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