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 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_stats {
  44         struct u64_stats_sync syncp;
  45         u64 tx_bytes;
  46         u64 tx_packets;
  47
  48         u64 rx_bytes;
  49         u64 rx_packets;
  50 };
  51
  52 struct virtnet_info {
  53         struct virtio_device *vdev;
  54         struct virtqueue *rvq, *svq, *cvq;
  55         struct net_device *dev;
  56         struct napi_struct napi;
  57         unsigned int status;
  58
  59         /* Number of input buffers, and max we've ever had. */
  60         unsigned int num, max;
  61
  62         /* I like... big packets and I cannot lie! */
  63         bool big_packets;
  64
  65         /* Host will merge rx buffers for big packets (shake it! shake it!) */
  66         bool mergeable_rx_bufs;
  67
  68         /* Active statistics */
  69         struct virtnet_stats __percpu *stats;
  70
  71         /* Work struct for refilling if we run low on memory. */
  72         struct delayed_work refill;
  73
  74         /* Chain pages by the private ptr. */
  75         struct page *pages;
  76
  77         /* fragments + linear part + virtio header */
  78         struct scatterlist rx_sg[MAX_SKB_FRAGS + 2];
  79         struct scatterlist tx_sg[MAX_SKB_FRAGS + 2];
  80 };
  81
  82 struct skb_vnet_hdr {
  83         union {
  84                 struct virtio_net_hdr hdr;
  85                 struct virtio_net_hdr_mrg_rxbuf mhdr;
  86         };
  87         unsigned int num_sg;
  88 };
  89
  90 struct padded_vnet_hdr {
  91         struct virtio_net_hdr hdr;
  92         /*
  93          * virtio_net_hdr should be in a separated sg buffer because of a
  94          * QEMU bug, and data sg buffer shares same page with this header sg.
  95          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
  96          */
  97         char padding[6];
  98 };
  99
 100 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 101 {
 102         return (struct skb_vnet_hdr *)skb->cb;
 103 }
 104
 105 /*
 106  * private is used to chain pages for big packets, put the whole
 107  * most recent used list in the beginning for reuse
 108  */
 109 static void give_pages(struct virtnet_info *vi, struct page *page)
 110 {
 111         struct page *end;
 112
 113         /* Find end of list, sew whole thing into vi->pages. */
 114         for (end = page; end->private; end = (struct page *)end->private);
 115         end->private = (unsigned long)vi->pages;
 116         vi->pages = page;
 117 }
 118
 119 static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
 120 {
 121         struct page *p = vi->pages;
 122
 123         if (p) {
 124                 vi->pages = (struct page *)p->private;
 125                 /* clear private here, it is used to chain pages */
 126                 p->private = 0;
 127         } else
 128                 p = alloc_page(gfp_mask);
 129         return p;
 130 }
 131
 132 static void skb_xmit_done(struct virtqueue *svq)
 133 {
 134         struct virtnet_info *vi = svq->vdev->priv;
 135
 136         /* Suppress further interrupts. */
 137         virtqueue_disable_cb(svq);
 138
 139         /* We were probably waiting for more output buffers. */
 140         netif_wake_queue(vi->dev);
 141 }
 142
 143 static void set_skb_frag(struct sk_buff *skb, struct page *page,
 144                          unsigned int offset, unsigned int *len)
 145 {
 146         int i = skb_shinfo(skb)->nr_frags;
 147         skb_frag_t *f;
 148
 149         f = &skb_shinfo(skb)->frags[i];
 150         f->size = min((unsigned)PAGE_SIZE - offset, *len);
 151         f->page_offset = offset;
 152         f->page = page;
 153
 154         skb->data_len += f->size;
 155         skb->len += f->size;
 156         skb_shinfo(skb)->nr_frags++;
 157         *len -= f->size;
 158 }
 159
 160 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 161                                    struct page *page, unsigned int len)
 162 {
 163         struct sk_buff *skb;
 164         struct skb_vnet_hdr *hdr;
 165         unsigned int copy, hdr_len, offset;
 166         char *p;
 167
 168         p = page_address(page);
 169
 170         /* copy small packet so we can reuse these pages for small data */
 171         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 172         if (unlikely(!skb))
 173                 return NULL;
 174
 175         hdr = skb_vnet_hdr(skb);
 176
 177         if (vi->mergeable_rx_bufs) {
 178                 hdr_len = sizeof hdr->mhdr;
 179                 offset = hdr_len;
 180         } else {
 181                 hdr_len = sizeof hdr->hdr;
 182                 offset = sizeof(struct padded_vnet_hdr);
 183         }
 184
 185         memcpy(hdr, p, hdr_len);
 186
 187         len -= hdr_len;
 188         p += offset;
 189
 190         copy = len;
 191         if (copy > skb_tailroom(skb))
 192                 copy = skb_tailroom(skb);
 193         memcpy(skb_put(skb, copy), p, copy);
 194
 195         len -= copy;
 196         offset += copy;
 197
 198         while (len) {
 199                 set_skb_frag(skb, page, offset, &len);
 200                 page = (struct page *)page->private;
 201                 offset = 0;
 202         }
 203
 204         if (page)
 205                 give_pages(vi, page);
 206
 207         return skb;
 208 }
 209
 210 static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
 211 {
 212         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 213         struct page *page;
 214         int num_buf, i, len;
 215
 216         num_buf = hdr->mhdr.num_buffers;
 217         while (--num_buf) {
 218                 i = skb_shinfo(skb)->nr_frags;
 219                 if (i >= MAX_SKB_FRAGS) {
 220                         pr_debug("%s: packet too long\n", skb->dev->name);
 221                         skb->dev->stats.rx_length_errors++;
 222                         return -EINVAL;
 223                 }
 224                 page = virtqueue_get_buf(vi->rvq, &len);
 225                 if (!page) {
 226                         pr_debug("%s: rx error: %d buffers missing\n",
 227                                  skb->dev->name, hdr->mhdr.num_buffers);
 228                         skb->dev->stats.rx_length_errors++;
 229                         return -EINVAL;
 230                 }
 231
 232                 if (len > PAGE_SIZE)
 233                         len = PAGE_SIZE;
 234
 235                 set_skb_frag(skb, page, 0, &len);
 236
 237                 --vi->num;
 238         }
 239         return 0;
 240 }
 241
 242 static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
 243 {
 244         struct virtnet_info *vi = netdev_priv(dev);
 245         struct virtnet_stats __percpu *stats = this_cpu_ptr(vi->stats);
 246         struct sk_buff *skb;
 247         struct page *page;
 248         struct skb_vnet_hdr *hdr;
 249
 250         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 251                 pr_debug("%s: short packet %i\n", dev->name, len);
 252                 dev->stats.rx_length_errors++;
 253                 if (vi->mergeable_rx_bufs || vi->big_packets)
 254                         give_pages(vi, buf);
 255                 else
 256                         dev_kfree_skb(buf);
 257                 return;
 258         }
 259
 260         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
 261                 skb = buf;
 262                 len -= sizeof(struct virtio_net_hdr);
 263                 skb_trim(skb, len);
 264         } else {
 265                 page = buf;
 266                 skb = page_to_skb(vi, page, len);
 267                 if (unlikely(!skb)) {
 268                         dev->stats.rx_dropped++;
 269                         give_pages(vi, page);
 270                         return;
 271                 }
 272                 if (vi->mergeable_rx_bufs)
 273                         if (receive_mergeable(vi, skb)) {
 274                                 dev_kfree_skb(skb);
 275                                 return;
 276                         }
 277         }
 278
 279         hdr = skb_vnet_hdr(skb);
 280         skb->truesize += skb->data_len;
 281
 282         u64_stats_update_begin(&stats->syncp);
 283         stats->rx_bytes += skb->len;
 284         stats->rx_packets++;
 285         u64_stats_update_end(&stats->syncp);
 286
 287         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 288                 pr_debug("Needs csum!\n");
 289                 if (!skb_partial_csum_set(skb,
 290                                           hdr->hdr.csum_start,
 291                                           hdr->hdr.csum_offset))
 292                         goto frame_err;
 293         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 294                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 295         }
 296
 297         skb->protocol = eth_type_trans(skb, dev);
 298         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 299                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 300
 301         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 302                 pr_debug("GSO!\n");
 303                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 304                 case VIRTIO_NET_HDR_GSO_TCPV4:
 305                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 306                         break;
 307                 case VIRTIO_NET_HDR_GSO_UDP:
 308                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 309                         break;
 310                 case VIRTIO_NET_HDR_GSO_TCPV6:
 311                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 312                         break;
 313                 default:
 314                         if (net_ratelimit())
 315                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
 316                                        dev->name, hdr->hdr.gso_type);
 317                         goto frame_err;
 318                 }
 319
 320                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 321                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 322
 323                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 324                 if (skb_shinfo(skb)->gso_size == 0) {
 325                         if (net_ratelimit())
 326                                 printk(KERN_WARNING "%s: zero gso size.\n",
 327                                        dev->name);
 328                         goto frame_err;
 329                 }
 330
 331                 /* Header must be checked, and gso_segs computed. */
 332                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 333                 skb_shinfo(skb)->gso_segs = 0;
 334         }
 335
 336         netif_receive_skb(skb);
 337         return;
 338
 339 frame_err:
 340         dev->stats.rx_frame_errors++;
 341         dev_kfree_skb(skb);
 342 }
 343
 344 static int add_recvbuf_small(struct virtnet_info *vi, gfp_t gfp)
 345 {
 346         struct sk_buff *skb;
 347         struct skb_vnet_hdr *hdr;
 348         int err;
 349
 350         skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
 351         if (unlikely(!skb))
 352                 return -ENOMEM;
 353
 354         skb_put(skb, MAX_PACKET_LEN);
 355
 356         hdr = skb_vnet_hdr(skb);
 357         sg_set_buf(vi->rx_sg, &hdr->hdr, sizeof hdr->hdr);
 358
 359         skb_to_sgvec(skb, vi->rx_sg + 1, 0, skb->len);
 360
 361         err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, 2, skb, gfp);
 362         if (err < 0)
 363                 dev_kfree_skb(skb);
 364
 365         return err;
 366 }
 367
 368 static int add_recvbuf_big(struct virtnet_info *vi, gfp_t gfp)
 369 {
 370         struct page *first, *list = NULL;
 371         char *p;
 372         int i, err, offset;
 373
 374         /* page in vi->rx_sg[MAX_SKB_FRAGS + 1] is list tail */
 375         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 376                 first = get_a_page(vi, gfp);
 377                 if (!first) {
 378                         if (list)
 379                                 give_pages(vi, list);
 380                         return -ENOMEM;
 381                 }
 382                 sg_set_buf(&vi->rx_sg[i], page_address(first), PAGE_SIZE);
 383
 384                 /* chain new page in list head to match sg */
 385                 first->private = (unsigned long)list;
 386                 list = first;
 387         }
 388
 389         first = get_a_page(vi, gfp);
 390         if (!first) {
 391                 give_pages(vi, list);
 392                 return -ENOMEM;
 393         }
 394         p = page_address(first);
 395
 396         /* vi->rx_sg[0], vi->rx_sg[1] share the same page */
 397         /* a separated vi->rx_sg[0] for virtio_net_hdr only due to QEMU bug */
 398         sg_set_buf(&vi->rx_sg[0], p, sizeof(struct virtio_net_hdr));
 399
 400         /* vi->rx_sg[1] for data packet, from offset */
 401         offset = sizeof(struct padded_vnet_hdr);
 402         sg_set_buf(&vi->rx_sg[1], p + offset, PAGE_SIZE - offset);
 403
 404         /* chain first in list head */
 405         first->private = (unsigned long)list;
 406         err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, MAX_SKB_FRAGS + 2,
 407                                     first, gfp);
 408         if (err < 0)
 409                 give_pages(vi, first);
 410
 411         return err;
 412 }
 413
 414 static int add_recvbuf_mergeable(struct virtnet_info *vi, gfp_t gfp)
 415 {
 416         struct page *page;
 417         int err;
 418
 419         page = get_a_page(vi, gfp);
 420         if (!page)
 421                 return -ENOMEM;
 422
 423         sg_init_one(vi->rx_sg, page_address(page), PAGE_SIZE);
 424
 425         err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, 1, page, gfp);
 426         if (err < 0)
 427                 give_pages(vi, page);
 428
 429         return err;
 430 }
 431
 432 /* Returns false if we couldn't fill entirely (OOM). */
 433 static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
 434 {
 435         int err;
 436         bool oom;
 437
 438         do {
 439                 if (vi->mergeable_rx_bufs)
 440                         err = add_recvbuf_mergeable(vi, gfp);
 441                 else if (vi->big_packets)
 442                         err = add_recvbuf_big(vi, gfp);
 443                 else
 444                         err = add_recvbuf_small(vi, gfp);
 445
 446                 oom = err == -ENOMEM;
 447                 if (err < 0)
 448                         break;
 449                 ++vi->num;
 450         } while (err > 0);
 451         if (unlikely(vi->num > vi->max))
 452                 vi->max = vi->num;
 453         virtqueue_kick(vi->rvq);
 454         return !oom;
 455 }
 456
 457 static void skb_recv_done(struct virtqueue *rvq)
 458 {
 459         struct virtnet_info *vi = rvq->vdev->priv;
 460         /* Schedule NAPI, Suppress further interrupts if successful. */
 461         if (napi_schedule_prep(&vi->napi)) {
 462                 virtqueue_disable_cb(rvq);
 463                 __napi_schedule(&vi->napi);
 464         }
 465 }
 466
 467 static void virtnet_napi_enable(struct virtnet_info *vi)
 468 {
 469         napi_enable(&vi->napi);
 470
 471         /* If all buffers were filled by other side before we napi_enabled, we
 472          * won't get another interrupt, so process any outstanding packets
 473          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 474          * We synchronize against interrupts via NAPI_STATE_SCHED */
 475         if (napi_schedule_prep(&vi->napi)) {
 476                 virtqueue_disable_cb(vi->rvq);
 477                 __napi_schedule(&vi->napi);
 478         }
 479 }
 480
 481 static void refill_work(struct work_struct *work)
 482 {
 483         struct virtnet_info *vi;
 484         bool still_empty;
 485
 486         vi = container_of(work, struct virtnet_info, refill.work);
 487         napi_disable(&vi->napi);
 488         still_empty = !try_fill_recv(vi, GFP_KERNEL);
 489         virtnet_napi_enable(vi);
 490
 491         /* In theory, this can happen: if we don't get any buffers in
 492          * we will *never* try to fill again. */
 493         if (still_empty)
 494                 schedule_delayed_work(&vi->refill, HZ/2);
 495 }
 496
 497 static int virtnet_poll(struct napi_struct *napi, int budget)
 498 {
 499         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
 500         void *buf;
 501         unsigned int len, received = 0;
 502
 503 again:
 504         while (received < budget &&
 505                (buf = virtqueue_get_buf(vi->rvq, &len)) != NULL) {
 506                 receive_buf(vi->dev, buf, len);
 507                 --vi->num;
 508                 received++;
 509         }
 510
 511         if (vi->num < vi->max / 2) {
 512                 if (!try_fill_recv(vi, GFP_ATOMIC))
 513                         schedule_delayed_work(&vi->refill, 0);
 514         }
 515
 516         /* Out of packets? */
 517         if (received < budget) {
 518                 napi_complete(napi);
 519                 if (unlikely(!virtqueue_enable_cb(vi->rvq)) &&
 520                     napi_schedule_prep(napi)) {
 521                         virtqueue_disable_cb(vi->rvq);
 522                         __napi_schedule(napi);
 523                         goto again;
 524                 }
 525         }
 526
 527         return received;
 528 }
 529
 530 static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
 531 {
 532         struct sk_buff *skb;
 533         unsigned int len, tot_sgs = 0;
 534         struct virtnet_stats __percpu *stats = this_cpu_ptr(vi->stats);
 535
 536         while ((skb = virtqueue_get_buf(vi->svq, &len)) != NULL) {
 537                 pr_debug("Sent skb %p\n", skb);
 538
 539                 u64_stats_update_begin(&stats->syncp);
 540                 stats->tx_bytes += skb->len;
 541                 stats->tx_packets++;
 542                 u64_stats_update_end(&stats->syncp);
 543
 544                 tot_sgs += skb_vnet_hdr(skb)->num_sg;
 545                 dev_kfree_skb_any(skb);
 546         }
 547         return tot_sgs;
 548 }
 549
 550 static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 551 {
 552         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 553         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 554
 555         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 556
 557         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 558                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 559                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 560                 hdr->hdr.csum_offset = skb->csum_offset;
 561         } else {
 562                 hdr->hdr.flags = 0;
 563                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 564         }
 565
 566         if (skb_is_gso(skb)) {
 567                 hdr->hdr.hdr_len = skb_headlen(skb);
 568                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 569                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 570                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 571                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 572                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 573                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 574                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 575                 else
 576                         BUG();
 577                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 578                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 579         } else {
 580                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 581                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 582         }
 583
 584         hdr->mhdr.num_buffers = 0;
 585
 586         /* Encode metadata header at front. */
 587         if (vi->mergeable_rx_bufs)
 588                 sg_set_buf(vi->tx_sg, &hdr->mhdr, sizeof hdr->mhdr);
 589         else
 590                 sg_set_buf(vi->tx_sg, &hdr->hdr, sizeof hdr->hdr);
 591
 592         hdr->num_sg = skb_to_sgvec(skb, vi->tx_sg + 1, 0, skb->len) + 1;
 593         return virtqueue_add_buf(vi->svq, vi->tx_sg, hdr->num_sg,
 594                                         0, skb);
 595 }
 596
 597 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 598 {
 599         struct virtnet_info *vi = netdev_priv(dev);
 600         int capacity;
 601
 602         /* Free up any pending old buffers before queueing new ones. */
 603         free_old_xmit_skbs(vi);
 604
 605         /* Try to transmit */
 606         capacity = xmit_skb(vi, skb);
 607
 608         /* This can happen with OOM and indirect buffers. */
 609         if (unlikely(capacity < 0)) {
 610                 if (net_ratelimit()) {
 611                         if (likely(capacity == -ENOMEM)) {
 612                                 dev_warn(&dev->dev,
 613                                          "TX queue failure: out of memory\n");
 614                         } else {
 615                                 dev->stats.tx_fifo_errors++;
 616                                 dev_warn(&dev->dev,
 617                                          "Unexpected TX queue failure: %d\n",
 618                                          capacity);
 619                         }
 620                 }
 621                 dev->stats.tx_dropped++;
 622                 kfree_skb(skb);
 623                 return NETDEV_TX_OK;
 624         }
 625         virtqueue_kick(vi->svq);
 626
 627         /* Don't wait up for transmitted skbs to be freed. */
 628         skb_orphan(skb);
 629         nf_reset(skb);
 630
 631         /* Apparently nice girls don't return TX_BUSY; stop the queue
 632          * before it gets out of hand.  Naturally, this wastes entries. */
 633         if (capacity < 2+MAX_SKB_FRAGS) {
 634                 netif_stop_queue(dev);
 635                 if (unlikely(!virtqueue_enable_cb_delayed(vi->svq))) {
 636                         /* More just got used, free them then recheck. */
 637                         capacity += free_old_xmit_skbs(vi);
 638                         if (capacity >= 2+MAX_SKB_FRAGS) {
 639                                 netif_start_queue(dev);
 640                                 virtqueue_disable_cb(vi->svq);
 641                         }
 642                 }
 643         }
 644
 645         return NETDEV_TX_OK;
 646 }
 647
 648 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 649 {
 650         struct virtnet_info *vi = netdev_priv(dev);
 651         struct virtio_device *vdev = vi->vdev;
 652         int ret;
 653
 654         ret = eth_mac_addr(dev, p);
 655         if (ret)
 656                 return ret;
 657
 658         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
 659                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 660                                   dev->dev_addr, dev->addr_len);
 661
 662         return 0;
 663 }
 664
 665 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 666                                                struct rtnl_link_stats64 *tot)
 667 {
 668         struct virtnet_info *vi = netdev_priv(dev);
 669         int cpu;
 670         unsigned int start;
 671
 672         for_each_possible_cpu(cpu) {
 673                 struct virtnet_stats __percpu *stats
 674                         = per_cpu_ptr(vi->stats, cpu);
 675                 u64 tpackets, tbytes, rpackets, rbytes;
 676
 677                 do {
 678                         start = u64_stats_fetch_begin(&stats->syncp);
 679                         tpackets = stats->tx_packets;
 680                         tbytes   = stats->tx_bytes;
 681                         rpackets = stats->rx_packets;
 682                         rbytes   = stats->rx_bytes;
 683                 } while (u64_stats_fetch_retry(&stats->syncp, start));
 684
 685                 tot->rx_packets += rpackets;
 686                 tot->tx_packets += tpackets;
 687                 tot->rx_bytes   += rbytes;
 688                 tot->tx_bytes   += tbytes;
 689         }
 690
 691         tot->tx_dropped = dev->stats.tx_dropped;
 692         tot->rx_dropped = dev->stats.rx_dropped;
 693         tot->rx_length_errors = dev->stats.rx_length_errors;
 694         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 695
 696         return tot;
 697 }
 698
 699 #ifdef CONFIG_NET_POLL_CONTROLLER
 700 static void virtnet_netpoll(struct net_device *dev)
 701 {
 702         struct virtnet_info *vi = netdev_priv(dev);
 703
 704         napi_schedule(&vi->napi);
 705 }
 706 #endif
 707
 708 static int virtnet_open(struct net_device *dev)
 709 {
 710         struct virtnet_info *vi = netdev_priv(dev);
 711
 712         virtnet_napi_enable(vi);
 713         return 0;
 714 }
 715
 716 /*
 717  * Send command via the control virtqueue and check status.  Commands
 718  * supported by the hypervisor, as indicated by feature bits, should
 719  * never fail unless improperly formated.
 720  */
 721 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 722                                  struct scatterlist *data, int out, int in)
 723 {
 724         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
 725         struct virtio_net_ctrl_hdr ctrl;
 726         virtio_net_ctrl_ack status = ~0;
 727         unsigned int tmp;
 728         int i;
 729
 730         /* Caller should know better */
 731         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
 732                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
 733
 734         out++; /* Add header */
 735         in++; /* Add return status */
 736
 737         ctrl.class = class;
 738         ctrl.cmd = cmd;
 739
 740         sg_init_table(sg, out + in);
 741
 742         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
 743         for_each_sg(data, s, out + in - 2, i)
 744                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
 745         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
 746
 747         BUG_ON(virtqueue_add_buf(vi->cvq, sg, out, in, vi) < 0);
 748
 749         virtqueue_kick(vi->cvq);
 750
 751         /*
 752          * Spin for a response, the kick causes an ioport write, trapping
 753          * into the hypervisor, so the request should be handled immediately.
 754          */
 755         while (!virtqueue_get_buf(vi->cvq, &tmp))
 756                 cpu_relax();
 757
 758         return status == VIRTIO_NET_OK;
 759 }
 760
 761 static int virtnet_close(struct net_device *dev)
 762 {
 763         struct virtnet_info *vi = netdev_priv(dev);
 764
 765         napi_disable(&vi->napi);
 766
 767         return 0;
 768 }
 769
 770 static void virtnet_set_rx_mode(struct net_device *dev)
 771 {
 772         struct virtnet_info *vi = netdev_priv(dev);
 773         struct scatterlist sg[2];
 774         u8 promisc, allmulti;
 775         struct virtio_net_ctrl_mac *mac_data;
 776         struct netdev_hw_addr *ha;
 777         int uc_count;
 778         int mc_count;
 779         void *buf;
 780         int i;
 781
 782         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 783         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 784                 return;
 785
 786         promisc = ((dev->flags & IFF_PROMISC) != 0);
 787         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 788
 789         sg_init_one(sg, &promisc, sizeof(promisc));
 790
 791         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 792                                   VIRTIO_NET_CTRL_RX_PROMISC,
 793                                   sg, 1, 0))
 794                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 795                          promisc ? "en" : "dis");
 796
 797         sg_init_one(sg, &allmulti, sizeof(allmulti));
 798
 799         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 800                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 801                                   sg, 1, 0))
 802                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
 803                          allmulti ? "en" : "dis");
 804
 805         uc_count = netdev_uc_count(dev);
 806         mc_count = netdev_mc_count(dev);
 807         /* MAC filter - use one buffer for both lists */
 808         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
 809                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
 810         mac_data = buf;
 811         if (!buf) {
 812                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
 813                 return;
 814         }
 815
 816         sg_init_table(sg, 2);
 817
 818         /* Store the unicast list and count in the front of the buffer */
 819         mac_data->entries = uc_count;
 820         i = 0;
 821         netdev_for_each_uc_addr(ha, dev)
 822                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 823
 824         sg_set_buf(&sg[0], mac_data,
 825                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
 826
 827         /* multicast list and count fill the end */
 828         mac_data = (void *)&mac_data->macs[uc_count][0];
 829
 830         mac_data->entries = mc_count;
 831         i = 0;
 832         netdev_for_each_mc_addr(ha, dev)
 833                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 834
 835         sg_set_buf(&sg[1], mac_data,
 836                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
 837
 838         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 839                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
 840                                   sg, 2, 0))
 841                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
 842
 843         kfree(buf);
 844 }
 845
 846 static void virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
 847 {
 848         struct virtnet_info *vi = netdev_priv(dev);
 849         struct scatterlist sg;
 850
 851         sg_init_one(&sg, &vid, sizeof(vid));
 852
 853         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 854                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
 855                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
 856 }
 857
 858 static void virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
 859 {
 860         struct virtnet_info *vi = netdev_priv(dev);
 861         struct scatterlist sg;
 862
 863         sg_init_one(&sg, &vid, sizeof(vid));
 864
 865         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 866                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
 867                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
 868 }
 869
 870 static const struct ethtool_ops virtnet_ethtool_ops = {
 871         .get_link = ethtool_op_get_link,
 872 };
 873
 874 #define MIN_MTU 68
 875 #define MAX_MTU 65535
 876
 877 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
 878 {
 879         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
 880                 return -EINVAL;
 881         dev->mtu = new_mtu;
 882         return 0;
 883 }
 884
 885 static const struct net_device_ops virtnet_netdev = {
 886         .ndo_open            = virtnet_open,
 887         .ndo_stop            = virtnet_close,
 888         .ndo_start_xmit      = start_xmit,
 889         .ndo_validate_addr   = eth_validate_addr,
 890         .ndo_set_mac_address = virtnet_set_mac_address,
 891         .ndo_set_rx_mode     = virtnet_set_rx_mode,
 892         .ndo_change_mtu      = virtnet_change_mtu,
 893         .ndo_get_stats64     = virtnet_stats,
 894         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
 895         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
 896 #ifdef CONFIG_NET_POLL_CONTROLLER
 897         .ndo_poll_controller = virtnet_netpoll,
 898 #endif
 899 };
 900
 901 static void virtnet_update_status(struct virtnet_info *vi)
 902 {
 903         u16 v;
 904
 905         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS))
 906                 return;
 907
 908         vi->vdev->config->get(vi->vdev,
 909                               offsetof(struct virtio_net_config, status),
 910                               &v, sizeof(v));
 911
 912         /* Ignore unknown (future) status bits */
 913         v &= VIRTIO_NET_S_LINK_UP;
 914
 915         if (vi->status == v)
 916                 return;
 917
 918         vi->status = v;
 919
 920         if (vi->status & VIRTIO_NET_S_LINK_UP) {
 921                 netif_carrier_on(vi->dev);
 922                 netif_wake_queue(vi->dev);
 923         } else {
 924                 netif_carrier_off(vi->dev);
 925                 netif_stop_queue(vi->dev);
 926         }
 927 }
 928
 929 static void virtnet_config_changed(struct virtio_device *vdev)
 930 {
 931         struct virtnet_info *vi = vdev->priv;
 932
 933         virtnet_update_status(vi);
 934 }
 935
 936 static int virtnet_probe(struct virtio_device *vdev)
 937 {
 938         int err;
 939         struct net_device *dev;
 940         struct virtnet_info *vi;
 941         struct virtqueue *vqs[3];
 942         vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
 943         const char *names[] = { "input", "output", "control" };
 944         int nvqs;
 945
 946         /* Allocate ourselves a network device with room for our info */
 947         dev = alloc_etherdev(sizeof(struct virtnet_info));
 948         if (!dev)
 949                 return -ENOMEM;
 950
 951         /* Set up network device as normal. */
 952         dev->netdev_ops = &virtnet_netdev;
 953         dev->features = NETIF_F_HIGHDMA;
 954
 955         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
 956         SET_NETDEV_DEV(dev, &vdev->dev);
 957
 958         /* Do we support "hardware" checksums? */
 959         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
 960                 /* This opens up the world of extra features. */
 961                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
 962                 if (csum)
 963                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
 964
 965                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
 966                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
 967                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
 968                 }
 969                 /* Individual feature bits: what can host handle? */
 970                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
 971                         dev->hw_features |= NETIF_F_TSO;
 972                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
 973                         dev->hw_features |= NETIF_F_TSO6;
 974                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
 975                         dev->hw_features |= NETIF_F_TSO_ECN;
 976                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
 977                         dev->hw_features |= NETIF_F_UFO;
 978
 979                 if (gso)
 980                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
 981                 /* (!csum && gso) case will be fixed by register_netdev() */
 982         }
 983
 984         /* Configuration may specify what MAC to use.  Otherwise random. */
 985         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 986                 vdev->config->get(vdev,
 987                                   offsetof(struct virtio_net_config, mac),
 988                                   dev->dev_addr, dev->addr_len);
 989         } else
 990                 random_ether_addr(dev->dev_addr);
 991
 992         /* Set up our device-specific information */
 993         vi = netdev_priv(dev);
 994         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
 995         vi->dev = dev;
 996         vi->vdev = vdev;
 997         vdev->priv = vi;
 998         vi->pages = NULL;
 999         vi->stats = alloc_percpu(struct virtnet_stats);
1000         err = -ENOMEM;
1001         if (vi->stats == NULL)
1002                 goto free;
1003
1004         INIT_DELAYED_WORK(&vi->refill, refill_work);
1005         sg_init_table(vi->rx_sg, ARRAY_SIZE(vi->rx_sg));
1006         sg_init_table(vi->tx_sg, ARRAY_SIZE(vi->tx_sg));
1007
1008         /* If we can receive ANY GSO packets, we must allocate large ones. */
1009         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1010             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1011             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1012                 vi->big_packets = true;
1013
1014         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1015                 vi->mergeable_rx_bufs = true;
1016
1017         /* We expect two virtqueues, receive then send,
1018          * and optionally control. */
1019         nvqs = virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ? 3 : 2;
1020
1021         err = vdev->config->find_vqs(vdev, nvqs, vqs, callbacks, names);
1022         if (err)
1023                 goto free_stats;
1024
1025         vi->rvq = vqs[0];
1026         vi->svq = vqs[1];
1027
1028         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)) {
1029                 vi->cvq = vqs[2];
1030
1031                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1032                         dev->features |= NETIF_F_HW_VLAN_FILTER;
1033         }
1034
1035         err = register_netdev(dev);
1036         if (err) {
1037                 pr_debug("virtio_net: registering device failed\n");
1038                 goto free_vqs;
1039         }
1040
1041         /* Last of all, set up some receive buffers. */
1042         try_fill_recv(vi, GFP_KERNEL);
1043
1044         /* If we didn't even get one input buffer, we're useless. */
1045         if (vi->num == 0) {
1046                 err = -ENOMEM;
1047                 goto unregister;
1048         }
1049
1050         /* Assume link up if device can't report link status,
1051            otherwise get link status from config. */
1052         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1053                 netif_carrier_off(dev);
1054                 virtnet_update_status(vi);
1055         } else {
1056                 vi->status = VIRTIO_NET_S_LINK_UP;
1057                 netif_carrier_on(dev);
1058         }
1059
1060         pr_debug("virtnet: registered device %s\n", dev->name);
1061         return 0;
1062
1063 unregister:
1064         unregister_netdev(dev);
1065         cancel_delayed_work_sync(&vi->refill);
1066 free_vqs:
1067         vdev->config->del_vqs(vdev);
1068 free_stats:
1069         free_percpu(vi->stats);
1070 free:
1071         free_netdev(dev);
1072         return err;
1073 }
1074
1075 static void free_unused_bufs(struct virtnet_info *vi)
1076 {
1077         void *buf;
1078         while (1) {
1079                 buf = virtqueue_detach_unused_buf(vi->svq);
1080                 if (!buf)
1081                         break;
1082                 dev_kfree_skb(buf);
1083         }
1084         while (1) {
1085                 buf = virtqueue_detach_unused_buf(vi->rvq);
1086                 if (!buf)
1087                         break;
1088                 if (vi->mergeable_rx_bufs || vi->big_packets)
1089                         give_pages(vi, buf);
1090                 else
1091                         dev_kfree_skb(buf);
1092                 --vi->num;
1093         }
1094         BUG_ON(vi->num != 0);
1095 }
1096
1097 static void __devexit virtnet_remove(struct virtio_device *vdev)
1098 {
1099         struct virtnet_info *vi = vdev->priv;
1100
1101         /* Stop all the virtqueues. */
1102         vdev->config->reset(vdev);
1103
1104
1105         unregister_netdev(vi->dev);
1106         cancel_delayed_work_sync(&vi->refill);
1107
1108         /* Free unused buffers in both send and recv, if any. */
1109         free_unused_bufs(vi);
1110
1111         vdev->config->del_vqs(vi->vdev);
1112
1113         while (vi->pages)
1114                 __free_pages(get_a_page(vi, GFP_KERNEL), 0);
1115
1116         free_percpu(vi->stats);
1117         free_netdev(vi->dev);
1118 }
1119
1120 static struct virtio_device_id id_table[] = {
1121         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1122         { 0 },
1123 };
1124
1125 static unsigned int features[] = {
1126         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1127         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1128         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1129         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1130         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1131         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1132         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1133 };
1134
1135 static struct virtio_driver virtio_net_driver = {
1136         .feature_table = features,
1137         .feature_table_size = ARRAY_SIZE(features),
1138         .driver.name =  KBUILD_MODNAME,
1139         .driver.owner = THIS_MODULE,
1140         .id_table =     id_table,
1141         .probe =        virtnet_probe,
1142         .remove =       __devexit_p(virtnet_remove),
1143         .config_changed = virtnet_config_changed,
1144 };
1145
1146 static int __init init(void)
1147 {
1148         return register_virtio_driver(&virtio_net_driver);
1149 }
1150
1151 static void __exit fini(void)
1152 {
1153         unregister_virtio_driver(&virtio_net_driver);
1154 }
1155 module_init(init);
1156 module_exit(fini);
1157
1158 MODULE_DEVICE_TABLE(virtio, id_table);
1159 MODULE_DESCRIPTION("Virtio network driver");
1160 MODULE_LICENSE("GPL");