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
  29 static int napi_weight = 128;
  30 module_param(napi_weight, int, 0444);
  31
  32 static int csum = 1, gso = 1;
  33 module_param(csum, bool, 0444);
  34 module_param(gso, bool, 0444);
  35
  36 /* FIXME: MTU in config. */
  37 #define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  38 #define GOOD_COPY_LEN   128
  39
  40 #define VIRTNET_SEND_COMMAND_SG_MAX    2
  41
  42 struct virtnet_info
  43 {
  44         struct virtio_device *vdev;
  45         struct virtqueue *rvq, *svq, *cvq;
  46         struct net_device *dev;
  47         struct napi_struct napi;
  48         unsigned int status;
  49
  50         /* Number of input buffers, and max we've ever had. */
  51         unsigned int num, max;
  52
  53         /* I like... big packets and I cannot lie! */
  54         bool big_packets;
  55
  56         /* Host will merge rx buffers for big packets (shake it! shake it!) */
  57         bool mergeable_rx_bufs;
  58
  59         /* Work struct for refilling if we run low on memory. */
  60         struct delayed_work refill;
  61
  62         /* Chain pages by the private ptr. */
  63         struct page *pages;
  64 };
  65
  66 struct skb_vnet_hdr {
  67         union {
  68                 struct virtio_net_hdr hdr;
  69                 struct virtio_net_hdr_mrg_rxbuf mhdr;
  70         };
  71         unsigned int num_sg;
  72 };
  73
  74 struct padded_vnet_hdr {
  75         struct virtio_net_hdr hdr;
  76         /*
  77          * virtio_net_hdr should be in a separated sg buffer because of a
  78          * QEMU bug, and data sg buffer shares same page with this header sg.
  79          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
  80          */
  81         char padding[6];
  82 };
  83
  84 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
  85 {
  86         return (struct skb_vnet_hdr *)skb->cb;
  87 }
  88
  89 /*
  90  * private is used to chain pages for big packets, put the whole
  91  * most recent used list in the beginning for reuse
  92  */
  93 static void give_pages(struct virtnet_info *vi, struct page *page)
  94 {
  95         struct page *end;
  96
  97         /* Find end of list, sew whole thing into vi->pages. */
  98         for (end = page; end->private; end = (struct page *)end->private);
  99         end->private = (unsigned long)vi->pages;
 100         vi->pages = page;
 101 }
 102
 103 static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
 104 {
 105         struct page *p = vi->pages;
 106
 107         if (p) {
 108                 vi->pages = (struct page *)p->private;
 109                 /* clear private here, it is used to chain pages */
 110                 p->private = 0;
 111         } else
 112                 p = alloc_page(gfp_mask);
 113         return p;
 114 }
 115
 116 static void skb_xmit_done(struct virtqueue *svq)
 117 {
 118         struct virtnet_info *vi = svq->vdev->priv;
 119
 120         /* Suppress further interrupts. */
 121         svq->vq_ops->disable_cb(svq);
 122
 123         /* We were probably waiting for more output buffers. */
 124         netif_wake_queue(vi->dev);
 125 }
 126
 127 static void set_skb_frag(struct sk_buff *skb, struct page *page,
 128                          unsigned int offset, unsigned int *len)
 129 {
 130         int i = skb_shinfo(skb)->nr_frags;
 131         skb_frag_t *f;
 132
 133         f = &skb_shinfo(skb)->frags[i];
 134         f->size = min((unsigned)PAGE_SIZE - offset, *len);
 135         f->page_offset = offset;
 136         f->page = page;
 137
 138         skb->data_len += f->size;
 139         skb->len += f->size;
 140         skb_shinfo(skb)->nr_frags++;
 141         *len -= f->size;
 142 }
 143
 144 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 145                                    struct page *page, unsigned int len)
 146 {
 147         struct sk_buff *skb;
 148         struct skb_vnet_hdr *hdr;
 149         unsigned int copy, hdr_len, offset;
 150         char *p;
 151
 152         p = page_address(page);
 153
 154         /* copy small packet so we can reuse these pages for small data */
 155         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 156         if (unlikely(!skb))
 157                 return NULL;
 158
 159         hdr = skb_vnet_hdr(skb);
 160
 161         if (vi->mergeable_rx_bufs) {
 162                 hdr_len = sizeof hdr->mhdr;
 163                 offset = hdr_len;
 164         } else {
 165                 hdr_len = sizeof hdr->hdr;
 166                 offset = sizeof(struct padded_vnet_hdr);
 167         }
 168
 169         memcpy(hdr, p, hdr_len);
 170
 171         len -= hdr_len;
 172         p += offset;
 173
 174         copy = len;
 175         if (copy > skb_tailroom(skb))
 176                 copy = skb_tailroom(skb);
 177         memcpy(skb_put(skb, copy), p, copy);
 178
 179         len -= copy;
 180         offset += copy;
 181
 182         while (len) {
 183                 set_skb_frag(skb, page, offset, &len);
 184                 page = (struct page *)page->private;
 185                 offset = 0;
 186         }
 187
 188         if (page)
 189                 give_pages(vi, page);
 190
 191         return skb;
 192 }
 193
 194 static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
 195 {
 196         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 197         struct page *page;
 198         int num_buf, i, len;
 199
 200         num_buf = hdr->mhdr.num_buffers;
 201         while (--num_buf) {
 202                 i = skb_shinfo(skb)->nr_frags;
 203                 if (i >= MAX_SKB_FRAGS) {
 204                         pr_debug("%s: packet too long\n", skb->dev->name);
 205                         skb->dev->stats.rx_length_errors++;
 206                         return -EINVAL;
 207                 }
 208
 209                 page = vi->rvq->vq_ops->get_buf(vi->rvq, &len);
 210                 if (!page) {
 211                         pr_debug("%s: rx error: %d buffers missing\n",
 212                                  skb->dev->name, hdr->mhdr.num_buffers);
 213                         skb->dev->stats.rx_length_errors++;
 214                         return -EINVAL;
 215                 }
 216                 if (len > PAGE_SIZE)
 217                         len = PAGE_SIZE;
 218
 219                 set_skb_frag(skb, page, 0, &len);
 220
 221                 --vi->num;
 222         }
 223         return 0;
 224 }
 225
 226 static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
 227 {
 228         struct virtnet_info *vi = netdev_priv(dev);
 229         struct sk_buff *skb;
 230         struct page *page;
 231         struct skb_vnet_hdr *hdr;
 232
 233         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 234                 pr_debug("%s: short packet %i\n", dev->name, len);
 235                 dev->stats.rx_length_errors++;
 236                 if (vi->mergeable_rx_bufs || vi->big_packets)
 237                         give_pages(vi, buf);
 238                 else
 239                         dev_kfree_skb(buf);
 240                 return;
 241         }
 242
 243         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
 244                 skb = buf;
 245                 len -= sizeof(struct virtio_net_hdr);
 246                 skb_trim(skb, len);
 247         } else {
 248                 page = buf;
 249                 skb = page_to_skb(vi, page, len);
 250                 if (unlikely(!skb)) {
 251                         dev->stats.rx_dropped++;
 252                         give_pages(vi, page);
 253                         return;
 254                 }
 255                 if (vi->mergeable_rx_bufs)
 256                         if (receive_mergeable(vi, skb)) {
 257                                 dev_kfree_skb(skb);
 258                                 return;
 259                         }
 260         }
 261
 262         hdr = skb_vnet_hdr(skb);
 263         skb->truesize += skb->data_len;
 264         dev->stats.rx_bytes += skb->len;
 265         dev->stats.rx_packets++;
 266
 267         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 268                 pr_debug("Needs csum!\n");
 269                 if (!skb_partial_csum_set(skb,
 270                                           hdr->hdr.csum_start,
 271                                           hdr->hdr.csum_offset))
 272                         goto frame_err;
 273         }
 274
 275         skb->protocol = eth_type_trans(skb, dev);
 276         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 277                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 278
 279         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 280                 pr_debug("GSO!\n");
 281                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 282                 case VIRTIO_NET_HDR_GSO_TCPV4:
 283                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 284                         break;
 285                 case VIRTIO_NET_HDR_GSO_UDP:
 286                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 287                         break;
 288                 case VIRTIO_NET_HDR_GSO_TCPV6:
 289                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 290                         break;
 291                 default:
 292                         if (net_ratelimit())
 293                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
 294                                        dev->name, hdr->hdr.gso_type);
 295                         goto frame_err;
 296                 }
 297
 298                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 299                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 300
 301                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 302                 if (skb_shinfo(skb)->gso_size == 0) {
 303                         if (net_ratelimit())
 304                                 printk(KERN_WARNING "%s: zero gso size.\n",
 305                                        dev->name);
 306                         goto frame_err;
 307                 }
 308
 309                 /* Header must be checked, and gso_segs computed. */
 310                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 311                 skb_shinfo(skb)->gso_segs = 0;
 312         }
 313
 314         netif_receive_skb(skb);
 315         return;
 316
 317 frame_err:
 318         dev->stats.rx_frame_errors++;
 319         dev_kfree_skb(skb);
 320 }
 321
 322 static int add_recvbuf_small(struct virtnet_info *vi, gfp_t gfp)
 323 {
 324         struct sk_buff *skb;
 325         struct skb_vnet_hdr *hdr;
 326         struct scatterlist sg[2];
 327         int err;
 328
 329         sg_init_table(sg, 2);
 330         skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
 331         if (unlikely(!skb))
 332                 return -ENOMEM;
 333
 334         skb_put(skb, MAX_PACKET_LEN);
 335
 336         hdr = skb_vnet_hdr(skb);
 337         sg_set_buf(sg, &hdr->hdr, sizeof hdr->hdr);
 338
 339         skb_to_sgvec(skb, sg + 1, 0, skb->len);
 340
 341         err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, 2, skb);
 342         if (err < 0)
 343                 dev_kfree_skb(skb);
 344
 345         return err;
 346 }
 347
 348 static int add_recvbuf_big(struct virtnet_info *vi, gfp_t gfp)
 349 {
 350         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 351         struct page *first, *list = NULL;
 352         char *p;
 353         int i, err, offset;
 354
 355         sg_init_table(sg, MAX_SKB_FRAGS + 2);
 356         /* page in sg[MAX_SKB_FRAGS + 1] is list tail */
 357         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 358                 first = get_a_page(vi, gfp);
 359                 if (!first) {
 360                         if (list)
 361                                 give_pages(vi, list);
 362                         return -ENOMEM;
 363                 }
 364                 sg_set_buf(&sg[i], page_address(first), PAGE_SIZE);
 365
 366                 /* chain new page in list head to match sg */
 367                 first->private = (unsigned long)list;
 368                 list = first;
 369         }
 370
 371         first = get_a_page(vi, gfp);
 372         if (!first) {
 373                 give_pages(vi, list);
 374                 return -ENOMEM;
 375         }
 376         p = page_address(first);
 377
 378         /* sg[0], sg[1] share the same page */
 379         /* a separated sg[0] for  virtio_net_hdr only during to QEMU bug*/
 380         sg_set_buf(&sg[0], p, sizeof(struct virtio_net_hdr));
 381
 382         /* sg[1] for data packet, from offset */
 383         offset = sizeof(struct padded_vnet_hdr);
 384         sg_set_buf(&sg[1], p + offset, PAGE_SIZE - offset);
 385
 386         /* chain first in list head */
 387         first->private = (unsigned long)list;
 388         err = vi->rvq->vq_ops->add_buf(vi->rvq, sg, 0, MAX_SKB_FRAGS + 2,
 389                                        first);
 390         if (err < 0)
 391                 give_pages(vi, first);
 392
 393         return err;
 394 }
 395
 396 static int add_recvbuf_mergeable(struct virtnet_info *vi, gfp_t gfp)
 397 {
 398         struct page *page;
 399         struct scatterlist sg;
 400         int err;
 401
 402         page = get_a_page(vi, gfp);
 403         if (!page)
 404                 return -ENOMEM;
 405
 406         sg_init_one(&sg, page_address(page), PAGE_SIZE);
 407
 408         err = vi->rvq->vq_ops->add_buf(vi->rvq, &sg, 0, 1, page);
 409         if (err < 0)
 410                 give_pages(vi, page);
 411
 412         return err;
 413 }
 414
 415 /* Returns false if we couldn't fill entirely (OOM). */
 416 static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
 417 {
 418         int err;
 419         bool oom = false;
 420
 421         do {
 422                 if (vi->mergeable_rx_bufs)
 423                         err = add_recvbuf_mergeable(vi, gfp);
 424                 else if (vi->big_packets)
 425                         err = add_recvbuf_big(vi, gfp);
 426                 else
 427                         err = add_recvbuf_small(vi, gfp);
 428
 429                 if (err < 0) {
 430                         oom = true;
 431                         break;
 432                 }
 433                 ++vi->num;
 434         } while (err > 0);
 435         if (unlikely(vi->num > vi->max))
 436                 vi->max = vi->num;
 437         vi->rvq->vq_ops->kick(vi->rvq);
 438         return !oom;
 439 }
 440
 441 static void skb_recv_done(struct virtqueue *rvq)
 442 {
 443         struct virtnet_info *vi = rvq->vdev->priv;
 444         /* Schedule NAPI, Suppress further interrupts if successful. */
 445         if (napi_schedule_prep(&vi->napi)) {
 446                 rvq->vq_ops->disable_cb(rvq);
 447                 __napi_schedule(&vi->napi);
 448         }
 449 }
 450
 451 static void refill_work(struct work_struct *work)
 452 {
 453         struct virtnet_info *vi;
 454         bool still_empty;
 455
 456         vi = container_of(work, struct virtnet_info, refill.work);
 457         napi_disable(&vi->napi);
 458         still_empty = !try_fill_recv(vi, GFP_KERNEL);
 459         napi_enable(&vi->napi);
 460
 461         /* In theory, this can happen: if we don't get any buffers in
 462          * we will *never* try to fill again. */
 463         if (still_empty)
 464                 schedule_delayed_work(&vi->refill, HZ/2);
 465 }
 466
 467 static int virtnet_poll(struct napi_struct *napi, int budget)
 468 {
 469         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
 470         void *buf;
 471         unsigned int len, received = 0;
 472
 473 again:
 474         while (received < budget &&
 475                (buf = vi->rvq->vq_ops->get_buf(vi->rvq, &len)) != NULL) {
 476                 receive_buf(vi->dev, buf, len);
 477                 --vi->num;
 478                 received++;
 479         }
 480
 481         if (vi->num < vi->max / 2) {
 482                 if (!try_fill_recv(vi, GFP_ATOMIC))
 483                         schedule_delayed_work(&vi->refill, 0);
 484         }
 485
 486         /* Out of packets? */
 487         if (received < budget) {
 488                 napi_complete(napi);
 489                 if (unlikely(!vi->rvq->vq_ops->enable_cb(vi->rvq)) &&
 490                     napi_schedule_prep(napi)) {
 491                         vi->rvq->vq_ops->disable_cb(vi->rvq);
 492                         __napi_schedule(napi);
 493                         goto again;
 494                 }
 495         }
 496
 497         return received;
 498 }
 499
 500 static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
 501 {
 502         struct sk_buff *skb;
 503         unsigned int len, tot_sgs = 0;
 504
 505         while ((skb = vi->svq->vq_ops->get_buf(vi->svq, &len)) != NULL) {
 506                 pr_debug("Sent skb %p\n", skb);
 507                 vi->dev->stats.tx_bytes += skb->len;
 508                 vi->dev->stats.tx_packets++;
 509                 tot_sgs += skb_vnet_hdr(skb)->num_sg;
 510                 dev_kfree_skb_any(skb);
 511         }
 512         return tot_sgs;
 513 }
 514
 515 static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 516 {
 517         struct scatterlist sg[2+MAX_SKB_FRAGS];
 518         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 519         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 520
 521         sg_init_table(sg, 2+MAX_SKB_FRAGS);
 522
 523         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 524
 525         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 526                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 527                 hdr->hdr.csum_start = skb->csum_start - skb_headroom(skb);
 528                 hdr->hdr.csum_offset = skb->csum_offset;
 529         } else {
 530                 hdr->hdr.flags = 0;
 531                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 532         }
 533
 534         if (skb_is_gso(skb)) {
 535                 hdr->hdr.hdr_len = skb_headlen(skb);
 536                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 537                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 538                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 539                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 540                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 541                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 542                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 543                 else
 544                         BUG();
 545                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 546                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 547         } else {
 548                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 549                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 550         }
 551
 552         hdr->mhdr.num_buffers = 0;
 553
 554         /* Encode metadata header at front. */
 555         if (vi->mergeable_rx_bufs)
 556                 sg_set_buf(sg, &hdr->mhdr, sizeof hdr->mhdr);
 557         else
 558                 sg_set_buf(sg, &hdr->hdr, sizeof hdr->hdr);
 559
 560         hdr->num_sg = skb_to_sgvec(skb, sg+1, 0, skb->len) + 1;
 561         return vi->svq->vq_ops->add_buf(vi->svq, sg, hdr->num_sg, 0, skb);
 562 }
 563
 564 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 565 {
 566         struct virtnet_info *vi = netdev_priv(dev);
 567         int capacity;
 568
 569 again:
 570         /* Free up any pending old buffers before queueing new ones. */
 571         free_old_xmit_skbs(vi);
 572
 573         /* Try to transmit */
 574         capacity = xmit_skb(vi, skb);
 575
 576         /* This can happen with OOM and indirect buffers. */
 577         if (unlikely(capacity < 0)) {
 578                 netif_stop_queue(dev);
 579                 dev_warn(&dev->dev, "Unexpected full queue\n");
 580                 if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
 581                         vi->svq->vq_ops->disable_cb(vi->svq);
 582                         netif_start_queue(dev);
 583                         goto again;
 584                 }
 585                 return NETDEV_TX_BUSY;
 586         }
 587         vi->svq->vq_ops->kick(vi->svq);
 588
 589         /* Don't wait up for transmitted skbs to be freed. */
 590         skb_orphan(skb);
 591         nf_reset(skb);
 592
 593         /* Apparently nice girls don't return TX_BUSY; stop the queue
 594          * before it gets out of hand.  Naturally, this wastes entries. */
 595         if (capacity < 2+MAX_SKB_FRAGS) {
 596                 netif_stop_queue(dev);
 597                 if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
 598                         /* More just got used, free them then recheck. */
 599                         capacity += free_old_xmit_skbs(vi);
 600                         if (capacity >= 2+MAX_SKB_FRAGS) {
 601                                 netif_start_queue(dev);
 602                                 vi->svq->vq_ops->disable_cb(vi->svq);
 603                         }
 604                 }
 605         }
 606
 607         return NETDEV_TX_OK;
 608 }
 609
 610 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 611 {
 612         struct virtnet_info *vi = netdev_priv(dev);
 613         struct virtio_device *vdev = vi->vdev;
 614         int ret;
 615
 616         ret = eth_mac_addr(dev, p);
 617         if (ret)
 618                 return ret;
 619
 620         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
 621                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 622                                   dev->dev_addr, dev->addr_len);
 623
 624         return 0;
 625 }
 626
 627 #ifdef CONFIG_NET_POLL_CONTROLLER
 628 static void virtnet_netpoll(struct net_device *dev)
 629 {
 630         struct virtnet_info *vi = netdev_priv(dev);
 631
 632         napi_schedule(&vi->napi);
 633 }
 634 #endif
 635
 636 static int virtnet_open(struct net_device *dev)
 637 {
 638         struct virtnet_info *vi = netdev_priv(dev);
 639
 640         napi_enable(&vi->napi);
 641
 642         /* If all buffers were filled by other side before we napi_enabled, we
 643          * won't get another interrupt, so process any outstanding packets
 644          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 645          * We synchronize against interrupts via NAPI_STATE_SCHED */
 646         if (napi_schedule_prep(&vi->napi)) {
 647                 vi->rvq->vq_ops->disable_cb(vi->rvq);
 648                 __napi_schedule(&vi->napi);
 649         }
 650         return 0;
 651 }
 652
 653 /*
 654  * Send command via the control virtqueue and check status.  Commands
 655  * supported by the hypervisor, as indicated by feature bits, should
 656  * never fail unless improperly formated.
 657  */
 658 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 659                                  struct scatterlist *data, int out, int in)
 660 {
 661         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
 662         struct virtio_net_ctrl_hdr ctrl;
 663         virtio_net_ctrl_ack status = ~0;
 664         unsigned int tmp;
 665         int i;
 666
 667         /* Caller should know better */
 668         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
 669                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
 670
 671         out++; /* Add header */
 672         in++; /* Add return status */
 673
 674         ctrl.class = class;
 675         ctrl.cmd = cmd;
 676
 677         sg_init_table(sg, out + in);
 678
 679         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
 680         for_each_sg(data, s, out + in - 2, i)
 681                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
 682         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
 683
 684         BUG_ON(vi->cvq->vq_ops->add_buf(vi->cvq, sg, out, in, vi) < 0);
 685
 686         vi->cvq->vq_ops->kick(vi->cvq);
 687
 688         /*
 689          * Spin for a response, the kick causes an ioport write, trapping
 690          * into the hypervisor, so the request should be handled immediately.
 691          */
 692         while (!vi->cvq->vq_ops->get_buf(vi->cvq, &tmp))
 693                 cpu_relax();
 694
 695         return status == VIRTIO_NET_OK;
 696 }
 697
 698 static int virtnet_close(struct net_device *dev)
 699 {
 700         struct virtnet_info *vi = netdev_priv(dev);
 701
 702         napi_disable(&vi->napi);
 703
 704         return 0;
 705 }
 706
 707 static int virtnet_set_tx_csum(struct net_device *dev, u32 data)
 708 {
 709         struct virtnet_info *vi = netdev_priv(dev);
 710         struct virtio_device *vdev = vi->vdev;
 711
 712         if (data && !virtio_has_feature(vdev, VIRTIO_NET_F_CSUM))
 713                 return -ENOSYS;
 714
 715         return ethtool_op_set_tx_hw_csum(dev, data);
 716 }
 717
 718 static void virtnet_set_rx_mode(struct net_device *dev)
 719 {
 720         struct virtnet_info *vi = netdev_priv(dev);
 721         struct scatterlist sg[2];
 722         u8 promisc, allmulti;
 723         struct virtio_net_ctrl_mac *mac_data;
 724         struct netdev_hw_addr *ha;
 725         int uc_count;
 726         int mc_count;
 727         void *buf;
 728         int i;
 729
 730         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 731         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 732                 return;
 733
 734         promisc = ((dev->flags & IFF_PROMISC) != 0);
 735         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 736
 737         sg_init_one(sg, &promisc, sizeof(promisc));
 738
 739         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 740                                   VIRTIO_NET_CTRL_RX_PROMISC,
 741                                   sg, 1, 0))
 742                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 743                          promisc ? "en" : "dis");
 744
 745         sg_init_one(sg, &allmulti, sizeof(allmulti));
 746
 747         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 748                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 749                                   sg, 1, 0))
 750                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
 751                          allmulti ? "en" : "dis");
 752
 753         uc_count = netdev_uc_count(dev);
 754         mc_count = netdev_mc_count(dev);
 755         /* MAC filter - use one buffer for both lists */
 756         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
 757                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
 758         mac_data = buf;
 759         if (!buf) {
 760                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
 761                 return;
 762         }
 763
 764         sg_init_table(sg, 2);
 765
 766         /* Store the unicast list and count in the front of the buffer */
 767         mac_data->entries = uc_count;
 768         i = 0;
 769         netdev_for_each_uc_addr(ha, dev)
 770                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 771
 772         sg_set_buf(&sg[0], mac_data,
 773                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
 774
 775         /* multicast list and count fill the end */
 776         mac_data = (void *)&mac_data->macs[uc_count][0];
 777
 778         mac_data->entries = mc_count;
 779         i = 0;
 780         netdev_for_each_mc_addr(ha, dev)
 781                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 782
 783         sg_set_buf(&sg[1], mac_data,
 784                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
 785
 786         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 787                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
 788                                   sg, 2, 0))
 789                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
 790
 791         kfree(buf);
 792 }
 793
 794 static void virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
 795 {
 796         struct virtnet_info *vi = netdev_priv(dev);
 797         struct scatterlist sg;
 798
 799         sg_init_one(&sg, &vid, sizeof(vid));
 800
 801         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 802                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
 803                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
 804 }
 805
 806 static void virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
 807 {
 808         struct virtnet_info *vi = netdev_priv(dev);
 809         struct scatterlist sg;
 810
 811         sg_init_one(&sg, &vid, sizeof(vid));
 812
 813         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 814                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
 815                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
 816 }
 817
 818 static const struct ethtool_ops virtnet_ethtool_ops = {
 819         .set_tx_csum = virtnet_set_tx_csum,
 820         .set_sg = ethtool_op_set_sg,
 821         .set_tso = ethtool_op_set_tso,
 822         .set_ufo = ethtool_op_set_ufo,
 823         .get_link = ethtool_op_get_link,
 824 };
 825
 826 #define MIN_MTU 68
 827 #define MAX_MTU 65535
 828
 829 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
 830 {
 831         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
 832                 return -EINVAL;
 833         dev->mtu = new_mtu;
 834         return 0;
 835 }
 836
 837 static const struct net_device_ops virtnet_netdev = {
 838         .ndo_open            = virtnet_open,
 839         .ndo_stop            = virtnet_close,
 840         .ndo_start_xmit      = start_xmit,
 841         .ndo_validate_addr   = eth_validate_addr,
 842         .ndo_set_mac_address = virtnet_set_mac_address,
 843         .ndo_set_rx_mode     = virtnet_set_rx_mode,
 844         .ndo_change_mtu      = virtnet_change_mtu,
 845         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
 846         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
 847 #ifdef CONFIG_NET_POLL_CONTROLLER
 848         .ndo_poll_controller = virtnet_netpoll,
 849 #endif
 850 };
 851
 852 static void virtnet_update_status(struct virtnet_info *vi)
 853 {
 854         u16 v;
 855
 856         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS))
 857                 return;
 858
 859         vi->vdev->config->get(vi->vdev,
 860                               offsetof(struct virtio_net_config, status),
 861                               &v, sizeof(v));
 862
 863         /* Ignore unknown (future) status bits */
 864         v &= VIRTIO_NET_S_LINK_UP;
 865
 866         if (vi->status == v)
 867                 return;
 868
 869         vi->status = v;
 870
 871         if (vi->status & VIRTIO_NET_S_LINK_UP) {
 872                 netif_carrier_on(vi->dev);
 873                 netif_wake_queue(vi->dev);
 874         } else {
 875                 netif_carrier_off(vi->dev);
 876                 netif_stop_queue(vi->dev);
 877         }
 878 }
 879
 880 static void virtnet_config_changed(struct virtio_device *vdev)
 881 {
 882         struct virtnet_info *vi = vdev->priv;
 883
 884         virtnet_update_status(vi);
 885 }
 886
 887 static int virtnet_probe(struct virtio_device *vdev)
 888 {
 889         int err;
 890         struct net_device *dev;
 891         struct virtnet_info *vi;
 892         struct virtqueue *vqs[3];
 893         vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
 894         const char *names[] = { "input", "output", "control" };
 895         int nvqs;
 896
 897         /* Allocate ourselves a network device with room for our info */
 898         dev = alloc_etherdev(sizeof(struct virtnet_info));
 899         if (!dev)
 900                 return -ENOMEM;
 901
 902         /* Set up network device as normal. */
 903         dev->netdev_ops = &virtnet_netdev;
 904         dev->features = NETIF_F_HIGHDMA;
 905         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
 906         SET_NETDEV_DEV(dev, &vdev->dev);
 907
 908         /* Do we support "hardware" checksums? */
 909         if (csum && virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
 910                 /* This opens up the world of extra features. */
 911                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
 912                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
 913                         dev->features |= NETIF_F_TSO | NETIF_F_UFO
 914                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
 915                 }
 916                 /* Individual feature bits: what can host handle? */
 917                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
 918                         dev->features |= NETIF_F_TSO;
 919                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
 920                         dev->features |= NETIF_F_TSO6;
 921                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
 922                         dev->features |= NETIF_F_TSO_ECN;
 923                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
 924                         dev->features |= NETIF_F_UFO;
 925         }
 926
 927         /* Configuration may specify what MAC to use.  Otherwise random. */
 928         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 929                 vdev->config->get(vdev,
 930                                   offsetof(struct virtio_net_config, mac),
 931                                   dev->dev_addr, dev->addr_len);
 932         } else
 933                 random_ether_addr(dev->dev_addr);
 934
 935         /* Set up our device-specific information */
 936         vi = netdev_priv(dev);
 937         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
 938         vi->dev = dev;
 939         vi->vdev = vdev;
 940         vdev->priv = vi;
 941         vi->pages = NULL;
 942         INIT_DELAYED_WORK(&vi->refill, refill_work);
 943
 944         /* If we can receive ANY GSO packets, we must allocate large ones. */
 945         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
 946             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
 947             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
 948                 vi->big_packets = true;
 949
 950         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
 951                 vi->mergeable_rx_bufs = true;
 952
 953         /* We expect two virtqueues, receive then send,
 954          * and optionally control. */
 955         nvqs = virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ? 3 : 2;
 956
 957         err = vdev->config->find_vqs(vdev, nvqs, vqs, callbacks, names);
 958         if (err)
 959                 goto free;
 960
 961         vi->rvq = vqs[0];
 962         vi->svq = vqs[1];
 963
 964         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)) {
 965                 vi->cvq = vqs[2];
 966
 967                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
 968                         dev->features |= NETIF_F_HW_VLAN_FILTER;
 969         }
 970
 971         err = register_netdev(dev);
 972         if (err) {
 973                 pr_debug("virtio_net: registering device failed\n");
 974                 goto free_vqs;
 975         }
 976
 977         /* Last of all, set up some receive buffers. */
 978         try_fill_recv(vi, GFP_KERNEL);
 979
 980         /* If we didn't even get one input buffer, we're useless. */
 981         if (vi->num == 0) {
 982                 err = -ENOMEM;
 983                 goto unregister;
 984         }
 985
 986         vi->status = VIRTIO_NET_S_LINK_UP;
 987         virtnet_update_status(vi);
 988         netif_carrier_on(dev);
 989
 990         pr_debug("virtnet: registered device %s\n", dev->name);
 991         return 0;
 992
 993 unregister:
 994         unregister_netdev(dev);
 995         cancel_delayed_work_sync(&vi->refill);
 996 free_vqs:
 997         vdev->config->del_vqs(vdev);
 998 free:
 999         free_netdev(dev);
1000         return err;
1001 }
1002
1003 static void free_unused_bufs(struct virtnet_info *vi)
1004 {
1005         void *buf;
1006         while (1) {
1007                 buf = vi->svq->vq_ops->detach_unused_buf(vi->svq);
1008                 if (!buf)
1009                         break;
1010                 dev_kfree_skb(buf);
1011         }
1012         while (1) {
1013                 buf = vi->rvq->vq_ops->detach_unused_buf(vi->rvq);
1014                 if (!buf)
1015                         break;
1016                 if (vi->mergeable_rx_bufs || vi->big_packets)
1017                         give_pages(vi, buf);
1018                 else
1019                         dev_kfree_skb(buf);
1020                 --vi->num;
1021         }
1022         BUG_ON(vi->num != 0);
1023 }
1024
1025 static void __devexit virtnet_remove(struct virtio_device *vdev)
1026 {
1027         struct virtnet_info *vi = vdev->priv;
1028
1029         /* Stop all the virtqueues. */
1030         vdev->config->reset(vdev);
1031
1032
1033         unregister_netdev(vi->dev);
1034         cancel_delayed_work_sync(&vi->refill);
1035
1036         /* Free unused buffers in both send and recv, if any. */
1037         free_unused_bufs(vi);
1038
1039         vdev->config->del_vqs(vi->vdev);
1040
1041         while (vi->pages)
1042                 __free_pages(get_a_page(vi, GFP_KERNEL), 0);
1043
1044         free_netdev(vi->dev);
1045 }
1046
1047 static struct virtio_device_id id_table[] = {
1048         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1049         { 0 },
1050 };
1051
1052 static unsigned int features[] = {
1053         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1054         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1055         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1056         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1057         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1058         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1059         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1060 };
1061
1062 static struct virtio_driver virtio_net_driver = {
1063         .feature_table = features,
1064         .feature_table_size = ARRAY_SIZE(features),
1065         .driver.name =  KBUILD_MODNAME,
1066         .driver.owner = THIS_MODULE,
1067         .id_table =     id_table,
1068         .probe =        virtnet_probe,
1069         .remove =       __devexit_p(virtnet_remove),
1070         .config_changed = virtnet_config_changed,
1071 };
1072
1073 static int __init init(void)
1074 {
1075         return register_virtio_driver(&virtio_net_driver);
1076 }
1077
1078 static void __exit fini(void)
1079 {
1080         unregister_virtio_driver(&virtio_net_driver);
1081 }
1082 module_init(init);
1083 module_exit(fini);
1084
1085 MODULE_DEVICE_TABLE(virtio, id_table);
1086 MODULE_DESCRIPTION("Virtio network driver");
1087 MODULE_LICENSE("GPL");