                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
 897                 }
 898                 /* Individual feature bits: what can host handle? */
 899                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
 900                         dev->features |= NETIF_F_TSO;
 901                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
 902                         dev->features |= NETIF_F_TSO6;
 903                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
 904                         dev->features |= NETIF_F_TSO_ECN;
 905                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
 906                         dev->features |= NETIF_F_UFO;
 907         }
 908
 909         /* Configuration may specify what MAC to use.  Otherwise random. */
 910         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 911                 vdev->config->get(vdev,
 912                                   offsetof(struct virtio_net_config, mac),
 913                                   dev->dev_addr, dev->addr_len);
 914         } else
 915                 random_ether_addr(dev->dev_addr);
 916
 917         /* Set up our device-specific information */
 918         vi = netdev_priv(dev);
 919         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
 920         vi->dev = dev;
 921         vi->vdev = vdev;
 922         vdev->priv = vi;
 923         vi->pages = NULL;
 924         INIT_DELAYED_WORK(&vi->refill, refill_work);
 925
 926         /* If they give us a callback when all buffers are done, we don't need
 927          * the timer. */
 928         vi->free_in_tasklet = virtio_has_feature(vdev,VIRTIO_F_NOTIFY_ON_EMPTY);
 929
 930         /* If we can receive ANY GSO packets, we must allocate large ones. */
 931         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4)
 932             || virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6)
 933             || virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
 934                 vi->big_packets = true;
 935
 936         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
 937                 vi->mergeable_rx_bufs = true;
 938
 939         /* We expect two virtqueues, receive then send,
 940          * and optionally control. */
 941         nvqs = virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ? 3 : 2;
 942
 943         err = vdev->config->find_vqs(vdev, nvqs, vqs, callbacks, names);
 944         if (err)
 945                 goto free;
 946
 947         vi->rvq = vqs[0];
 948         vi->svq = vqs[1];
 949
 950         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)) {
 951                 vi->cvq = vqs[2];
 952
 953                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
 954                         dev->features |= NETIF_F_HW_VLAN_FILTER;
 955         }
 956
 957         /* Initialize our empty receive and send queues. */
 958         skb_queue_head_init(&vi->recv);
 959         skb_queue_head_init(&vi->send);
 960
 961         tasklet_init(&vi->tasklet, xmit_tasklet, (unsigned long)vi);
 962
 963         if (!vi->free_in_tasklet)
 964                 setup_timer(&vi->xmit_free_timer, xmit_free, (unsigned long)vi);
 965
 966         err = register_netdev(dev);
 967         if (err) {
 968                 pr_debug("virtio_net: registering device failed\n");
 969                 goto free_vqs;
 970         }
 971
 972         /* Last of all, set up some receive buffers. */
 973         try_fill_recv(vi, GFP_KERNEL);
 974
 975         /* If we didn't even get one input buffer, we're useless. */
 976         if (vi->num == 0) {
 977                 err = -ENOMEM;
 978                 goto unregister;
 979         }
 980
 981         vi->status = VIRTIO_NET_S_LINK_UP;
 982         virtnet_update_status(vi);
 983         netif_carrier_on(dev);
 984
 985         pr_debug("virtnet: registered device %s\n", dev->name);
 986         return 0;
 987
 988 unregister:
 989         unregister_netdev(dev);
 990         cancel_delayed_work_sync(&vi->refill);
 991 free_vqs:
 992         vdev->config->del_vqs(vdev);
 993 free:
 994         free_netdev(dev);
 995         return err;
 996 }
 997
 998 static void virtnet_remove(struct virtio_device *vdev)
 999 {
1000         struct virtnet_info *vi = vdev->priv;
1001         struct sk_buff *skb;
1002
1003         /* Stop all the virtqueues. */
1004         vdev->config->reset(vdev);
1005
1006         if (!vi->free_in_tasklet)
1007                 del_timer_sync(&vi->xmit_free_timer);
1008
1009         /* Free our skbs in send and recv queues, if any. */
1010         while ((skb = __skb_dequeue(&vi->recv)) != NULL) {
1011                 kfree_skb(skb);
1012                 vi->num--;
1013         }
1014         __skb_queue_purge(&vi->send);
1015
1016         BUG_ON(vi->num != 0);
1017
1018         unregister_netdev(vi->dev);
1019         cancel_delayed_work_sync(&vi->refill);
1020
1021         vdev->config->del_vqs(vi->vdev);
1022
1023         while (vi->pages)
1024                 __free_pages(get_a_page(vi, GFP_KERNEL), 0);
1025
1026         free_netdev(vi->dev);
1027 }
1028
1029 static struct virtio_device_id id_table[] = {
1030         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1031         { 0 },
1032 };
1033
1034 static unsigned int features[] = {
1035         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1036         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1037         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1038         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1039         VIRTIO_NET_F_GUEST_ECN, /* We don't yet handle UFO input. */
1040         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1041         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1042         VIRTIO_F_NOTIFY_ON_EMPTY,
1043 };
1044
1045 static struct virtio_driver virtio_net = {
1046         .feature_table = features,
1047         .feature_table_size = ARRAY_SIZE(features),
1048         .driver.name =  KBUILD_MODNAME,
1049         .driver.owner = THIS_MODULE,
1050         .id_table =     id_table,
1051         .probe =        virtnet_probe,
1052         .remove =       __devexit_p(virtnet_remove),
1053         .config_changed = virtnet_config_changed,
1054 };
1055
1056 static int __init init(void)
1057 {
1058         return register_virtio_driver(&virtio_net);
1059 }
1060
1061 static void __exit fini(void)
1062 {
1063         unregister_virtio_driver(&virtio_net);
1064 }
1065 module_init(init);
1066 module_exit(fini);
1067
1068 MODULE_DEVICE_TABLE(virtio, id_table);
1069 MODULE_DESCRIPTION("Virtio network driver");
1070 MODULE_LICENSE("GPL");