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_info {
  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         /* fragments + linear part + virtio header */
  66         struct scatterlist rx_sg[MAX_SKB_FRAGS + 2];
  67         struct scatterlist tx_sg[MAX_SKB_FRAGS + 2];
  68 };
  69
  70 struct skb_vnet_hdr {
  71         union {
  72                 struct virtio_net_hdr hdr;
  73                 struct virtio_net_hdr_mrg_rxbuf mhdr;
  74         };
  75         unsigned int num_sg;
  76 };
  77
  78 struct padded_vnet_hdr {
  79         struct virtio_net_hdr hdr;
  80         /*
  81          * virtio_net_hdr should be in a separated sg buffer because of a
  82          * QEMU bug, and data sg buffer shares same page with this header sg.
  83          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
  84          */
  85         char padding[6];
  86 };
  87
  88 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
  89 {
  90         return (struct skb_vnet_hdr *)skb->cb;
  91 }
  92
  93 /*
  94  * private is used to chain pages for big packets, put the whole
  95  * most recent used list in the beginning for reuse
  96  */
  97 static void give_pages(struct virtnet_info *vi, struct page *page)
  98 {
  99         struct page *end;
 100
 101         /* Find end of list, sew whole thing into vi->pages. */
 102         for (end = page; end->private; end = (struct page *)end->private);
 103         end->private = (unsigned long)vi->pages;
 104         vi->pages = page;
 105 }
 106
 107 static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
 108 {
 109         struct page *p = vi->pages;
 110
 111         if (p) {
 112                 vi->pages = (struct page *)p->private;
 113                 /* clear private here, it is used to chain pages */
 114                 p->private = 0;
 115         } else
 116                 p = alloc_page(gfp_mask);
 117         return p;
 118 }
 119
 120 static void skb_xmit_done(struct virtqueue *svq)
 121 {
 122         struct virtnet_info *vi = svq->vdev->priv;
 123
 124         /* Suppress further interrupts. */
 125         virtqueue_disable_cb(svq);
 126
 127         /* We were probably waiting for more output buffers. */
 128         netif_wake_queue(vi->dev);
 129 }
 130
 131 static void set_skb_frag(struct sk_buff *skb, struct page *page,
 132                          unsigned int offset, unsigned int *len)
 133 {
 134         int i = skb_shinfo(skb)->nr_frags;
 135         skb_frag_t *f;
 136
 137         f = &skb_shinfo(skb)->frags[i];
 138         f->size = min((unsigned)PAGE_SIZE - offset, *len);
 139         f->page_offset = offset;
 140         f->page = page;
 141
 142         skb->data_len += f->size;
 143         skb->len += f->size;
 144         skb_shinfo(skb)->nr_frags++;
 145         *len -= f->size;
 146 }
 147
 148 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 149                                    struct page *page, unsigned int len)
 150 {
 151         struct sk_buff *skb;
 152         struct skb_vnet_hdr *hdr;
 153         unsigned int copy, hdr_len, offset;
 154         char *p;
 155
 156         p = page_address(page);
 157
 158         /* copy small packet so we can reuse these pages for small data */
 159         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 160         if (unlikely(!skb))
 161                 return NULL;
 162
 163         hdr = skb_vnet_hdr(skb);
 164
 165         if (vi->mergeable_rx_bufs) {
 166                 hdr_len = sizeof hdr->mhdr;
 167                 offset = hdr_len;
 168         } else {
 169                 hdr_len = sizeof hdr->hdr;
 170                 offset = sizeof(struct padded_vnet_hdr);
 171         }
 172
 173         memcpy(hdr, p, hdr_len);
 174
 175         len -= hdr_len;
 176         p += offset;
 177
 178         copy = len;
 179         if (copy > skb_tailroom(skb))
 180                 copy = skb_tailroom(skb);
 181         memcpy(skb_put(skb, copy), p, copy);
 182
 183         len -= copy;
 184         offset += copy;
 185
 186         while (len) {
 187                 set_skb_frag(skb, page, offset, &len);
 188                 page = (struct page *)page->private;
 189                 offset = 0;
 190         }
 191
 192         if (page)
 193                 give_pages(vi, page);
 194
 195         return skb;
 196 }
 197
 198 static int receive_mergeable(struct virtnet_info *vi, struct sk_buff *skb)
 199 {
 200         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 201         struct page *page;
 202         int num_buf, i, len;
 203
 204         num_buf = hdr->mhdr.num_buffers;
 205         while (--num_buf) {
 206                 i = skb_shinfo(skb)->nr_frags;
 207                 if (i >= MAX_SKB_FRAGS) {
 208                         pr_debug("%s: packet too long\n", skb->dev->name);
 209                         skb->dev->stats.rx_length_errors++;
 210                         return -EINVAL;
 211                 }
 212
 213                 page = virtqueue_get_buf(vi->rvq, &len);
 214                 if (!page) {
 215                         pr_debug("%s: rx error: %d buffers missing\n",
 216                                  skb->dev->name, hdr->mhdr.num_buffers);
 217                         skb->dev->stats.rx_length_errors++;
 218                         return -EINVAL;
 219                 }
 220                 if (len > PAGE_SIZE)
 221                         len = PAGE_SIZE;
 222
 223                 set_skb_frag(skb, page, 0, &len);
 224
 225                 --vi->num;
 226         }
 227         return 0;
 228 }
 229
 230 static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
 231 {
 232         struct virtnet_info *vi = netdev_priv(dev);
 233         struct sk_buff *skb;
 234         struct page *page;
 235         struct skb_vnet_hdr *hdr;
 236
 237         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 238                 pr_debug("%s: short packet %i\n", dev->name, len);
 239                 dev->stats.rx_length_errors++;
 240                 if (vi->mergeable_rx_bufs || vi->big_packets)
 241                         give_pages(vi, buf);
 242                 else
 243                         dev_kfree_skb(buf);
 244                 return;
 245         }
 246
 247         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
 248                 skb = buf;
 249                 len -= sizeof(struct virtio_net_hdr);
 250                 skb_trim(skb, len);
 251         } else {
 252                 page = buf;
 253                 skb = page_to_skb(vi, page, len);
 254                 if (unlikely(!skb)) {
 255                         dev->stats.rx_dropped++;
 256                         give_pages(vi, page);
 257                         return;
 258                 }
 259                 if (vi->mergeable_rx_bufs)
 260                         if (receive_mergeable(vi, skb)) {
 261                                 dev_kfree_skb(skb);
 262                                 return;
 263                         }
 264         }
 265
 266         hdr = skb_vnet_hdr(skb);
 267         skb->truesize += skb->data_len;
 268         dev->stats.rx_bytes += skb->len;
 269         dev->stats.rx_packets++;
 270
 271         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 272                 pr_debug("Needs csum!\n");
 273                 if (!skb_partial_csum_set(skb,
 274                                           hdr->hdr.csum_start,
 275                                           hdr->hdr.csum_offset))
 276                         goto frame_err;
 277         }
 278
 279         skb->protocol = eth_type_trans(skb, dev);
 280         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 281                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 282
 283         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 284                 pr_debug("GSO!\n");
 285                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 286                 case VIRTIO_NET_HDR_GSO_TCPV4:
 287                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 288                         break;
 289                 case VIRTIO_NET_HDR_GSO_UDP:
 290                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 291                         break;
 292                 case VIRTIO_NET_HDR_GSO_TCPV6:
 293                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 294                         break;
 295                 default:
 296                         if (net_ratelimit())
 297                                 printk(KERN_WARNING "%s: bad gso type %u.\n",
 298                                        dev->name, hdr->hdr.gso_type);
 299                         goto frame_err;
 300                 }
 301
 302                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 303                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 304
 305                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 306                 if (skb_shinfo(skb)->gso_size == 0) {
 307                         if (net_ratelimit())
 308                                 printk(KERN_WARNING "%s: zero gso size.\n",
 309                                        dev->name);
 310                         goto frame_err;
 311                 }
 312
 313                 /* Header must be checked, and gso_segs computed. */
 314                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 315                 skb_shinfo(skb)->gso_segs = 0;
 316         }
 317
 318         netif_receive_skb(skb);
 319         return;
 320
 321 frame_err:
 322         dev->stats.rx_frame_errors++;
 323         dev_kfree_skb(skb);
 324 }
 325
 326 static int add_recvbuf_small(struct virtnet_info *vi, gfp_t gfp)
 327 {
 328         struct sk_buff *skb;
 329         struct skb_vnet_hdr *hdr;
 330         int err;
 331
 332         skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
 333         if (unlikely(!skb))
 334                 return -ENOMEM;
 335
 336         skb_put(skb, MAX_PACKET_LEN);
 337
 338         hdr = skb_vnet_hdr(skb);
 339         sg_set_buf(vi->rx_sg, &hdr->hdr, sizeof hdr->hdr);
 340
 341         skb_to_sgvec(skb, vi->rx_sg + 1, 0, skb->len);
 342
 343         err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, 2, skb, gfp);
 344         if (err < 0)
 345                 dev_kfree_skb(skb);
 346
 347         return err;
 348 }
 349
 350 static int add_recvbuf_big(struct virtnet_info *vi, gfp_t gfp)
 351 {
 352         struct page *first, *list = NULL;
 353         char *p;
 354         int i, err, offset;
 355
 356         /* page in vi->rx_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(&vi->rx_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         /* vi->rx_sg[0], vi->rx_sg[1] share the same page */
 379         /* a separated vi->rx_sg[0] for virtio_net_hdr only due to QEMU bug */
 380         sg_set_buf(&vi->rx_sg[0], p, sizeof(struct virtio_net_hdr));
 381
 382         /* vi->rx_sg[1] for data packet, from offset */
 383         offset = sizeof(struct padded_vnet_hdr);
 384         sg_set_buf(&vi->rx_sg[1], p + offset, PAGE_SIZE - offset);
 385
 386         /* chain first in list head */
 387         first->private = (unsigned long)list;
 388         err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, MAX_SKB_FRAGS + 2,
 389                                     first, gfp);
 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         int err;
 400
 401         page = get_a_page(vi, gfp);
 402         if (!page)
 403                 return -ENOMEM;
 404
 405         sg_init_one(vi->rx_sg, page_address(page), PAGE_SIZE);
 406
 407         err = virtqueue_add_buf_gfp(vi->rvq, vi->rx_sg, 0, 1, page, gfp);
 408         if (err < 0)
 409                 give_pages(vi, page);
 410
 411         return err;
 412 }
 413
 414 /* Returns false if we couldn't fill entirely (OOM). */
 415 static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
 416 {
 417         int err;
 418         bool oom;
 419
 420         do {
 421                 if (vi->mergeable_rx_bufs)
 422                         err = add_recvbuf_mergeable(vi, gfp);
 423                 else if (vi->big_packets)
 424                         err = add_recvbuf_big(vi, gfp);
 425                 else
 426                         err = add_recvbuf_small(vi, gfp);
 427
 428                 oom = err == -ENOMEM;
 429                 if (err < 0)
 430                         break;
 431                 ++vi->num;
 432         } while (err > 0);
 433         if (unlikely(vi->num > vi->max))
 434                 vi->max = vi->num;
 435         virtqueue_kick(vi->rvq);
 436         return !oom;
 437 }
 438
 439 static void skb_recv_done(struct virtqueue *rvq)
 440 {
 441         struct virtnet_info *vi = rvq->vdev->priv;
 442         /* Schedule NAPI, Suppress further interrupts if successful. */
 443         if (napi_schedule_prep(&vi->napi)) {
 444                 virtqueue_disable_cb(rvq);
 445                 __napi_schedule(&vi->napi);
 446         }
 447 }
 448
 449 static void refill_work(struct work_struct *work)
 450 {
 451         struct virtnet_info *vi;
 452         bool still_empty;
 453
 454         vi = container_of(work, struct virtnet_info, refill.work);
 455         napi_disable(&vi->napi);
 456         still_empty = !try_fill_recv(vi, GFP_KERNEL);
 457         napi_enable(&vi->napi);
 458
 459         /* In theory, this can happen: if we don't get any buffers in
 460          * we will *never* try to fill again. */
 461         if (still_empty)
 462                 schedule_delayed_work(&vi->refill, HZ/2);
 463 }
 464
 465 static int virtnet_poll(struct napi_struct *napi, int budget)
 466 {
 467         struct virtnet_info *vi = container_of(napi, struct virtnet_info, napi);
 468         void *buf;
 469         unsigned int len, received = 0;
 470
 471 again:
 472         while (received < budget &&
 473                (buf = virtqueue_get_buf(vi->rvq, &len)) != NULL) {
 474                 receive_buf(vi->dev, buf, len);
 475                 --vi->num;
 476                 received++;
 477         }
 478
 479         if (vi->num < vi->max / 2) {
 480                 if (!try_fill_recv(vi, GFP_ATOMIC))
 481                         schedule_delayed_work(&vi->refill, 0);
 482         }
 483
 484         /* Out of packets? */
 485         if (received < budget) {
 486                 napi_complete(napi);
 487                 if (unlikely(!virtqueue_enable_cb(vi->rvq)) &&
 488                     napi_schedule_prep(napi)) {
 489                         virtqueue_disable_cb(vi->rvq);
 490                         __napi_schedule(napi);
 491                         goto again;
 492                 }
 493         }
 494
 495         return received;
 496 }
 497
 498 static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
 499 {
 500         struct sk_buff *skb;
 501         unsigned int len, tot_sgs = 0;
 502
 503         while ((skb = virtqueue_get_buf(vi->svq, &len)) != NULL) {
 504                 pr_debug("Sent skb %p\n", skb);
 505                 vi->dev->stats.tx_bytes += skb->len;
 506                 vi->dev->stats.tx_packets++;
 507                 tot_sgs += skb_vnet_hdr(skb)->num_sg;
 508                 dev_kfree_skb_any(skb);
 509         }
 510         return tot_sgs;
 511 }
 512
 513 static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
 514 {
 515         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 516         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 517
 518         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 519
 520         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 521                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 522                 hdr->hdr.csum_start = skb->csum_start - skb_headroom(skb);
 523                 hdr->hdr.csum_offset = skb->csum_offset;
 524         } else {
 525                 hdr->hdr.flags = 0;
 526                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 527         }
 528
 529         if (skb_is_gso(skb)) {
 530                 hdr->hdr.hdr_len = skb_headlen(skb);
 531                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 532                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 533                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 534                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 535                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 536                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 537                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 538                 else
 539                         BUG();
 540                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 541                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 542         } else {
 543                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 544                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 545         }
 546
 547         hdr->mhdr.num_buffers = 0;
 548
 549         /* Encode metadata header at front. */
 550         if (vi->mergeable_rx_bufs)
 551                 sg_set_buf(vi->tx_sg, &hdr->mhdr, sizeof hdr->mhdr);
 552         else
 553                 sg_set_buf(vi->tx_sg, &hdr->hdr, sizeof hdr->hdr);
 554
 555         hdr->num_sg = skb_to_sgvec(skb, vi->tx_sg + 1, 0, skb->len) + 1;
 556         return virtqueue_add_buf(vi->svq, vi->tx_sg, hdr->num_sg,
 557                                         0, skb);
 558 }
 559
 560 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 561 {
 562         struct virtnet_info *vi = netdev_priv(dev);
 563         int capacity;
 564
 565         /* Free up any pending old buffers before queueing new ones. */
 566         free_old_xmit_skbs(vi);
 567
 568         /* Try to transmit */
 569         capacity = xmit_skb(vi, skb);
 570
 571         /* This can happen with OOM and indirect buffers. */
 572         if (unlikely(capacity < 0)) {
 573                 if (net_ratelimit()) {
 574                         if (likely(capacity == -ENOMEM)) {
 575                                 dev_warn(&dev->dev,
 576                                          "TX queue failure: out of memory\n");
 577                         } else {
 578                                 dev->stats.tx_fifo_errors++;
 579                                 dev_warn(&dev->dev,
 580                                          "Unexpected TX queue failure: %d\n",
 581                                          capacity);
 582                         }
 583                 }
 584                 dev->stats.tx_dropped++;
 585                 kfree_skb(skb);
 586                 return NETDEV_TX_OK;
 587         }
 588         virtqueue_kick(vi->svq);
 589
 590         /* Don't wait up for transmitted skbs to be freed. */
 591         skb_orphan(skb);
 592         nf_reset(skb);
 593
 594         /* Apparently nice girls don't return TX_BUSY; stop the queue
 595          * before it gets out of hand.  Naturally, this wastes entries. */
 596         if (capacity < 2+MAX_SKB_FRAGS) {
 597                 netif_stop_queue(dev);
 598                 if (unlikely(!virtqueue_enable_cb(vi->svq))) {
 599                         /* More just got used, free them then recheck. */
 600                         capacity += free_old_xmit_skbs(vi);
 601                         if (capacity >= 2+MAX_SKB_FRAGS) {
 602                                 netif_start_queue(dev);
 603                                 virtqueue_disable_cb(vi->svq);
 604                         }
 605                 }
 606         }
 607
 608         return NETDEV_TX_OK;
 609 }
 610
 611 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 612 {
 613         struct virtnet_info *vi = netdev_priv(dev);
 614         struct virtio_device *vdev = vi->vdev;
 615         int ret;
 616
 617         ret = eth_mac_addr(dev, p);
 618         if (ret)
 619                 return ret;
 620
 621         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
 622                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 623                                   dev->dev_addr, dev->addr_len);
 624
 625         return 0;
 626 }
 627
 628 #ifdef CONFIG_NET_POLL_CONTROLLER
 629 static void virtnet_netpoll(struct net_device *dev)
 630 {
 631         struct virtnet_info *vi = netdev_priv(dev);
 632
 633         napi_schedule(&vi->napi);
 634 }
 635 #endif
 636
 637 static int virtnet_open(struct net_device *dev)
 638 {
 639         struct virtnet_info *vi = netdev_priv(dev);
 640
 641         napi_enable(&vi->napi);
 642
 643         /* If all buffers were filled by other side before we napi_enabled, we
 644          * won't get another interrupt, so process any outstanding packets
 645          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 646          * We synchronize against interrupts via NAPI_STATE_SCHED */
 647         if (napi_schedule_prep(&vi->napi)) {
 648                 virtqueue_disable_cb(vi->rvq);
 649                 __napi_schedule(&vi->napi);
 650         }
 651         return 0;
 652 }
 653
 654 /*
 655  * Send command via the control virtqueue and check status.  Commands
 656  * supported by the hypervisor, as indicated by feature bits, should
 657  * never fail unless improperly formated.
 658  */
 659 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 660                                  struct scatterlist *data, int out, int in)
 661 {
 662         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
 663         struct virtio_net_ctrl_hdr ctrl;
 664         virtio_net_ctrl_ack status = ~0;
 665         unsigned int tmp;
 666         int i;
 667
 668         /* Caller should know better */
 669         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
 670                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
 671
 672         out++; /* Add header */
 673         in++; /* Add return status */
 674
 675         ctrl.class = class;
 676         ctrl.cmd = cmd;
 677
 678         sg_init_table(sg, out + in);
 679
 680         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
 681         for_each_sg(data, s, out + in - 2, i)
 682                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
 683         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
 684
 685         BUG_ON(virtqueue_add_buf(vi->cvq, sg, out, in, vi) < 0);
 686
 687         virtqueue_kick(vi->cvq);
 688
 689         /*
 690          * Spin for a response, the kick causes an ioport write, trapping
 691          * into the hypervisor, so the request should be handled immediately.
 692          */
 693         while (!virtqueue_get_buf(vi->cvq, &tmp))
 694                 cpu_relax();
 695
 696         return status == VIRTIO_NET_OK;
 697 }
 698
 699 static int virtnet_close(struct net_device *dev)
 700 {
 701         struct virtnet_info *vi = netdev_priv(dev);
 702
 703         napi_disable(&vi->napi);
 704
 705         return 0;
 706 }
 707
 708 static int virtnet_set_tx_csum(struct net_device *dev, u32 data)
 709 {
 710         struct virtnet_info *vi = netdev_priv(dev);
 711         struct virtio_device *vdev = vi->vdev;
 712
 713         if (data && !virtio_has_feature(vdev, VIRTIO_NET_F_CSUM))
 714                 return -ENOSYS;
 715
 716         return ethtool_op_set_tx_hw_csum(dev, data);
 717 }
 718
 719 static void virtnet_set_rx_mode(struct net_device *dev)
 720 {
 721         struct virtnet_info *vi = netdev_priv(dev);
 722         struct scatterlist sg[2];
 723         u8 promisc, allmulti;
 724         struct virtio_net_ctrl_mac *mac_data;
 725         struct netdev_hw_addr *ha;
 726         int uc_count;
 727         int mc_count;
 728         void *buf;
 729         int i;
 730
 731         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 732         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 733                 return;
 734
 735         promisc = ((dev->flags & IFF_PROMISC) != 0);
 736         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 737
 738         sg_init_one(sg, &promisc, sizeof(promisc));
 739
 740         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 741                                   VIRTIO_NET_CTRL_RX_PROMISC,
 742                                   sg, 1, 0))
 743                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 744                          promisc ? "en" : "dis");
 745
 746         sg_init_one(sg, &allmulti, sizeof(allmulti));
 747
 748         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 749                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 750                                   sg, 1, 0))
 751                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
 752                          allmulti ? "en" : "dis");
 753
 754         uc_count = netdev_uc_count(dev);
 755         mc_count = netdev_mc_count(dev);
 756         /* MAC filter - use one buffer for both lists */
 757         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
 758                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
 759         mac_data = buf;
 760         if (!buf) {
 761                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
 762                 return;
 763         }
 764
 765         sg_init_table(sg, 2);
 766
 767         /* Store the unicast list and count in the front of the buffer */
 768         mac_data->entries = uc_count;
 769         i = 0;
 770         netdev_for_each_uc_addr(ha, dev)
 771                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 772
 773         sg_set_buf(&sg[0], mac_data,
 774                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
 775
 776         /* multicast list and count fill the end */
 777         mac_data = (void *)&mac_data->macs[uc_count][0];
 778
 779         mac_data->entries = mc_count;
 780         i = 0;
 781         netdev_for_each_mc_addr(ha, dev)
 782                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 783
 784         sg_set_buf(&sg[1], mac_data,
 785                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
 786
 787         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 788                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
 789                                   sg, 2, 0))
 790                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
 791
 792         kfree(buf);
 793 }
 794
 795 static void virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
 796 {
 797         struct virtnet_info *vi = netdev_priv(dev);
 798         struct scatterlist sg;
 799
 800         sg_init_one(&sg, &vid, sizeof(vid));
 801
 802         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 803                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
 804                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
 805 }
 806
 807 static void virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
 808 {
 809         struct virtnet_info *vi = netdev_priv(dev);
 810         struct scatterlist sg;
 811
 812         sg_init_one(&sg, &vid, sizeof(vid));
 813
 814         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 815                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
 816                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
 817 }
 818
 819 static const struct ethtool_ops virtnet_ethtool_ops = {
 820         .set_tx_csum = virtnet_set_tx_csum,
 821         .set_sg = ethtool_op_set_sg,
 822         .set_tso = ethtool_op_set_tso,
 823         .set_ufo = ethtool_op_set_ufo,
 824         .get_link = ethtool_op_get_link,
 825 };
 826
 827 #define MIN_MTU 68
 828 #define MAX_MTU 65535
 829
 830 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
 831 {
 832         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
 833                 return -EINVAL;
 834         dev->mtu = new_mtu;
 835         return 0;
 836 }
 837
 838 static const struct net_device_ops virtnet_netdev = {
 839         .ndo_open            = virtnet_open,
 840         .ndo_stop            = virtnet_close,
 841         .ndo_start_xmit      = start_xmit,
 842         .ndo_validate_addr   = eth_validate_addr,
 843         .ndo_set_mac_address = virtnet_set_mac_address,
 844         .ndo_set_rx_mode     = virtnet_set_rx_mode,
 845         .ndo_change_mtu      = virtnet_change_mtu,
 846         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
 847         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
 848 #ifdef CONFIG_NET_POLL_CONTROLLER
 849         .ndo_poll_controller = virtnet_netpoll,
 850 #endif
 851 };
 852
 853 static void virtnet_update_status(struct virtnet_info *vi)
 854 {
 855         u16 v;
 856
 857         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS))
 858                 return;
 859
 860         vi->vdev->config->get(vi->vdev,
 861                               offsetof(struct virtio_net_config, status),
 862                               &v, sizeof(v));
 863
 864         /* Ignore unknown (future) status bits */
 865         v &= VIRTIO_NET_S_LINK_UP;
 866
 867         if (vi->status == v)
 868                 return;
 869
 870         vi->status = v;
 871
 872         if (vi->status & VIRTIO_NET_S_LINK_UP) {
 873                 netif_carrier_on(vi->dev);
 874                 netif_wake_queue(vi->dev);
 875         } else {
 876                 netif_carrier_off(vi->dev);
 877                 netif_stop_queue(vi->dev);
 878         }
 879 }
 880
 881 static void virtnet_config_changed(struct virtio_device *vdev)
 882 {
 883         struct virtnet_info *vi = vdev->priv;
 884
 885         virtnet_update_status(vi);
 886 }
 887
 888 static int virtnet_probe(struct virtio_device *vdev)
 889 {
 890         int err;
 891         struct net_device *dev;
 892         struct virtnet_info *vi;
 893         struct virtqueue *vqs[3];
 894         vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
 895         const char *names[] = { "input", "output", "control" };
 896         int nvqs;
 897
 898         /* Allocate ourselves a network device with room for our info */
 899         dev = alloc_etherdev(sizeof(struct virtnet_info));
 900         if (!dev)
 901                 return -ENOMEM;
 902
 903         /* Set up network device as normal. */
 904         dev->netdev_ops = &virtnet_netdev;
 905         dev->features = NETIF_F_HIGHDMA;
 906         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
 907         SET_NETDEV_DEV(dev, &vdev->dev);
 908
 909         /* Do we support "hardware" checksums? */
 910         if (csum && virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
 911                 /* This opens up the world of extra features. */
 912                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
 913                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
 914                         dev->features |= NETIF_F_TSO | NETIF_F_UFO
 915                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
 916                 }
 917                 /* Individual feature bits: what can host handle? */
 918                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
 919                         dev->features |= NETIF_F_TSO;
 920                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
 921                         dev->features |= NETIF_F_TSO6;
 922                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
 923                         dev->features |= NETIF_F_TSO_ECN;
 924                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
 925                         dev->features |= NETIF_F_UFO;
 926         }
 927
 928         /* Configuration may specify what MAC to use.  Otherwise random. */
 929         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 930                 vdev->config->get(vdev,
 931                                   offsetof(struct virtio_net_config, mac),
 932                                   dev->dev_addr, dev->addr_len);
 933         } else
 934                 random_ether_addr(dev->dev_addr);
 935
 936         /* Set up our device-specific information */
 937         vi = netdev_priv(dev);
 938         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
 939         vi->dev = dev;
 940         vi->vdev = vdev;
 941         vdev->priv = vi;
 942         vi->pages = NULL;
 943         INIT_DELAYED_WORK(&vi->refill, refill_work);
 944         sg_init_table(vi->rx_sg, ARRAY_SIZE(vi->rx_sg));
 945         sg_init_table(vi->tx_sg, ARRAY_SIZE(vi->tx_sg));
 946
 947         /* If we can receive ANY GSO packets, we must allocate large ones. */
 948         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
 949             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
 950             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
 951                 vi->big_packets = true;
 952
 953         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
 954                 vi->mergeable_rx_bufs = true;
 955
 956         /* We expect two virtqueues, receive then send,
 957          * and optionally control. */
 958         nvqs = virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ? 3 : 2;
 959
 960         err = vdev->config->find_vqs(vdev, nvqs, vqs, callbacks, names);
 961         if (err)
 962                 goto free;
 963
 964         vi->rvq = vqs[0];
 965         vi->svq = vqs[1];
 966
 967         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)) {
 968                 vi->cvq = vqs[2];
 969
 970                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
 971                         dev->features |= NETIF_F_HW_VLAN_FILTER;
 972         }
 973
 974         err = register_netdev(dev);
 975         if (err) {
 976                 pr_debug("virtio_net: registering device failed\n");
 977                 goto free_vqs;
 978         }
 979
 980         /* Last of all, set up some receive buffers. */
 981         try_fill_recv(vi, GFP_KERNEL);
 982
 983         /* If we didn't even get one input buffer, we're useless. */
 984         if (vi->num == 0) {
 985                 err = -ENOMEM;
 986                 goto unregister;
 987         }
 988
 989         vi->status = VIRTIO_NET_S_LINK_UP;
 990         virtnet_update_status(vi);
 991         netif_carrier_on(dev);
 992
 993         pr_debug("virtnet: registered device %s\n", dev->name);
 994         return 0;
 995
 996 unregister:
 997         unregister_netdev(dev);
 998         cancel_delayed_work_sync(&vi->refill);
 999 free_vqs:
1000         vdev->config->del_vqs(vdev);
1001 free:
1002         free_netdev(dev);
1003         return err;
1004 }
1005
1006 static void free_unused_bufs(struct virtnet_info *vi)
1007 {
1008         void *buf;
1009         while (1) {
1010                 buf = virtqueue_detach_unused_buf(vi->svq);
1011                 if (!buf)
1012                         break;
1013                 dev_kfree_skb(buf);
1014         }
1015         while (1) {
1016                 buf = virtqueue_detach_unused_buf(vi->rvq);
1017                 if (!buf)
1018                         break;
1019                 if (vi->mergeable_rx_bufs || vi->big_packets)
1020                         give_pages(vi, buf);
1021                 else
1022                         dev_kfree_skb(buf);
1023                 --vi->num;
1024         }
1025         BUG_ON(vi->num != 0);
1026 }
1027
1028 static void __devexit virtnet_remove(struct virtio_device *vdev)
1029 {
1030         struct virtnet_info *vi = vdev->priv;
1031
1032         /* Stop all the virtqueues. */
1033         vdev->config->reset(vdev);
1034
1035
1036         unregister_netdev(vi->dev);
1037         cancel_delayed_work_sync(&vi->refill);
1038
1039         /* Free unused buffers in both send and recv, if any. */
1040         free_unused_bufs(vi);
1041
1042         vdev->config->del_vqs(vi->vdev);
1043
1044         while (vi->pages)
1045                 __free_pages(get_a_page(vi, GFP_KERNEL), 0);
1046
1047         free_netdev(vi->dev);
1048 }
1049
1050 static struct virtio_device_id id_table[] = {
1051         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1052         { 0 },
1053 };
1054
1055 static unsigned int features[] = {
1056         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1057         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1058         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1059         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1060         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1061         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1062         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1063 };
1064
1065 static struct virtio_driver virtio_net_driver = {
1066         .feature_table = features,
1067         .feature_table_size = ARRAY_SIZE(features),
1068         .driver.name =  KBUILD_MODNAME,
1069         .driver.owner = THIS_MODULE,
1070         .id_table =     id_table,
1071         .probe =        virtnet_probe,
1072         .remove =       __devexit_p(virtnet_remove),
1073         .config_changed = virtnet_config_changed,
1074 };
1075
1076 static int __init init(void)
1077 {
1078         return register_virtio_driver(&virtio_net_driver);
1079 }
1080
1081 static void __exit fini(void)
1082 {
1083         unregister_virtio_driver(&virtio_net_driver);
1084 }
1085 module_init(init);
1086 module_exit(fini);
1087
1088 MODULE_DEVICE_TABLE(virtio, id_table);
1089 MODULE_DESCRIPTION("Virtio network driver");
1090 MODULE_LICENSE("GPL");