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_ids.h>
  26 #include <linux/virtio_net.h>
  27 #include <linux/scatterlist.h>
  28 #include <linux/if_vlan.h>
  29
  30 static int napi_weight = 128;
  31 module_param(napi_weight, int, 0444);
  32
  33 static int csum = 1, gso = 1;
  34 module_param(csum, bool, 0444);
  35 module_param(gso, bool, 0444);
  36
  37 /* FIXME: MTU in config. */
  38 #define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  39 #define GOOD_COPY_LEN   128
  40
  41 #define VIRTNET_SEND_COMMAND_SG_MAX    2
  42
  43 struct virtnet_info
  44 {
  45         struct virtio_device *vdev;
  46         struct virtqueue *rvq, *svq, *cvq;
  47         struct net_device *dev;
  48         struct napi_struct napi;
  49         unsigned int status;
  50
  51         /* The skb we couldn't send because buffers were full. */
  52         struct sk_buff *last_xmit_skb;
  53
  54         /* If we need to free in a timer, this is it. */
  55         struct timer_list xmit_free_timer;
  56
  57         /* Number of input buffers, and max we've ever had. */
  58         unsigned int num, max;
  59
  60         /* For cleaning up after transmission. */
  61         struct tasklet_struct tasklet;
  62         bool free_in_tasklet;
  63
  64         /* I like... big packets and I cannot lie! */
  65         bool big_packets;
  66
  67         /* Host will merge rx buffers for big packets (shake it! shake it!) */
  68         bool mergeable_rx_bufs;
  69
  70         /* Receive & send queues. */
  71         struct sk_buff_head recv;
  72         struct sk_buff_head send;
  73
  74         /* Work struct for refilling if we run low on memory. */
  75         struct delayed_work refill;
  76
  77         /* Chain pages by the private ptr. */
  78         struct page *pages;
  79 };
  80
  81 static inline void *skb_vnet_hdr(struct sk_buff *skb)
  82 {
  83         return (struct virtio_net_hdr *)skb->cb;
  84 }
  85
  86 static void give_a_page(struct virtnet_info *vi, struct page *page)
  87 {
  88         page->private = (unsigned long)vi->pages;
  89         vi->pages = page;
  90 }
  91
  92 static void trim_pages(struct virtnet_info *vi, struct sk_buff *skb)
  93 {
  94         unsigned int i;
  95
  96         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
  97                 give_a_page(vi, skb_shinfo(skb)->frags[i].page);
  98         skb_shinfo(skb)->nr_frags = 0;
  99         skb->data_len = 0;
 100 }
 101
 102 static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
 103 {
 104         struct page *p = vi->pages;
 105
 106         if (p)
 107                 vi->pages = (struct page *)p->private;
 108         else
 109                 p = alloc_page(gfp_mask);
 110         return p;
 111 }
 112
 113 static void skb_xmit_done(struct virtqueue *svq)
 114 {
 115         struct virtnet_info *vi = svq->vdev->priv;
 116
 117         /* Suppress further interrupts. */
 118         svq->vq_ops->disable_cb(svq);
 119
 120         /* We were probably waiting for more output buffers. */
 121         netif_wake_queue(vi->dev);
 122
 123         /* Make sure we re-xmit last_xmit_skb: if there are no more packets
 124          * queued, start_xmit won't be called. */
 125         tasklet_schedule(&vi->tasklet);
 126 }
 127
 128 static void receive_skb(struct net_device *dev, struct sk_buff *skb,
 129                         unsigned len)
 130 {
 131         struct virtnet_info *vi = netdev_priv(dev);
 132         struct virtio_net_hdr *hdr = skb_vnet_hdr(skb);
 133         int err;
 134         int i;
 135
 136         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 137                 pr_debug("%s: short packet %i\n", dev->name, len);
 138                 dev->stats.rx_length_errors++;
 139                 goto drop;
 140         }
 141
 142         if (vi->mergeable_rx_bufs) {
 143                 struct virtio_net_hdr_mrg_rxbuf *mhdr = skb_vnet_hdr(skb);
 144                 unsigned int copy;
 145                 char *p = page_address(skb_shinfo(skb)->frags[0].page);
 146
 147                 if (len > PAGE_SIZE)
 148                         len = PAGE_SIZE;
 149                 len -= sizeof(struct virtio_net_hdr_mrg_rxbuf);
 150
 151                 memcpy(hdr, p, sizeof(*mhdr));
 152                 p += sizeof(*mhdr);
 153
 154                 copy = len;
 155                 if (copy > skb_tailroom(skb))
 156                         copy = skb_tailroom(skb);
 157
 158                 memcpy(skb_put(skb, copy), p, copy);
 159
 160                 len -= copy;
 161
 162                 if (!len) {
 163                         give_a_page(vi, skb_shinfo(skb)->frags[0].page);
 164                         skb_shinfo(skb)->nr_frags--;
 165                 } else {
 166                         skb_shinfo(skb)->frags[0].page_offset +=
 167                                 sizeof(*mhdr) + copy;
 168                         skb_shinfo(skb)->frags[0].size = len;
 169                         skb->data_len += len;
 170                         skb->len += len;
 171                 }
 172
 173                 while (--mhdr->num_buffers) {
 174                         struct sk_buff *nskb;
 175
 176                         i = skb_shinfo(skb)->nr_frags;
 177                         if (i >= MAX_SKB_FRAGS) {
 178                                 pr_debug("%s: packet too long %d\n", dev->name,
 179                                          len);
 180                                 dev->stats.rx_length_errors++;
 181                                 goto drop;
 182                         }
 183
 184                         nskb = vi->rvq->vq_ops->get_buf(vi->rvq, &len);
 185                         if (!nskb) {
 186                                 pr_debug("%s: rx error: %d buffers missing\n",
 187                                          dev->name, mhdr->num_buffers);
 188                                 dev->stats.rx_length_errors++;
 189                                 goto drop;
 190                         }
 191
 192                         __skb_unlink(nskb, &vi->recv);
 193                         vi->num--;
 194
 195                         skb_shinfo(skb)->frags[i] = skb_shinfo(nskb)->frags[0];
 196                         skb_shinfo(nskb)->nr_frags = 0;
 197                         kfree_skb(nskb);
 198
 199                         if (len > PAGE_SIZE)
 200                                 len = PAGE_SIZE;
 201
 202                         skb_shinfo(skb)->frags[i].size = len;
 203                         skb_shinfo(skb)->nr_frags++;
 204                         skb->data_len += len;
 205                         skb->len += len;
 206                 }
 207         } else {
 208                 len -= sizeof(struct virtio_net_hdr);
 209
 210                 if (len <= MAX_PACKET_LEN)
 211                         trim_pages(vi, skb);
 212
 213                 err = pskb_trim(skb, len);
 214                 if (err) {
 215                         pr_debug("%s: pskb_trim failed %i %d\n", dev->name,
 216                                  len, err);
 217                         dev->stats.rx_dropped++;
 218                         goto drop;
 219                 }
 220         }
 221
 222         skb->truesize += skb->data_len;
 223         dev->stats.rx_bytes += skb->len;
 224         dev->stats.rx_packets++;
 225
 226         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 227                 pr_debug("Needs csum!\n");
 228                 if (!skb_partial_csum_set(skb,hdr->csum_start,hdr->csum_offset))
 229                         goto frame_err;
 230         }
 231
 232         skb->protocol = eth_type_trans(skb, dev);
 233         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 234                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 235
 236         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 237                 pr_debug("GSO!\n");
 238                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 239                 case VIRTIO_NET_HDR_GSO_TCPV4:
 240                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 241                         break;
 242                 case VIRTIO_NET_HDR_GSO_UDP:
 243                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 244                         break;
 245                 case VIRTIO_NET_HDR_GSO_TCPV6:
 246                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 247                         break;
 248                 default:
 249                         if (net_ratelimit())
 250                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
 251                                        dev->name, hdr->gso_type);
 252                         goto frame_err;
 253                 }
 254
 255                 if (hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN)
 256                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 257
 258                 skb_shinfo(skb)->gso_size = hdr->gso_size;
 259                 if (skb_shinfo(skb)->gso_size == 0) {
 260                         if (net_ratelimit())
 261                                 printk(KERN_WARNING "%s: zero gso size.\n",
 262                                        dev->name);
 263                         goto frame_err;
 264                 }
 265
 266                 /* Header must be checked, and gso_segs computed. */
 267                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 268                 skb_shinfo(skb)->gso_segs = 0;
 269         }
 270
 271         netif_receive_skb(skb);
 272         return;
 273
 274 frame_err:
 275         dev->stats.rx_frame_errors++;
 276 drop:
 277         dev_kfree_skb(skb);
 278 }
 279
 280 static bool try_fill_recv_maxbufs(struct virtnet_info *vi, gfp_t gfp)
 281 {
 282         struct sk_buff *skb;
 283         struct scatterlist sg[2+MAX_SKB_FRAGS];
 284         int num, err, i;
 285         bool oom = false;
 286
 287         sg_init_table(sg, 2+MAX_SKB_FRAGS);
 288         for (;;) {
 289                 struct virtio_net_hdr *hdr;
 290
 291                 skb = netdev_alloc_skb(vi->dev, MAX_PACKET_LEN + NET_IP_ALIGN);
 292                 if (unlikely(!skb)) {
 293                         oom = true;
 294                         break;
 295                 }
 296
 297                 skb_reserve(skb, NET_IP_ALIGN);
 298                 skb_put(skb, MAX_PACKET_LEN);
 299
 300                 hdr = skb_vnet_hdr(skb);
 301                 sg_set_buf(sg, hdr, sizeof(*hdr));
 302
 303                 if (vi->big_packets) {
 304                         for (i = 0; i < MAX_SKB_FRAGS; i++) {
 305                                 skb_frag_t *f = &skb_shinfo(skb)->frags[i];
 306                                 f->page = get_a_page(vi, gfp);
 307                                 if (!f->page)
 308                                         break;
 309
 310                                 f->page_offset = 0;
 311                                 f->size = PAGE_SIZE;
 312
 313                                 skb->data_len += PAGE_SIZE;
 314                                 skb->len += PAGE_SIZE;
 315
 316                                 skb_shinfo(skb)->nr_frags++;
 317                         }
 318                 }
 319
 320                 num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 321                 skb_queue_head(&vi->recv, skb);
 322
 323                 err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, num, skb);
 324                 if (err < 0) {
 325                         skb_unlink(skb, &vi->recv);
 326                         trim_pages(vi, skb);
 327                         kfree_skb(skb);
 328                         break;
 329                 }
 330                 vi->num++;
 331         }
 332         if (unlikely(vi->num > vi->max))
 333                 vi->max = vi->num;
 334         vi->rvq->vq_ops->kick(vi->rvq);
 335         return !oom;
 336 }
 337
 338 /* Returns false if we couldn't fill entirely (OOM). */
 339 static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
 340 {
 341         struct sk_buff *skb;
 342         struct scatterlist sg[1];
 343         int err;
 344         bool oom = false;
 345
 346         if (!vi->mergeable_rx_bufs)
 347                 return try_fill_recv_maxbufs(vi, gfp);
 348
 349         for (;;) {
 350                 skb_frag_t *f;
 351
 352                 skb = netdev_alloc_skb(vi->dev, GOOD_COPY_LEN + NET_IP_ALIGN);
 353                 if (unlikely(!skb)) {
 354                         oom = true;
 355                         break;
 356                 }
 357
 358                 skb_reserve(skb, NET_IP_ALIGN);
 359
 360                 f = &skb_shinfo(skb)->frags[0];
 361                 f->page = get_a_page(vi, gfp);
 362                 if (!f->page) {
 363                         oom = true;
 364                         kfree_skb(skb);
 365                         break;
 366                 }
 367
 368                 f->page_offset = 0;
 369                 f->size = PAGE_SIZE;
 370
 371                 skb_shinfo(skb)->nr_frags++;
 372
 373                 sg_init_one(sg, page_address(f->page), PAGE_SIZE);
 374                 skb_queue_head(&vi->recv, skb);
 375
 376                 err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, 1, skb);
 377                 if (err < 0) {
 378                         skb_unlink(skb, &vi->recv);
 379                         kfree_skb(skb);
 380                         break;
 381                 }
 382                 vi->num++;
 383         }
 384         if (unlikely(vi->num > vi->max))
 385                 vi->max = vi->num;
 386         vi->rvq->vq_ops->kick(vi->rvq);
 387         return !oom;
 388 }
 389
 390 static void skb_recv_done(struct virtqueue *rvq)
 391 {
 392         struct virtnet_info *vi = rvq->vdev->priv;
 393         /* Schedule NAPI, Suppress further interrupts if successful. */
 394         if (napi_schedule_prep(&vi->napi)) {
 395                 rvq->vq_ops->disable_cb(rvq);
 396                 __napi_schedule(&vi->napi);
 397         }
 398 }
 399
 400 static void refill_work(struct work_struct *work)
 401 {
 402         struct virtnet_info *vi;
 403         bool still_empty;
 404
 405         vi = container_of(work, struct virtnet_info, refill.work);
 406         napi_disable(&vi->napi);
 407         try_fill_recv(vi, GFP_KERNEL);
 408         still_empty = (vi->num == 0);
 409         napi_enable(&vi->napi);
 410
 411         /* In theory, this can happen: if we don't get any buffers in
 412          * we will *never* try to fill again. */
 413         if (still_empty)
 414                 schedule_delayed_work(&vi->refill, HZ/2);
 415 }
 416
 417 static int virtnet_poll(struct napi_struct *napi, int budget)
 418 {
 419         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
 420         struct sk_buff *skb = NULL;
 421         unsigned int len, received = 0;
 422
 423 again:
 424         while (received < budget &&
 425                (skb = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
 426                 __skb_unlink(skb, &vi->recv);
 427                 receive_skb(vi->dev, skb, len);
 428                 vi->num--;
 429                 received++;
 430         }
 431
 432         if (vi->num < vi->max / 2) {
 433                 if (!try_fill_recv(vi, GFP_ATOMIC))
 434                         schedule_delayed_work(&vi->refill, 0);
 435         }
 436
 437         /* Out of packets? */
 438         if (received < budget) {
 439                 napi_complete(napi);
 440                 if (unlikely(!vi->rvq->vq_ops->enable_cb(vi->rvq))
 441                     && napi_schedule_prep(napi)) {
 442                         vi->rvq->vq_ops->disable_cb(vi->rvq);
 443                         __napi_schedule(napi);
 444                         goto again;
 445                 }
 446         }
 447
 448         return received;
 449 }
 450
 451 static void free_old_xmit_skbs(struct virtnet_info *vi)
 452 {
 453         struct sk_buff *skb;
 454         unsigned int len;
 455
 456         while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
 457                 pr_debug("Sent skb %p\n", skb);
 458                 __skb_unlink(skb, &vi->send);
 459                 vi->dev->stats.tx_bytes += skb->len;
 460                 vi->dev->stats.tx_packets++;
 461                 kfree_skb(skb);
 462         }
 463 }
 464
 465 /* If the virtio transport doesn't always notify us when all in-flight packets
 466  * are consumed, we fall back to using this function on a timer to free them. */
 467 static void xmit_free(unsigned long data)
 468 {
 469         struct virtnet_info *vi = (void *)data;
 470
 471         netif_tx_lock(vi->dev);
 472
 473         free_old_xmit_skbs(vi);
 474
 475         if (!skb_queue_empty(&vi->send))
 476                 mod_timer(&vi->xmit_free_timer, jiffies + (HZ/10));
 477
 478         netif_tx_unlock(vi->dev);
 479 }
 480
 481 static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 482 {
 483         int num, err;
 484         struct scatterlist sg[2+MAX_SKB_FRAGS];
 485         struct virtio_net_hdr_mrg_rxbuf *mhdr = skb_vnet_hdr(skb);
 486         struct virtio_net_hdr *hdr = skb_vnet_hdr(skb);
 487         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 488
 489         sg_init_table(sg, 2+MAX_SKB_FRAGS);
 490
 491         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 492
 493         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 494                 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 495                 hdr->csum_start = skb->csum_start - skb_headroom(skb);
 496                 hdr->csum_offset = skb->csum_offset;
 497         } else {
 498                 hdr->flags = 0;
 499                 hdr->csum_offset = hdr->csum_start = 0;
 500         }
 501
 502         if (skb_is_gso(skb)) {
 503                 hdr->hdr_len = skb_headlen(skb);
 504                 hdr->gso_size = skb_shinfo(skb)->gso_size;
 505                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 506                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 507                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 508                         hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 509                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 510                         hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 511                 else
 512                         BUG();
 513                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 514                         hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 515         } else {
 516                 hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
 517                 hdr->gso_size = hdr->hdr_len = 0;
 518         }
 519
 520         mhdr->num_buffers = 0;
 521
 522         /* Encode metadata header at front. */
 523         if (vi->mergeable_rx_bufs)
 524                 sg_set_buf(sg, mhdr, sizeof(*mhdr));
 525         else
 526                 sg_set_buf(sg, hdr, sizeof(*hdr));
 527
 528         num = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 529
 530         err = vi->svq->vq_ops->add_buf(vi->svq, sg, num, 0, skb);
 531         if (err >= 0 && !vi->free_in_tasklet)
 532                 mod_timer(&vi->xmit_free_timer, jiffies + (HZ/10));
 533
 534         return err;
 535 }
 536
 537 static void xmit_tasklet(unsigned long data)
 538 {
 539         struct virtnet_info *vi = (void *)data;
 540
 541         netif_tx_lock_bh(vi->dev);
 542         if (vi->last_xmit_skb && xmit_skb(vi, vi->last_xmit_skb) >= 0) {
 543                 vi->svq->vq_ops->kick(vi->svq);
 544                 vi->last_xmit_skb = NULL;
 545         }
 546         if (vi->free_in_tasklet)
 547                 free_old_xmit_skbs(vi);
 548         netif_tx_unlock_bh(vi->dev);
 549 }
 550
 551 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 552 {
 553         struct virtnet_info *vi = netdev_priv(dev);
 554
 555 again:
 556         /* Free up any pending old buffers before queueing new ones. */
 557         free_old_xmit_skbs(vi);
 558
 559         /* If we has a buffer left over from last time, send it now. */
 560         if (unlikely(vi->last_xmit_skb) &&
 561             xmit_skb(vi, vi->last_xmit_skb) < 0)
 562                 goto stop_queue;
 563
 564         vi->last_xmit_skb = NULL;
 565
 566         /* Put new one in send queue and do transmit */
 567         if (likely(skb)) {
 568                 __skb_queue_head(&vi->send, skb);
 569                 if (xmit_skb(vi, skb) < 0) {
 570                         vi->last_xmit_skb = skb;
 571                         skb = NULL;
 572                         goto stop_queue;
 573                 }
 574         }
 575 done:
 576         vi->svq->vq_ops->kick(vi->svq);
 577         return NETDEV_TX_OK;
 578
 579 stop_queue:
 580         pr_debug("%s: virtio not prepared to send\n", dev->name);
 581         netif_stop_queue(dev);
 582
 583         /* Activate callback for using skbs: if this returns false it
 584          * means some were used in the meantime. */
 585         if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
 586                 vi->svq->vq_ops->disable_cb(vi->svq);
 587                 netif_start_queue(dev);
 588                 goto again;
 589         }
 590         if (skb) {
 591                 /* Drop this skb: we only queue one. */
 592                 vi->dev->stats.tx_dropped++;
 593                 kfree_skb(skb);
 594         }
 595         goto done;
 596 }
 597
 598 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 599 {
 600         struct virtnet_info *vi = netdev_priv(dev);
 601         struct virtio_device *vdev = vi->vdev;
 602         int ret;
 603
 604         ret = eth_mac_addr(dev, p);
 605         if (ret)
 606                 return ret;
 607
 608         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
 609                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 610                                   dev->dev_addr, dev->addr_len);
 611
 612         return 0;
 613 }
 614
 615 #ifdef CONFIG_NET_POLL_CONTROLLER
 616 static void virtnet_netpoll(struct net_device *dev)
 617 {
 618         struct virtnet_info *vi = netdev_priv(dev);
 619
 620         napi_schedule(&vi->napi);
 621 }
 622 #endif
 623
 624 static int virtnet_open(struct net_device *dev)
 625 {
 626         struct virtnet_info *vi = netdev_priv(dev);
 627
 628         napi_enable(&vi->napi);
 629
 630         /* If all buffers were filled by other side before we napi_enabled, we
 631          * won't get another interrupt, so process any outstanding packets
 632          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 633          * We synchronize against interrupts via NAPI_STATE_SCHED */
 634         if (napi_schedule_prep(&vi->napi)) {
 635                 vi->rvq->vq_ops->disable_cb(vi->rvq);
 636                 __napi_schedule(&vi->napi);
 637         }
 638         return 0;
 639 }
 640
 641 /*
 642  * Send command via the control virtqueue and check status.  Commands
 643  * supported by the hypervisor, as indicated by feature bits, should
 644  * never fail unless improperly formated.
 645  */
 646 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 647                                  struct scatterlist *data, int out, int in)
 648 {
 649         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
 650         struct virtio_net_ctrl_hdr ctrl;
 651         virtio_net_ctrl_ack status = ~0;
 652         unsigned int tmp;
 653         int i;
 654
 655         /* Caller should know better */
 656         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
 657                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
 658
 659         out++; /* Add header */
 660         in++; /* Add return status */
 661
 662         ctrl.class = class;
 663         ctrl.cmd = cmd;
 664
 665         sg_init_table(sg, out + in);
 666
 667         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
 668         for_each_sg(data, s, out + in - 2, i)
 669                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
 670         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
 671
 672         BUG_ON(vi->cvq->vq_ops->add_buf(vi->cvq, sg, out, in, vi) < 0);
 673
 674         vi->cvq->vq_ops->kick(vi->cvq);
 675
 676         /*
 677          * Spin for a response, the kick causes an ioport write, trapping
 678          * into the hypervisor, so the request should be handled immediately.
 679          */
 680         while (!vi->cvq->vq_ops->get_buf(vi->cvq, &tmp))
 681                 cpu_relax();
 682
 683         return status == VIRTIO_NET_OK;
 684 }
 685
 686 static int virtnet_close(struct net_device *dev)
 687 {
 688         struct virtnet_info *vi = netdev_priv(dev);
 689
 690         napi_disable(&vi->napi);
 691
 692         return 0;
 693 }
 694
 695 static int virtnet_set_tx_csum(struct net_device *dev, u32 data)
 696 {
 697         struct virtnet_info *vi = netdev_priv(dev);
 698         struct virtio_device *vdev = vi->vdev;
 699
 700         if (data && !virtio_has_feature(vdev, VIRTIO_NET_F_CSUM))
 701                 return -ENOSYS;
 702
 703         return ethtool_op_set_tx_hw_csum(dev, data);
 704 }
 705
 706 static void virtnet_set_rx_mode(struct net_device *dev)
 707 {
 708         struct virtnet_info *vi = netdev_priv(dev);
 709         struct scatterlist sg[2];
 710         u8 promisc, allmulti;
 711         struct virtio_net_ctrl_mac *mac_data;
 712         struct dev_addr_list *addr;
 713         struct netdev_hw_addr *ha;
 714         void *buf;
 715         int i;
 716
 717         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 718         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 719                 return;
 720
 721         promisc = ((dev->flags & IFF_PROMISC) != 0);
 722         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 723
 724         sg_init_one(sg, &promisc, sizeof(promisc));
 725
 726         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 727                                   VIRTIO_NET_CTRL_RX_PROMISC,
 728                                   sg, 1, 0))
 729                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 730                          promisc ? "en" : "dis");
 731
 732         sg_init_one(sg, &allmulti, sizeof(allmulti));
 733
 734         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 735                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 736                                   sg, 1, 0))
 737                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
 738                          allmulti ? "en" : "dis");
 739
 740         /* MAC filter - use one buffer for both lists */
 741         mac_data = buf = kzalloc(((dev->uc.count + dev->mc_count) * ETH_ALEN) +
 742                                  (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
 743         if (!buf) {
 744                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
 745                 return;
 746         }
 747
 748         sg_init_table(sg, 2);
 749
 750         /* Store the unicast list and count in the front of the buffer */
 751         mac_data->entries = dev->uc.count;
 752         i = 0;
 753         list_for_each_entry(ha, &dev->uc.list, list)
 754                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 755
 756         sg_set_buf(&sg[0], mac_data,
 757                    sizeof(mac_data->entries) + (dev->uc.count * ETH_ALEN));
 758
 759         /* multicast list and count fill the end */
 760         mac_data = (void *)&mac_data->macs[dev->uc.count][0];
 761
 762         mac_data->entries = dev->mc_count;
 763         addr = dev->mc_list;
 764         for (i = 0; i < dev->mc_count; i++, addr = addr->next)
 765                 memcpy(&mac_data->macs[i][0], addr->da_addr, ETH_ALEN);
 766
 767         sg_set_buf(&sg[1], mac_data,
 768                    sizeof(mac_data->entries) + (dev->mc_count * ETH_ALEN));
 769
 770         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 771                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
 772                                   sg, 2, 0))
 773                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
 774
 775         kfree(buf);
 776 }
 777
 778 static void virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
 779 {
 780         struct virtnet_info *vi = netdev_priv(dev);
 781         struct scatterlist sg;
 782
 783         sg_init_one(&sg, &vid, sizeof(vid));
 784
 785         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 786                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
 787                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
 788 }
 789
 790 static void virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
 791 {
 792         struct virtnet_info *vi = netdev_priv(dev);
 793         struct scatterlist sg;
 794
 795         sg_init_one(&sg, &vid, sizeof(vid));
 796
 797         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 798                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
 799                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
 800 }
 801
 802 static const struct ethtool_ops virtnet_ethtool_ops = {
 803         .set_tx_csum = virtnet_set_tx_csum,
 804         .set_sg = ethtool_op_set_sg,
 805         .set_tso = ethtool_op_set_tso,
 806         .set_ufo = ethtool_op_set_ufo,
 807         .get_link = ethtool_op_get_link,
 808 };
 809
 810 #define MIN_MTU 68
 811 #define MAX_MTU 65535
 812
 813 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
 814 {
 815         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
 816                 return -EINVAL;
 817         dev->mtu = new_mtu;
 818         return 0;
 819 }
 820
 821 static const struct net_device_ops virtnet_netdev = {
 822         .ndo_open            = virtnet_open,
 823         .ndo_stop            = virtnet_close,
 824         .ndo_start_xmit      = start_xmit,
 825         .ndo_validate_addr   = eth_validate_addr,
 826         .ndo_set_mac_address = virtnet_set_mac_address,
 827         .ndo_set_rx_mode     = virtnet_set_rx_mode,
 828         .ndo_change_mtu      = virtnet_change_mtu,
 829         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
 830         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
 831 #ifdef CONFIG_NET_POLL_CONTROLLER
 832         .ndo_poll_controller = virtnet_netpoll,
 833 #endif
 834 };
 835
 836 static void virtnet_update_status(struct virtnet_info *vi)
 837 {
 838         u16 v;
 839
 840         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS))
 841                 return;
 842
 843         vi->vdev->config->get(vi->vdev,
 844                               offsetof(struct virtio_net_config, status),
 845                               &v, sizeof(v));
 846
 847         /* Ignore unknown (future) status bits */
 848         v &= VIRTIO_NET_S_LINK_UP;
 849
 850         if (vi->status == v)
 851                 return;
 852
 853         vi->status = v;
 854
 855         if (vi->status & VIRTIO_NET_S_LINK_UP) {
 856                 netif_carrier_on(vi->dev);
 857                 netif_wake_queue(vi->dev);
 858         } else {
 859                 netif_carrier_off(vi->dev);
 860                 netif_stop_queue(vi->dev);
 861         }
 862 }
 863
 864 static void virtnet_config_changed(struct virtio_device *vdev)
 865 {
 866         struct virtnet_info *vi = vdev->priv;
 867
 868         virtnet_update_status(vi);
 869 }
 870
 871 static int virtnet_probe(struct virtio_device *vdev)
 872 {
 873         int err;
 874         struct net_device *dev;
 875         struct virtnet_info *vi;
 876         struct virtqueue *vqs[3];
 877         vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
 878         const char *names[] = { "input", "output", "control" };
 879         int nvqs;
 880
 881         /* Allocate ourselves a network device with room for our info */
 882         dev = alloc_etherdev(sizeof(struct virtnet_info));
 883         if (!dev)
 884                 return -ENOMEM;
 885
 886         /* Set up network device as normal. */
 887         dev->netdev_ops = &virtnet_netdev;
 888         dev->features = NETIF_F_HIGHDMA;
 889         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
 890         SET_NETDEV_DEV(dev, &vdev->dev);
 891
 892         /* Do we support "hardware" checksums? */
 893         if (csum && virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
 894                 /* This opens up the world of extra features. */
 895                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
 896                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
 897                         dev->features |= NETIF_F_TSO | NETIF_F_UFO
 898                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
 899                 }
 900                 /* Individual feature bits: what can host handle? */
 901                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
 902                         dev->features |= NETIF_F_TSO;
 903                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
 904                         dev->features |= NETIF_F_TSO6;
 905                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
 906                         dev->features |= NETIF_F_TSO_ECN;
 907                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
 908                         dev->features |= NETIF_F_UFO;
 909         }
 910
 911         /* Configuration may specify what MAC to use.  Otherwise random. */
 912         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 913                 vdev->config->get(vdev,
 914                                   offsetof(struct virtio_net_config, mac),
 915                                   dev->dev_addr, dev->addr_len);
 916         } else
 917                 random_ether_addr(dev->dev_addr);
 918
 919         /* Set up our device-specific information */
 920         vi = netdev_priv(dev);
 921         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
 922         vi->dev = dev;
 923         vi->vdev = vdev;
 924         vdev->priv = vi;
 925         vi->pages = NULL;
 926         INIT_DELAYED_WORK(&vi->refill, refill_work);
 927
 928         /* If they give us a callback when all buffers are done, we don't need
 929          * the timer. */
 930         vi->free_in_tasklet = virtio_has_feature(vdev,VIRTIO_F_NOTIFY_ON_EMPTY);
 931
 932         /* If we can receive ANY GSO packets, we must allocate large ones. */
 933         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4)
 934             || virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6)
 935             || virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
 936                 vi->big_packets = true;
 937
 938         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
 939                 vi->mergeable_rx_bufs = true;
 940
 941         /* We expect two virtqueues, receive then send,
 942          * and optionally control. */
 943         nvqs = virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ? 3 : 2;
 944
 945         err = vdev->config->find_vqs(vdev, nvqs, vqs, callbacks, names);
 946         if (err)
 947                 goto free;
 948
 949         vi->rvq = vqs[0];
 950         vi->svq = vqs[1];
 951
 952         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)) {
 953                 vi->cvq = vqs[2];
 954
 955                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
 956                         dev->features |= NETIF_F_HW_VLAN_FILTER;
 957         }
 958
 959         /* Initialize our empty receive and send queues. */
 960         skb_queue_head_init(&vi->recv);
 961         skb_queue_head_init(&vi->send);
 962
 963         tasklet_init(&vi->tasklet, xmit_tasklet, (unsigned long)vi);
 964
 965         if (!vi->free_in_tasklet)
 966                 setup_timer(&vi->xmit_free_timer, xmit_free, (unsigned long)vi);
 967
 968         err = register_netdev(dev);
 969         if (err) {
 970                 pr_debug("virtio_net: registering device failed\n");
 971                 goto free_vqs;
 972         }
 973
 974         /* Last of all, set up some receive buffers. */
 975         try_fill_recv(vi, GFP_KERNEL);
 976
 977         /* If we didn't even get one input buffer, we're useless. */
 978         if (vi->num == 0) {
 979                 err = -ENOMEM;
 980                 goto unregister;
 981         }
 982
 983         vi->status = VIRTIO_NET_S_LINK_UP;
 984         virtnet_update_status(vi);
 985         netif_carrier_on(dev);
 986
 987         pr_debug("virtnet: registered device %s\n", dev->name);
 988         return 0;
 989
 990 unregister:
 991         unregister_netdev(dev);
 992         cancel_delayed_work_sync(&vi->refill);
 993 free_vqs:
 994         vdev->config->del_vqs(vdev);
 995 free:
 996         free_netdev(dev);
 997         return err;
 998 }
 999
1000 static void virtnet_remove(struct virtio_device *vdev)
1001 {
1002         struct virtnet_info *vi = vdev->priv;
1003         struct sk_buff *skb;
1004
1005         /* Stop all the virtqueues. */
1006         vdev->config->reset(vdev);
1007
1008         if (!vi->free_in_tasklet)
1009                 del_timer_sync(&vi->xmit_free_timer);
1010
1011         /* Free our skbs in send and recv queues, if any. */
1012         while ((skb = __skb_dequeue(&vi->recv)) != NULL) {
1013                 kfree_skb(skb);
1014                 vi->num--;
1015         }
1016         __skb_queue_purge(&vi->send);
1017
1018         BUG_ON(vi->num != 0);
1019
1020         unregister_netdev(vi->dev);
1021         cancel_delayed_work_sync(&vi->refill);
1022
1023         vdev->config->del_vqs(vi->vdev);
1024
1025         while (vi->pages)
1026                 __free_pages(get_a_page(vi, GFP_KERNEL), 0);
1027
1028         free_netdev(vi->dev);
1029 }
1030
1031 static struct virtio_device_id id_table[] = {
1032         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1033         { 0 },
1034 };
1035
1036 static unsigned int features[] = {
1037         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1038         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1039         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1040         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1041         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1042         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1043         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1044         VIRTIO_F_NOTIFY_ON_EMPTY,
1045 };
1046
1047 static struct virtio_driver virtio_net = {
1048         .feature_table = features,
1049         .feature_table_size = ARRAY_SIZE(features),
1050         .driver.name =  KBUILD_MODNAME,
1051         .driver.owner = THIS_MODULE,
1052         .id_table =     id_table,
1053         .probe =        virtnet_probe,
1054         .remove =       __devexit_p(virtnet_remove),
1055         .config_changed = virtnet_config_changed,
1056 };
1057
1058 static int __init init(void)
1059 {
1060         return register_virtio_driver(&virtio_net);
1061 }
1062
1063 static void __exit fini(void)
1064 {
1065         unregister_virtio_driver(&virtio_net);
1066 }
1067 module_init(init);
1068 module_exit(fini);
1069
1070 MODULE_DEVICE_TABLE(virtio, id_table);
1071 MODULE_DESCRIPTION("Virtio network driver");
1072 MODULE_LICENSE("GPL");