drivers/net/virtio_net.c

   1 /* A 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 #include <linux/slab.h>
  29
  30 static int napi_weight = 128;
  31 module_param(napi_weight, int, 0444);
  32
  33 static bool csum = true, gso = true;
  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 #define VIRTNET_DRIVER_VERSION "1.0.0"
  43
  44 struct virtnet_stats {
  45         struct u64_stats_sync syncp;
  46         u64 tx_bytes;
  47         u64 tx_packets;
  48
  49         u64 rx_bytes;
  50         u64 rx_packets;
  51 };
  52
  53 struct virtnet_info {
  54         struct virtio_device *vdev;
  55         struct virtqueue *rvq, *svq, *cvq;
  56         struct net_device *dev;
  57         struct napi_struct napi;
  58         unsigned int status;
  59
  60         /* Number of input buffers, and max we've ever had. */
  61         unsigned int num, max;
  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         /* Active statistics */
  70         struct virtnet_stats __percpu *stats;
  71
  72         /* Work struct for refilling if we run low on memory. */
  73         struct delayed_work refill;
  74
  75         /* Chain pages by the private ptr. */
  76         struct page *pages;
  77
  78         /* fragments + linear part + virtio header */
  79         struct scatterlist rx_sg[MAX_SKB_FRAGS + 2];
  80         struct scatterlist tx_sg[MAX_SKB_FRAGS + 2];
  81 };
  82
  83 struct skb_vnet_hdr {
  84         union {
  85                 struct virtio_net_hdr hdr;
  86                 struct virtio_net_hdr_mrg_rxbuf mhdr;
  87         };
  88         unsigned int num_sg;
  89 };
  90
  91 struct padded_vnet_hdr {
  92         struct virtio_net_hdr hdr;
  93         /*
  94          * virtio_net_hdr should be in a separated sg buffer because of a
  95          * QEMU bug, and data sg buffer shares same page with this header sg.
  96          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
  97          */
  98         char padding[6];
  99 };
 100
 101 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 102 {
 103         return (struct skb_vnet_hdr *)skb->cb;
 104 }
 105
 106 /*
 107  * private is used to chain pages for big packets, put the whole
 108  * most recent used list in the beginning for reuse
 109  */
 110 static void give_pages(struct virtnet_info *vi, struct page *page)
 111 {
 112         struct page *end;
 113
 114         /* Find end of list, sew whole thing into vi->pages. */
 115         for (end = page; end->private; end = (struct page *)end->private);
 116         end->private = (unsigned long)vi->pages;
 117         vi->pages = page;
 118 }
 119
 120 static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
 121 {
 122         struct page *p = vi->pages;
 123
 124         if (p) {
 125                 vi->pages = (struct page *)p->private;
 126                 /* clear private here, it is used to chain pages */
 127                 p->private = 0;
 128         } else
 129                 p = alloc_page(gfp_mask);
 130         return p;
 131 }
 132
 133 static void skb_xmit_done(struct virtqueue *svq)
 134 {
 135         struct virtnet_info *vi = svq->vdev->priv;
 136
 137         /* Suppress further interrupts. */
 138         virtqueue_disable_cb(svq);
 139
 140         /* We were probably waiting for more output buffers. */
 141         netif_wake_queue(vi->dev);
 142 }
 143
 144 static void set_skb_frag(struct sk_buff *skb, struct page *page,
 145                          unsigned int offset, unsigned int *len)
 146 {
 147         int size = min((unsigned)PAGE_SIZE - offset, *len);
 148         int i = skb_shinfo(skb)->nr_frags;
 149
 150         __skb_fill_page_desc(skb, i, page, offset, size);
 151
 152         skb->data_len += size;
 153         skb->len += size;
 154         skb->truesize += PAGE_SIZE;
 155         skb_shinfo(skb)->nr_frags++;
 156         *len -= size;
 157 }
 158
 159 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 160                                    struct page *page, unsigned int len)
 161 {
 162         struct sk_buff *skb;
 163         struct skb_vnet_hdr *hdr;
 164         unsigned int copy, hdr_len, offset;
 165         char *p;
 166
 167         p = page_address(page);
 168
 169         /* copy small packet so we can reuse these pages for small data */
 170         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 171         if (unlikely(!skb))
 172                 return NULL;
 173
 174         hdr = skb_vnet_hdr(skb);
 175
 176         if (vi->mergeable_rx_bufs) {
 177                 hdr_len = sizeof hdr->mhdr;
 178                 offset = hdr_len;
 179         } else {
 180                 hdr_len = sizeof hdr->hdr;
 181                 offset = sizeof(struct padded_vnet_hdr);
 182         }
 183
 184         memcpy(hdr, p, hdr_len);
 185
 186         len -= hdr_len;
 187         p += offset;
 188
 189         copy = len;
 190         if (copy > skb_tailroom(skb))
 191                 copy = skb_tailroom(skb);
 192         memcpy(skb_put(skb, copy), p, copy);
 193
 194         len -= copy;
 195         offset += copy;
 196
 197         /*
 198          * Verify that we can indeed put this data into a skb.
 199          * This is here to handle cases when the device erroneously
 200          * tries to receive more than is possible. This is usually
 201          * the case of a broken device.
 202          */
 203         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 204                 if (net_ratelimit())
 205                         pr_debug("%s: too much data\n", skb->dev->name);
 206                 dev_kfree_skb(skb);
 207                 return NULL;
 208         }
 209
 210         while (len) {
 211                 set_skb_frag(skb, page, offset, &len);
 212                 page = (struct page *)page->private;
 213                 offset = 0;
 214         }
 215
 216         if (page)
 217                 give_pages(vi, page);
 218
 219         return skb;
 220 }
 221
 222 static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
 223 {
 224         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 225         struct page *page;
 226         int num_buf, i, len;
 227
 228         num_buf = hdr->mhdr.num_buffers;
 229         while (--num_buf) {
 230                 i = skb_shinfo(skb)->nr_frags;
 231                 if (i >= MAX_SKB_FRAGS) {
 232                         pr_debug("%s: packet too long\n", skb->dev->name);
 233                         skb->dev->stats.rx_length_errors++;
 234                         return -EINVAL;
 235                 }
 236                 page = virtqueue_get_buf(vi->rvq, &len);
 237                 if (!page) {
 238                         pr_debug("%s: rx error: %d buffers missing\n",
 239                                  skb->dev->name, hdr->mhdr.num_buffers);
 240                         skb->dev->stats.rx_length_errors++;
 241                         return -EINVAL;
 242                 }
 243
 244                 if (len > PAGE_SIZE)
 245                         len = PAGE_SIZE;
 246
 247                 set_skb_frag(skb, page, 0, &len);
 248
 249                 --vi->num;
 250         }
 251         return 0;
 252 }
 253
 254 static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
 255 {
 256         struct virtnet_info *vi = netdev_priv(dev);
 257         struct virtnet_stats __percpu *stats = this_cpu_ptr(vi->stats);
 258         struct sk_buff *skb;
 259         struct page *page;
 260         struct skb_vnet_hdr *hdr;
 261
 262         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 263                 pr_debug("%s: short packet %i\n", dev->name, len);
 264                 dev->stats.rx_length_errors++;
 265                 if (vi->mergeable_rx_bufs || vi->big_packets)
 266                         give_pages(vi, buf);
 267                 else
 268                         dev_kfree_skb(buf);
 269                 return;
 270         }
 271
 272         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
 273                 skb = buf;
 274                 len -= sizeof(struct virtio_net_hdr);
 275                 skb_trim(skb, len);
 276         } else {
 277                 page = buf;
 278                 skb = page_to_skb(vi, page, len);
 279                 if (unlikely(!skb)) {
 280                         dev->stats.rx_dropped++;
 281                         give_pages(vi, page);
 282                         return;
 283                 }
 284                 if (vi->mergeable_rx_bufs)
 285                         if (receive_mergeable(vi, skb)) {
 286                                 dev_kfree_skb(skb);
 287                                 return;
 288                         }
 289         }
 290
 291         hdr = skb_vnet_hdr(skb);
 292
 293         u64_stats_update_begin(&stats->syncp);
 294         stats->rx_bytes += skb->len;
 295         stats->rx_packets++;
 296         u64_stats_update_end(&stats->syncp);
 297
 298         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 299                 pr_debug("Needs csum!\n");
 300                 if (!skb_partial_csum_set(skb,
 301                                           hdr->hdr.csum_start,
 302                                           hdr->hdr.csum_offset))
 303                         goto frame_err;
 304         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 305                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 306         }
 307
 308         skb->protocol = eth_type_trans(skb, dev);
 309         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 310                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 311
 312         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 313                 pr_debug("GSO!\n");
 314                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 315                 case VIRTIO_NET_HDR_GSO_TCPV4:
 316                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 317                         break;
 318                 case VIRTIO_NET_HDR_GSO_UDP:
 319                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 320                         break;
 321                 case VIRTIO_NET_HDR_GSO_TCPV6:
 322                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 323                         break;
 324                 default:
 325                         if (net_ratelimit())
 326                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
 327                                        dev->name, hdr->hdr.gso_type);
 328                         goto frame_err;
 329                 }
 330
 331                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 332                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 333
 334                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 335                 if (skb_shinfo(skb)->gso_size == 0) {
 336                         if (net_ratelimit())
 337                                 printk(KERN_WARNING "%s: zero gso size.\n",
 338                                        dev->name);
 339                         goto frame_err;
 340                 }
 341
 342                 /* Header must be checked, and gso_segs computed. */
 343                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 344                 skb_shinfo(skb)->gso_segs = 0;
 345         }
 346
 347         netif_receive_skb(skb);
 348         return;
 349
 350 frame_err:
 351         dev->stats.rx_frame_errors++;
 352         dev_kfree_skb(skb);
 353 }
 354
 355 static int add_recvbuf_small(struct virtnet_info *vi, gfp_t gfp)
 356 {
 357         struct sk_buff *skb;
 358         struct skb_vnet_hdr *hdr;
 359         int err;
 360
 361         skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
 362         if (unlikely(!skb))
 363                 return -ENOMEM;
 364
 365         skb_put(skb, MAX_PACKET_LEN);
 366
 367         hdr = skb_vnet_hdr(skb);
 368         sg_set_buf(vi->rx_sg, &hdr->hdr, sizeof hdr->hdr);
 369
 370         skb_to_sgvec(skb, vi->rx_sg + 1, 0, skb->len);
 371
 372         err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, 2, skb, gfp);
 373         if (err < 0)
 374                 dev_kfree_skb(skb);
 375
 376         return err;
 377 }
 378
 379 static int add_recvbuf_big(struct virtnet_info *vi, gfp_t gfp)
 380 {
 381         struct page *first, *list = NULL;
 382         char *p;
 383         int i, err, offset;
 384
 385         /* page in vi->rx_sg[MAX_SKB_FRAGS + 1] is list tail */
 386         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 387                 first = get_a_page(vi, gfp);
 388                 if (!first) {
 389                         if (list)
 390                                 give_pages(vi, list);
 391                         return -ENOMEM;
 392                 }
 393                 sg_set_buf(&vi->rx_sg[i], page_address(first), PAGE_SIZE);
 394
 395                 /* chain new page in list head to match sg */
 396                 first->private = (unsigned long)list;
 397                 list = first;
 398         }
 399
 400         first = get_a_page(vi, gfp);
 401         if (!first) {
 402                 give_pages(vi, list);
 403                 return -ENOMEM;
 404         }
 405         p = page_address(first);
 406
 407         /* vi->rx_sg[0], vi->rx_sg[1] share the same page */
 408         /* a separated vi->rx_sg[0] for virtio_net_hdr only due to QEMU bug */
 409         sg_set_buf(&vi->rx_sg[0], p, sizeof(struct virtio_net_hdr));
 410
 411         /* vi->rx_sg[1] for data packet, from offset */
 412         offset = sizeof(struct padded_vnet_hdr);
 413         sg_set_buf(&vi->rx_sg[1], p + offset, PAGE_SIZE - offset);
 414
 415         /* chain first in list head */
 416         first->private = (unsigned long)list;
 417         err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, MAX_SKB_FRAGS + 2,
 418                                     first, gfp);
 419         if (err < 0)
 420                 give_pages(vi, first);
 421
 422         return err;
 423 }
 424
 425 static int add_recvbuf_mergeable(struct virtnet_info *vi, gfp_t gfp)
 426 {
 427         struct page *page;
 428         int err;
 429
 430         page = get_a_page(vi, gfp);
 431         if (!page)
 432                 return -ENOMEM;
 433
 434         sg_init_one(vi->rx_sg, page_address(page), PAGE_SIZE);
 435
 436         err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, 1, page, gfp);
 437         if (err < 0)
 438                 give_pages(vi, page);
 439
 440         return err;
 441 }
 442
 443 /*
 444  * Returns false if we couldn't fill entirely (OOM).
 445  *
 446  * Normally run in the receive path, but can also be run from ndo_open
 447  * before we're receiving packets, or from refill_work which is
 448  * careful to disable receiving (using napi_disable).
 449  */
 450 static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
 451 {
 452         int err;
 453         bool oom;
 454
 455         do {
 456                 if (vi->mergeable_rx_bufs)
 457                         err = add_recvbuf_mergeable(vi, gfp);
 458                 else if (vi->big_packets)
 459                         err = add_recvbuf_big(vi, gfp);
 460                 else
 461                         err = add_recvbuf_small(vi, gfp);
 462
 463                 oom = err == -ENOMEM;
 464                 if (err < 0)
 465                         break;
 466                 ++vi->num;
 467         } while (err > 0);
 468         if (unlikely(vi->num > vi->max))
 469                 vi->max = vi->num;
 470         virtqueue_kick(vi->rvq);
 471         return !oom;
 472 }
 473
 474 static void skb_recv_done(struct virtqueue *rvq)
 475 {
 476         struct virtnet_info *vi = rvq->vdev->priv;
 477         /* Schedule NAPI, Suppress further interrupts if successful. */
 478         if (napi_schedule_prep(&vi->napi)) {
 479                 virtqueue_disable_cb(rvq);
 480                 __napi_schedule(&vi->napi);
 481         }
 482 }
 483
 484 static void virtnet_napi_enable(struct virtnet_info *vi)
 485 {
 486         napi_enable(&vi->napi);
 487
 488         /* If all buffers were filled by other side before we napi_enabled, we
 489          * won't get another interrupt, so process any outstanding packets
 490          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 491          * We synchronize against interrupts via NAPI_STATE_SCHED */
 492         if (napi_schedule_prep(&vi->napi)) {
 493                 virtqueue_disable_cb(vi->rvq);
 494                 __napi_schedule(&vi->napi);
 495         }
 496 }
 497
 498 static void refill_work(struct work_struct *work)
 499 {
 500         struct virtnet_info *vi;
 501         bool still_empty;
 502
 503         vi = container_of(work, struct virtnet_info, refill.work);
 504         napi_disable(&vi->napi);
 505         still_empty = !try_fill_recv(vi, GFP_KERNEL);
 506         virtnet_napi_enable(vi);
 507
 508         /* In theory, this can happen: if we don't get any buffers in
 509          * we will *never* try to fill again. */
 510         if (still_empty)
 511                 queue_delayed_work(system_nrt_wq, &vi->refill, HZ/2);
 512 }
 513
 514 static int virtnet_poll(struct napi_struct *napi, int budget)
 515 {
 516         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
 517         void *buf;
 518         unsigned int len, received = 0;
 519
 520 again:
 521         while (received < budget &&
 522                (buf = virtqueue_get_buf(vi->rvq, &len)) != NULL) {
 523                 receive_buf(vi->dev, buf, len);
 524                 --vi->num;
 525                 received++;
 526         }
 527
 528         if (vi->num < vi->max / 2) {
 529                 if (!try_fill_recv(vi, GFP_ATOMIC))
 530                         queue_delayed_work(system_nrt_wq, &vi->refill, 0);
 531         }
 532
 533         /* Out of packets? */
 534         if (received < budget) {
 535                 napi_complete(napi);
 536                 if (unlikely(!virtqueue_enable_cb(vi->rvq)) &&
 537                     napi_schedule_prep(napi)) {
 538                         virtqueue_disable_cb(vi->rvq);
 539                         __napi_schedule(napi);
 540                         goto again;
 541                 }
 542         }
 543
 544         return received;
 545 }
 546
 547 static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
 548 {
 549         struct sk_buff *skb;
 550         unsigned int len, tot_sgs = 0;
 551         struct virtnet_stats __percpu *stats = this_cpu_ptr(vi->stats);
 552
 553         while ((skb = virtqueue_get_buf(vi->svq, &len)) != NULL) {
 554                 pr_debug("Sent skb %p\n", skb);
 555
 556                 u64_stats_update_begin(&stats->syncp);
 557                 stats->tx_bytes += skb->len;
 558                 stats->tx_packets++;
 559                 u64_stats_update_end(&stats->syncp);
 560
 561                 tot_sgs += skb_vnet_hdr(skb)->num_sg;
 562                 dev_kfree_skb_any(skb);
 563         }
 564         return tot_sgs;
 565 }
 566
 567 static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 568 {
 569         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 570         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 571
 572         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 573
 574         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 575                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 576                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 577                 hdr->hdr.csum_offset = skb->csum_offset;
 578         } else {
 579                 hdr->hdr.flags = 0;
 580                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 581         }
 582
 583         if (skb_is_gso(skb)) {
 584                 hdr->hdr.hdr_len = skb_headlen(skb);
 585                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 586                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 587                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 588                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 589                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 590                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 591                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 592                 else
 593                         BUG();
 594                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 595                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 596         } else {
 597                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 598                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 599         }
 600
 601         hdr->mhdr.num_buffers = 0;
 602
 603         /* Encode metadata header at front. */
 604         if (vi->mergeable_rx_bufs)
 605                 sg_set_buf(vi->tx_sg, &hdr->mhdr, sizeof hdr->mhdr);
 606         else
 607                 sg_set_buf(vi->tx_sg, &hdr->hdr, sizeof hdr->hdr);
 608
 609         hdr->num_sg = skb_to_sgvec(skb, vi->tx_sg + 1, 0, skb->len) + 1;
 610         return virtqueue_add_buf(vi->svq, vi->tx_sg, hdr->num_sg,
 611                                         0, skb);
 612 }
 613
 614 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 615 {
 616         struct virtnet_info *vi = netdev_priv(dev);
 617         int capacity;
 618
 619         /* Free up any pending old buffers before queueing new ones. */
 620         free_old_xmit_skbs(vi);
 621
 622         /* Try to transmit */
 623         capacity = xmit_skb(vi, skb);
 624
 625         /* This can happen with OOM and indirect buffers. */
 626         if (unlikely(capacity < 0)) {
 627                 if (net_ratelimit()) {
 628                         if (likely(capacity == -ENOMEM)) {
 629                                 dev_warn(&dev->dev,
 630                                          "TX queue failure: out of memory\n");
 631                         } else {
 632                                 dev->stats.tx_fifo_errors++;
 633                                 dev_warn(&dev->dev,
 634                                          "Unexpected TX queue failure: %d\n",
 635                                          capacity);
 636                         }
 637                 }
 638                 dev->stats.tx_dropped++;
 639                 kfree_skb(skb);
 640                 return NETDEV_TX_OK;
 641         }
 642         virtqueue_kick(vi->svq);
 643
 644         /* Don't wait up for transmitted skbs to be freed. */
 645         skb_orphan(skb);
 646         nf_reset(skb);
 647
 648         /* Apparently nice girls don't return TX_BUSY; stop the queue
 649          * before it gets out of hand.  Naturally, this wastes entries. */
 650         if (capacity < 2+MAX_SKB_FRAGS) {
 651                 netif_stop_queue(dev);
 652                 if (unlikely(!virtqueue_enable_cb_delayed(vi->svq))) {
 653                         /* More just got used, free them then recheck. */
 654                         capacity += free_old_xmit_skbs(vi);
 655                         if (capacity >= 2+MAX_SKB_FRAGS) {
 656                                 netif_start_queue(dev);
 657                                 virtqueue_disable_cb(vi->svq);
 658                         }
 659                 }
 660         }
 661
 662         return NETDEV_TX_OK;
 663 }
 664
 665 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 666 {
 667         struct virtnet_info *vi = netdev_priv(dev);
 668         struct virtio_device *vdev = vi->vdev;
 669         int ret;
 670
 671         ret = eth_mac_addr(dev, p);
 672         if (ret)
 673                 return ret;
 674
 675         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
 676                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 677                                   dev->dev_addr, dev->addr_len);
 678
 679         return 0;
 680 }
 681
 682 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 683                                                struct rtnl_link_stats64 *tot)
 684 {
 685         struct virtnet_info *vi = netdev_priv(dev);
 686         int cpu;
 687         unsigned int start;
 688
 689         for_each_possible_cpu(cpu) {
 690                 struct virtnet_stats __percpu *stats
 691                         = per_cpu_ptr(vi->stats, cpu);
 692                 u64 tpackets, tbytes, rpackets, rbytes;
 693
 694                 do {
 695                         start = u64_stats_fetch_begin(&stats->syncp);
 696                         tpackets = stats->tx_packets;
 697                         tbytes   = stats->tx_bytes;
 698                         rpackets = stats->rx_packets;
 699                         rbytes   = stats->rx_bytes;
 700                 } while (u64_stats_fetch_retry(&stats->syncp, start));
 701
 702                 tot->rx_packets += rpackets;
 703                 tot->tx_packets += tpackets;
 704                 tot->rx_bytes   += rbytes;
 705                 tot->tx_bytes   += tbytes;
 706         }
 707
 708         tot->tx_dropped = dev->stats.tx_dropped;
 709         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 710         tot->rx_dropped = dev->stats.rx_dropped;
 711         tot->rx_length_errors = dev->stats.rx_length_errors;
 712         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 713
 714         return tot;
 715 }
 716
 717 #ifdef CONFIG_NET_POLL_CONTROLLER
 718 static void virtnet_netpoll(struct net_device *dev)
 719 {
 720         struct virtnet_info *vi = netdev_priv(dev);
 721
 722         napi_schedule(&vi->napi);
 723 }
 724 #endif
 725
 726 static int virtnet_open(struct net_device *dev)
 727 {
 728         struct virtnet_info *vi = netdev_priv(dev);
 729
 730         /* Make sure we have some buffers: if oom use wq. */
 731         if (!try_fill_recv(vi, GFP_KERNEL))
 732                 queue_delayed_work(system_nrt_wq, &vi->refill, 0);
 733
 734         virtnet_napi_enable(vi);
 735         return 0;
 736 }
 737
 738 /*
 739  * Send command via the control virtqueue and check status.  Commands
 740  * supported by the hypervisor, as indicated by feature bits, should
 741  * never fail unless improperly formated.
 742  */
 743 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 744                                  struct scatterlist *data, int out, int in)
 745 {
 746         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
 747         struct virtio_net_ctrl_hdr ctrl;
 748         virtio_net_ctrl_ack status = ~0;
 749         unsigned int tmp;
 750         int i;
 751
 752         /* Caller should know better */
 753         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
 754                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
 755
 756         out++; /* Add header */
 757         in++; /* Add return status */
 758
 759         ctrl.class = class;
 760         ctrl.cmd = cmd;
 761
 762         sg_init_table(sg, out + in);
 763
 764         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
 765         for_each_sg(data, s, out + in - 2, i)
 766                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
 767         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
 768
 769         BUG_ON(virtqueue_add_buf(vi->cvq, sg, out, in, vi) < 0);
 770
 771         virtqueue_kick(vi->cvq);
 772
 773         /*
 774          * Spin for a response, the kick causes an ioport write, trapping
 775          * into the hypervisor, so the request should be handled immediately.
 776          */
 777         while (!virtqueue_get_buf(vi->cvq, &tmp))
 778                 cpu_relax();
 779
 780         return status == VIRTIO_NET_OK;
 781 }
 782
 783 static int virtnet_close(struct net_device *dev)
 784 {
 785         struct virtnet_info *vi = netdev_priv(dev);
 786
 787         /* Make sure refill_work doesn't re-enable napi! */
 788         cancel_delayed_work_sync(&vi->refill);
 789         napi_disable(&vi->napi);
 790
 791         return 0;
 792 }
 793
 794 static void virtnet_set_rx_mode(struct net_device *dev)
 795 {
 796         struct virtnet_info *vi = netdev_priv(dev);
 797         struct scatterlist sg[2];
 798         u8 promisc, allmulti;
 799         struct virtio_net_ctrl_mac *mac_data;
 800         struct netdev_hw_addr *ha;
 801         int uc_count;
 802         int mc_count;
 803         void *buf;
 804         int i;
 805
 806         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 807         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 808                 return;
 809
 810         promisc = ((dev->flags & IFF_PROMISC) != 0);
 811         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 812
 813         sg_init_one(sg, &promisc, sizeof(promisc));
 814
 815         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 816                                   VIRTIO_NET_CTRL_RX_PROMISC,
 817                                   sg, 1, 0))
 818                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 819                          promisc ? "en" : "dis");
 820
 821         sg_init_one(sg, &allmulti, sizeof(allmulti));
 822
 823         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 824                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 825                                   sg, 1, 0))
 826                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
 827                          allmulti ? "en" : "dis");
 828
 829         uc_count = netdev_uc_count(dev);
 830         mc_count = netdev_mc_count(dev);
 831         /* MAC filter - use one buffer for both lists */
 832         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
 833                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
 834         mac_data = buf;
 835         if (!buf) {
 836                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
 837                 return;
 838         }
 839
 840         sg_init_table(sg, 2);
 841
 842         /* Store the unicast list and count in the front of the buffer */
 843         mac_data->entries = uc_count;
 844         i = 0;
 845         netdev_for_each_uc_addr(ha, dev)
 846                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 847
 848         sg_set_buf(&sg[0], mac_data,
 849                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
 850
 851         /* multicast list and count fill the end */
 852         mac_data = (void *)&mac_data->macs[uc_count][0];
 853
 854         mac_data->entries = mc_count;
 855         i = 0;
 856         netdev_for_each_mc_addr(ha, dev)
 857                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 858
 859         sg_set_buf(&sg[1], mac_data,
 860                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
 861
 862         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 863                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
 864                                   sg, 2, 0))
 865                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
 866
 867         kfree(buf);
 868 }
 869
 870 static int virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
 871 {
 872         struct virtnet_info *vi = netdev_priv(dev);
 873         struct scatterlist sg;
 874
 875         sg_init_one(&sg, &vid, sizeof(vid));
 876
 877         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 878                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
 879                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
 880         return 0;
 881 }
 882
 883 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
 884 {
 885         struct virtnet_info *vi = netdev_priv(dev);
 886         struct scatterlist sg;
 887
 888         sg_init_one(&sg, &vid, sizeof(vid));
 889
 890         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 891                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
 892                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
 893         return 0;
 894 }
 895
 896 static void virtnet_get_ringparam(struct net_device *dev,
 897                                 struct ethtool_ringparam *ring)
 898 {
 899         struct virtnet_info *vi = netdev_priv(dev);
 900
 901         ring->rx_max_pending = virtqueue_get_vring_size(vi->rvq);
 902         ring->tx_max_pending = virtqueue_get_vring_size(vi->svq);
 903         ring->rx_pending = ring->rx_max_pending;
 904         ring->tx_pending = ring->tx_max_pending;
 905
 906 }
 907
 908
 909 static void virtnet_get_drvinfo(struct net_device *dev,
 910                                 struct ethtool_drvinfo *info)
 911 {
 912         struct virtnet_info *vi = netdev_priv(dev);
 913         struct virtio_device *vdev = vi->vdev;
 914
 915         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
 916         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
 917         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
 918
 919 }
 920
 921 static const struct ethtool_ops virtnet_ethtool_ops = {
 922         .get_drvinfo = virtnet_get_drvinfo,
 923         .get_link = ethtool_op_get_link,
 924         .get_ringparam = virtnet_get_ringparam,
 925 };
 926
 927 #define MIN_MTU 68
 928 #define MAX_MTU 65535
 929
 930 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
 931 {
 932         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
 933                 return -EINVAL;
 934         dev->mtu = new_mtu;
 935         return 0;
 936 }
 937
 938 static const struct net_device_ops virtnet_netdev = {
 939         .ndo_open            = virtnet_open,
 940         .ndo_stop            = virtnet_close,
 941         .ndo_start_xmit      = start_xmit,
 942         .ndo_validate_addr   = eth_validate_addr,
 943         .ndo_set_mac_address = virtnet_set_mac_address,
 944         .ndo_set_rx_mode     = virtnet_set_rx_mode,
 945         .ndo_change_mtu      = virtnet_change_mtu,
 946         .ndo_get_stats64     = virtnet_stats,
 947         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
 948         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
 949 #ifdef CONFIG_NET_POLL_CONTROLLER
 950         .ndo_poll_controller = virtnet_netpoll,
 951 #endif
 952 };
 953
 954 static void virtnet_update_status(struct virtnet_info *vi)
 955 {
 956         u16 v;
 957
 958         if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
 959                               offsetof(struct virtio_net_config, status),
 960                               &v) < 0)
 961                 return;
 962
 963         /* Ignore unknown (future) status bits */
 964         v &= VIRTIO_NET_S_LINK_UP;
 965
 966         if (vi->status == v)
 967                 return;
 968
 969         vi->status = v;
 970
 971         if (vi->status & VIRTIO_NET_S_LINK_UP) {
 972                 netif_carrier_on(vi->dev);
 973                 netif_wake_queue(vi->dev);
 974         } else {
 975                 netif_carrier_off(vi->dev);
 976                 netif_stop_queue(vi->dev);
 977         }
 978 }
 979
 980 static void virtnet_config_changed(struct virtio_device *vdev)
 981 {
 982         struct virtnet_info *vi = vdev->priv;
 983
 984         virtnet_update_status(vi);
 985 }
 986
 987 static int virtnet_probe(struct virtio_device *vdev)
 988 {
 989         int err;
 990         struct net_device *dev;
 991         struct virtnet_info *vi;
 992         struct virtqueue *vqs[3];
 993         vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
 994         const char *names[] = { "input", "output", "control" };
 995         int nvqs;
 996
 997         /* Allocate ourselves a network device with room for our info */
 998         dev = alloc_etherdev(sizeof(struct virtnet_info));
 999         if (!dev)
1000                 return -ENOMEM;
1001
1002         /* Set up network device as normal. */
1003         dev->priv_flags |= IFF_UNICAST_FLT;
1004         dev->netdev_ops = &virtnet_netdev;
1005         dev->features = NETIF_F_HIGHDMA;
1006
1007         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1008         SET_NETDEV_DEV(dev, &vdev->dev);
1009
1010         /* Do we support "hardware" checksums? */
1011         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1012                 /* This opens up the world of extra features. */
1013                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1014                 if (csum)
1015                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1016
1017                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1018                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1019                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1020                 }
1021                 /* Individual feature bits: what can host handle? */
1022                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1023                         dev->hw_features |= NETIF_F_TSO;
1024                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1025                         dev->hw_features |= NETIF_F_TSO6;
1026                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1027                         dev->hw_features |= NETIF_F_TSO_ECN;
1028                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1029                         dev->hw_features |= NETIF_F_UFO;
1030
1031                 if (gso)
1032                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1033                 /* (!csum && gso) case will be fixed by register_netdev() */
1034         }
1035
1036         /* Configuration may specify what MAC to use.  Otherwise random. */
1037         if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1038                                   offsetof(struct virtio_net_config, mac),
1039                                   dev->dev_addr, dev->addr_len) < 0)
1040                 random_ether_addr(dev->dev_addr);
1041
1042         /* Set up our device-specific information */
1043         vi = netdev_priv(dev);
1044         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
1045         vi->dev = dev;
1046         vi->vdev = vdev;
1047         vdev->priv = vi;
1048         vi->pages = NULL;
1049         vi->stats = alloc_percpu(struct virtnet_stats);
1050         err = -ENOMEM;
1051         if (vi->stats == NULL)
1052                 goto free;
1053
1054         INIT_DELAYED_WORK(&vi->refill, refill_work);
1055         sg_init_table(vi->rx_sg, ARRAY_SIZE(vi->rx_sg));
1056         sg_init_table(vi->tx_sg, ARRAY_SIZE(vi->tx_sg));
1057
1058         /* If we can receive ANY GSO packets, we must allocate large ones. */
1059         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1060             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1061             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1062                 vi->big_packets = true;
1063
1064         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1065                 vi->mergeable_rx_bufs = true;
1066
1067         /* We expect two virtqueues, receive then send,
1068          * and optionally control. */
1069         nvqs = virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ? 3 : 2;
1070
1071         err = vdev->config->find_vqs(vdev, nvqs, vqs, callbacks, names);
1072         if (err)
1073                 goto free_stats;
1074
1075         vi->rvq = vqs[0];
1076         vi->svq = vqs[1];
1077
1078         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)) {
1079                 vi->cvq = vqs[2];
1080
1081                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1082                         dev->features |= NETIF_F_HW_VLAN_FILTER;
1083         }
1084
1085         err = register_netdev(dev);
1086         if (err) {
1087                 pr_debug("virtio_net: registering device failed\n");
1088                 goto free_vqs;
1089         }
1090
1091         /* Last of all, set up some receive buffers. */
1092         try_fill_recv(vi, GFP_KERNEL);
1093
1094         /* If we didn't even get one input buffer, we're useless. */
1095         if (vi->num == 0) {
1096                 err = -ENOMEM;
1097                 goto unregister;
1098         }
1099
1100         /* Assume link up if device can't report link status,
1101            otherwise get link status from config. */
1102         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1103                 netif_carrier_off(dev);
1104                 virtnet_update_status(vi);
1105         } else {
1106                 vi->status = VIRTIO_NET_S_LINK_UP;
1107                 netif_carrier_on(dev);
1108         }
1109
1110         pr_debug("virtnet: registered device %s\n", dev->name);
1111         return 0;
1112
1113 unregister:
1114         unregister_netdev(dev);
1115 free_vqs:
1116         vdev->config->del_vqs(vdev);
1117 free_stats:
1118         free_percpu(vi->stats);
1119 free:
1120         free_netdev(dev);
1121         return err;
1122 }
1123
1124 static void free_unused_bufs(struct virtnet_info *vi)
1125 {
1126         void *buf;
1127         while (1) {
1128                 buf = virtqueue_detach_unused_buf(vi->svq);
1129                 if (!buf)
1130                         break;
1131                 dev_kfree_skb(buf);
1132         }
1133         while (1) {
1134                 buf = virtqueue_detach_unused_buf(vi->rvq);
1135                 if (!buf)
1136                         break;
1137                 if (vi->mergeable_rx_bufs || vi->big_packets)
1138                         give_pages(vi, buf);
1139                 else
1140                         dev_kfree_skb(buf);
1141                 --vi->num;
1142         }
1143         BUG_ON(vi->num != 0);
1144 }
1145
1146 static void __devexit virtnet_remove(struct virtio_device *vdev)
1147 {
1148         struct virtnet_info *vi = vdev->priv;
1149
1150         /* Stop all the virtqueues. */
1151         vdev->config->reset(vdev);
1152
1153         unregister_netdev(vi->dev);
1154
1155         /* Free unused buffers in both send and recv, if any. */
1156         free_unused_bufs(vi);
1157
1158         vdev->config->del_vqs(vi->vdev);
1159
1160         while (vi->pages)
1161                 __free_pages(get_a_page(vi, GFP_KERNEL), 0);
1162
1163         free_percpu(vi->stats);
1164         free_netdev(vi->dev);
1165 }
1166
1167 static struct virtio_device_id id_table[] = {
1168         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1169         { 0 },
1170 };
1171
1172 static unsigned int features[] = {
1173         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1174         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1175         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1176         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1177         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1178         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1179         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1180 };
1181
1182 static struct virtio_driver virtio_net_driver = {
1183         .feature_table = features,
1184         .feature_table_size = ARRAY_SIZE(features),
1185         .driver.name =  KBUILD_MODNAME,
1186         .driver.owner = THIS_MODULE,
1187         .id_table =     id_table,
1188         .probe =        virtnet_probe,
1189         .remove =       __devexit_p(virtnet_remove),
1190         .config_changed = virtnet_config_changed,
1191 };
1192
1193 static int __init init(void)
1194 {
1195         return register_virtio_driver(&virtio_net_driver);
1196 }
1197
1198 static void __exit fini(void)
1199 {
1200         unregister_virtio_driver(&virtio_net_driver);
1201 }
1202 module_init(init);
1203 module_exit(fini);
1204
1205 MODULE_DEVICE_TABLE(virtio, id_table);
1206 MODULE_DESCRIPTION("Virtio network driver");
1207 MODULE_LICENSE("GPL");