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