SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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         svq->vq_ops->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 = vi->rvq->vq_ops->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         sg_init_table(sg, 2);
 333         skb = netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN);
 334         if (unlikely(!skb))
 335                 return -ENOMEM;
 336
 337         skb_put(skb, MAX_PACKET_LEN);
 338
 339         hdr = skb_vnet_hdr(skb);
 340         sg_set_buf(vi->rx_sg, &hdr->hdr, sizeof hdr->hdr);
 341
 342         skb_to_sgvec(skb, vi->rx_sg + 1, 0, skb->len);
 343
 344         err = vi->rvq->vq_ops->add_buf(vi->rvq, vi->rx_sg, 0, 2, skb);
 345         if (err < 0)
 346                 dev_kfree_skb(skb);
 347
 348         return err;
 349 }
 350
 351 static int add_recvbuf_big(struct virtnet_info *vi, gfp_t gfp)
 352 {
 353         struct page *first, *list = NULL;
 354         char *p;
 355         int i, err, offset;
 356
 357         /* page in vi->rx_sg[MAX_SKB_FRAGS + 1] is list tail */
 358         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 359                 first = get_a_page(vi, gfp);
 360                 if (!first) {
 361                         if (list)
 362                                 give_pages(vi, list);
 363                         return -ENOMEM;
 364                 }
 365                 sg_set_buf(&vi->rx_sg[i], page_address(first), PAGE_SIZE);
 366
 367                 /* chain new page in list head to match sg */
 368                 first->private = (unsigned long)list;
 369                 list = first;
 370         }
 371
 372         first = get_a_page(vi, gfp);
 373         if (!first) {
 374                 give_pages(vi, list);
 375                 return -ENOMEM;
 376         }
 377         p = page_address(first);
 378
 379         /* vi->rx_sg[0], vi->rx_sg[1] share the same page */
 380         /* a separated vi->rx_sg[0] for virtio_net_hdr only due to QEMU bug */
 381         sg_set_buf(&vi->rx_sg[0], p, sizeof(struct virtio_net_hdr));
 382
 383         /* vi->rx_sg[1] for data packet, from offset */
 384         offset = sizeof(struct padded_vnet_hdr);
 385         sg_set_buf(&vi->rx_sg[1], p + offset, PAGE_SIZE - offset);
 386
 387         /* chain first in list head */
 388         first->private = (unsigned long)list;
 389         err = vi->rvq->vq_ops->add_buf(vi->rvq, vi->rx_sg, 0, MAX_SKB_FRAGS + 2,
 390                                        first);
 391         if (err < 0)
 392                 give_pages(vi, first);
 393
 394         return err;
 395 }
 396
 397 static int add_recvbuf_mergeable(struct virtnet_info *vi, gfp_t gfp)
 398 {
 399         struct page *page;
 400         int err;
 401
 402         page = get_a_page(vi, gfp);
 403         if (!page)
 404                 return -ENOMEM;
 405
 406         sg_init_one(vi->rx_sg, page_address(page), PAGE_SIZE);
 407
 408         err = vi->rvq->vq_ops->add_buf(vi->rvq, vi->rx_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 skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 518         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 519
 520         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 521
 522         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 523                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 524                 hdr->hdr.csum_start = skb->csum_start - skb_headroom(skb);
 525                 hdr->hdr.csum_offset = skb->csum_offset;
 526         } else {
 527                 hdr->hdr.flags = 0;
 528                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 529         }
 530
 531         if (skb_is_gso(skb)) {
 532                 hdr->hdr.hdr_len = skb_headlen(skb);
 533                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 534                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 535                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 536                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 537                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 538                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 539                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 540                 else
 541                         BUG();
 542                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 543                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 544         } else {
 545                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 546                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 547         }
 548
 549         hdr->mhdr.num_buffers = 0;
 550
 551         /* Encode metadata header at front. */
 552         if (vi->mergeable_rx_bufs)
 553                 sg_set_buf(vi->tx_sg, &hdr->mhdr, sizeof hdr->mhdr);
 554         else
 555                 sg_set_buf(vi->tx_sg, &hdr->hdr, sizeof hdr->hdr);
 556
 557         hdr->num_sg = skb_to_sgvec(skb, vi->tx_sg + 1, 0, skb->len) + 1;
 558         return vi->svq->vq_ops->add_buf(vi->svq, vi->tx_sg, hdr->num_sg,
 559                                         0, skb);
 560 }
 561
 562 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 563 {
 564         struct virtnet_info *vi = netdev_priv(dev);
 565         int capacity;
 566
 567 again:
 568         /* Free up any pending old buffers before queueing new ones. */
 569         free_old_xmit_skbs(vi);
 570
 571         /* Try to transmit */
 572         capacity = xmit_skb(vi, skb);
 573
 574         /* This can happen with OOM and indirect buffers. */
 575         if (unlikely(capacity < 0)) {
 576                 netif_stop_queue(dev);
 577                 dev_warn(&dev->dev, "Unexpected full queue\n");
 578                 if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
 579                         vi->svq->vq_ops->disable_cb(vi->svq);
 580                         netif_start_queue(dev);
 581                         goto again;
 582                 }
 583                 return NETDEV_TX_BUSY;
 584         }
 585         vi->svq->vq_ops->kick(vi->svq);
 586
 587         /* Don't wait up for transmitted skbs to be freed. */
 588         skb_orphan(skb);
 589         nf_reset(skb);
 590
 591         /* Apparently nice girls don't return TX_BUSY; stop the queue
 592          * before it gets out of hand.  Naturally, this wastes entries. */
 593         if (capacity < 2+MAX_SKB_FRAGS) {
 594                 netif_stop_queue(dev);
 595                 if (unlikely(!vi->svq->vq_ops->enable_cb(vi->svq))) {
 596                         /* More just got used, free them then recheck. */
 597                         capacity += free_old_xmit_skbs(vi);
 598                         if (capacity >= 2+MAX_SKB_FRAGS) {
 599                                 netif_start_queue(dev);
 600                                 vi->svq->vq_ops->disable_cb(vi->svq);
 601                         }
 602                 }
 603         }
 604
 605         return NETDEV_TX_OK;
 606 }
 607
 608 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 609 {
 610         struct virtnet_info *vi = netdev_priv(dev);
 611         struct virtio_device *vdev = vi->vdev;
 612         int ret;
 613
 614         ret = eth_mac_addr(dev, p);
 615         if (ret)
 616                 return ret;
 617
 618         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
 619                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 620                                   dev->dev_addr, dev->addr_len);
 621
 622         return 0;
 623 }
 624
 625 #ifdef CONFIG_NET_POLL_CONTROLLER
 626 static void virtnet_netpoll(struct net_device *dev)
 627 {
 628         struct virtnet_info *vi = netdev_priv(dev);
 629
 630         napi_schedule(&vi->napi);
 631 }
 632 #endif
 633
 634 static int virtnet_open(struct net_device *dev)
 635 {
 636         struct virtnet_info *vi = netdev_priv(dev);
 637
 638         napi_enable(&vi->napi);
 639
 640         /* If all buffers were filled by other side before we napi_enabled, we
 641          * won't get another interrupt, so process any outstanding packets
 642          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 643          * We synchronize against interrupts via NAPI_STATE_SCHED */
 644         if (napi_schedule_prep(&vi->napi)) {
 645                 vi->rvq->vq_ops->disable_cb(vi->rvq);
 646                 __napi_schedule(&vi->napi);
 647         }
 648         return 0;
 649 }
 650
 651 /*
 652  * Send command via the control virtqueue and check status.  Commands
 653  * supported by the hypervisor, as indicated by feature bits, should
 654  * never fail unless improperly formated.
 655  */
 656 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 657                                  struct scatterlist *data, int out, int in)
 658 {
 659         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
 660         struct virtio_net_ctrl_hdr ctrl;
 661         virtio_net_ctrl_ack status = ~0;
 662         unsigned int tmp;
 663         int i;
 664
 665         /* Caller should know better */
 666         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
 667                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
 668
 669         out++; /* Add header */
 670         in++; /* Add return status */
 671
 672         ctrl.class = class;
 673         ctrl.cmd = cmd;
 674
 675         sg_init_table(sg, out + in);
 676
 677         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
 678         for_each_sg(data, s, out + in - 2, i)
 679                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
 680         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
 681
 682         BUG_ON(vi->cvq->vq_ops->add_buf(vi->cvq, sg, out, in, vi) < 0);
 683
 684         vi->cvq->vq_ops->kick(vi->cvq);
 685
 686         /*
 687          * Spin for a response, the kick causes an ioport write, trapping
 688          * into the hypervisor, so the request should be handled immediately.
 689          */
 690         while (!vi->cvq->vq_ops->get_buf(vi->cvq, &tmp))
 691                 cpu_relax();
 692
 693         return status == VIRTIO_NET_OK;
 694 }
 695
 696 static int virtnet_close(struct net_device *dev)
 697 {
 698         struct virtnet_info *vi = netdev_priv(dev);
 699
 700         napi_disable(&vi->napi);
 701
 702         return 0;
 703 }
 704
 705 static int virtnet_set_tx_csum(struct net_device *dev, u32 data)
 706 {
 707         struct virtnet_info *vi = netdev_priv(dev);
 708         struct virtio_device *vdev = vi->vdev;
 709
 710         if (data && !virtio_has_feature(vdev, VIRTIO_NET_F_CSUM))
 711                 return -ENOSYS;
 712
 713         return ethtool_op_set_tx_hw_csum(dev, data);
 714 }
 715
 716 static void virtnet_set_rx_mode(struct net_device *dev)
 717 {
 718         struct virtnet_info *vi = netdev_priv(dev);
 719         struct scatterlist sg[2];
 720         u8 promisc, allmulti;
 721         struct virtio_net_ctrl_mac *mac_data;
 722         struct netdev_hw_addr *ha;
 723         int uc_count;
 724         int mc_count;
 725         void *buf;
 726         int i;
 727
 728         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 729         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 730                 return;
 731
 732         promisc = ((dev->flags & IFF_PROMISC) != 0);
 733         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 734
 735         sg_init_one(sg, &promisc, sizeof(promisc));
 736
 737         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 738                                   VIRTIO_NET_CTRL_RX_PROMISC,
 739                                   sg, 1, 0))
 740                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 741                          promisc ? "en" : "dis");
 742
 743         sg_init_one(sg, &allmulti, sizeof(allmulti));
 744
 745         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 746                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 747                                   sg, 1, 0))
 748                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
 749                          allmulti ? "en" : "dis");
 750
 751         uc_count = netdev_uc_count(dev);
 752         mc_count = netdev_mc_count(dev);
 753         /* MAC filter - use one buffer for both lists */
 754         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
 755                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
 756         mac_data = buf;
 757         if (!buf) {
 758                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
 759                 return;
 760         }
 761
 762         sg_init_table(sg, 2);
 763
 764         /* Store the unicast list and count in the front of the buffer */
 765         mac_data->entries = uc_count;
 766         i = 0;
 767         netdev_for_each_uc_addr(ha, dev)
 768                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 769
 770         sg_set_buf(&sg[0], mac_data,
 771                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
 772
 773         /* multicast list and count fill the end */
 774         mac_data = (void *)&mac_data->macs[uc_count][0];
 775
 776         mac_data->entries = mc_count;
 777         i = 0;
 778         netdev_for_each_mc_addr(ha, dev)
 779                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 780
 781         sg_set_buf(&sg[1], mac_data,
 782                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
 783
 784         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 785                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
 786                                   sg, 2, 0))
 787                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
 788
 789         kfree(buf);
 790 }
 791
 792 static void virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
 793 {
 794         struct virtnet_info *vi = netdev_priv(dev);
 795         struct scatterlist sg;
 796
 797         sg_init_one(&sg, &vid, sizeof(vid));
 798
 799         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 800                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
 801                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
 802 }
 803
 804 static void virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
 805 {
 806         struct virtnet_info *vi = netdev_priv(dev);
 807         struct scatterlist sg;
 808
 809         sg_init_one(&sg, &vid, sizeof(vid));
 810
 811         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
 812                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
 813                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
 814 }
 815
 816 static const struct ethtool_ops virtnet_ethtool_ops = {
 817         .set_tx_csum = virtnet_set_tx_csum,
 818         .set_sg = ethtool_op_set_sg,
 819         .set_tso = ethtool_op_set_tso,
 820         .set_ufo = ethtool_op_set_ufo,
 821         .get_link = ethtool_op_get_link,
 822 };
 823
 824 #define MIN_MTU 68
 825 #define MAX_MTU 65535
 826
 827 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
 828 {
 829         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
 830                 return -EINVAL;
 831         dev->mtu = new_mtu;
 832         return 0;
 833 }
 834
 835 static const struct net_device_ops virtnet_netdev = {
 836         .ndo_open            = virtnet_open,
 837         .ndo_stop            = virtnet_close,
 838         .ndo_start_xmit      = start_xmit,
 839         .ndo_validate_addr   = eth_validate_addr,
 840         .ndo_set_mac_address = virtnet_set_mac_address,
 841         .ndo_set_rx_mode     = virtnet_set_rx_mode,
 842         .ndo_change_mtu      = virtnet_change_mtu,
 843         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
 844         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
 845 #ifdef CONFIG_NET_POLL_CONTROLLER
 846         .ndo_poll_controller = virtnet_netpoll,
 847 #endif
 848 };
 849
 850 static void virtnet_update_status(struct virtnet_info *vi)
 851 {
 852         u16 v;
 853
 854         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS))
 855                 return;
 856
 857         vi->vdev->config->get(vi->vdev,
 858                               offsetof(struct virtio_net_config, status),
 859                               &v, sizeof(v));
 860
 861         /* Ignore unknown (future) status bits */
 862         v &= VIRTIO_NET_S_LINK_UP;
 863
 864         if (vi->status == v)
 865                 return;
 866
 867         vi->status = v;
 868
 869         if (vi->status & VIRTIO_NET_S_LINK_UP) {
 870                 netif_carrier_on(vi->dev);
 871                 netif_wake_queue(vi->dev);
 872         } else {
 873                 netif_carrier_off(vi->dev);
 874                 netif_stop_queue(vi->dev);
 875         }
 876 }
 877
 878 static void virtnet_config_changed(struct virtio_device *vdev)
 879 {
 880         struct virtnet_info *vi = vdev->priv;
 881
 882         virtnet_update_status(vi);
 883 }
 884
 885 static int virtnet_probe(struct virtio_device *vdev)
 886 {
 887         int err;
 888         struct net_device *dev;
 889         struct virtnet_info *vi;
 890         struct virtqueue *vqs[3];
 891         vq_callback_t *callbacks[] = { skb_recv_done, skb_xmit_done, NULL};
 892         const char *names[] = { "input", "output", "control" };
 893         int nvqs;
 894
 895         /* Allocate ourselves a network device with room for our info */
 896         dev = alloc_etherdev(sizeof(struct virtnet_info));
 897         if (!dev)
 898                 return -ENOMEM;
 899
 900         /* Set up network device as normal. */
 901         dev->netdev_ops = &virtnet_netdev;
 902         dev->features = NETIF_F_HIGHDMA;
 903         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
 904         SET_NETDEV_DEV(dev, &vdev->dev);
 905
 906         /* Do we support "hardware" checksums? */
 907         if (csum && virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
 908                 /* This opens up the world of extra features. */
 909                 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
 910                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
 911                         dev->features |= NETIF_F_TSO | NETIF_F_UFO
 912                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
 913                 }
 914                 /* Individual feature bits: what can host handle? */
 915                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
 916                         dev->features |= NETIF_F_TSO;
 917                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
 918                         dev->features |= NETIF_F_TSO6;
 919                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
 920                         dev->features |= NETIF_F_TSO_ECN;
 921                 if (gso && virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
 922                         dev->features |= NETIF_F_UFO;
 923         }
 924
 925         /* Configuration may specify what MAC to use.  Otherwise random. */
 926         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 927                 vdev->config->get(vdev,
 928                                   offsetof(struct virtio_net_config, mac),
 929                                   dev->dev_addr, dev->addr_len);
 930         } else
 931                 random_ether_addr(dev->dev_addr);
 932
 933         /* Set up our device-specific information */
 934         vi = netdev_priv(dev);
 935         netif_napi_add(dev, &vi->napi, virtnet_poll, napi_weight);
 936         vi->dev = dev;
 937         vi->vdev = vdev;
 938         vdev->priv = vi;
 939         vi->pages = NULL;
 940         INIT_DELAYED_WORK(&vi->refill, refill_work);
 941         sg_init_table(vi->rx_sg, ARRAY_SIZE(vi->rx_sg));
 942         sg_init_table(vi->tx_sg, ARRAY_SIZE(vi->tx_sg));
 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");