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 bool csum = true, gso = true;
  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 #define VIRTNET_DRIVER_VERSION "1.0.0"
  43
  44 struct virtnet_stats {
  45         struct u64_stats_sync tx_syncp;
  46         struct u64_stats_sync rx_syncp;
  47         u64 tx_bytes;
  48         u64 tx_packets;
  49
  50         u64 rx_bytes;
  51         u64 rx_packets;
  52 };
  53
  54 /* Internal representation of a send virtqueue */
  55 struct send_queue {
  56         /* Virtqueue associated with this send _queue */
  57         struct virtqueue *vq;
  58
  59         /* TX: fragments + linear part + virtio header */
  60         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  61
  62         /* Name of the send queue: output.$index */
  63         char name[40];
  64 };
  65
  66 /* Internal representation of a receive virtqueue */
  67 struct receive_queue {
  68         /* Virtqueue associated with this receive_queue */
  69         struct virtqueue *vq;
  70
  71         struct napi_struct napi;
  72
  73         /* Number of input buffers, and max we've ever had. */
  74         unsigned int num, max;
  75
  76         /* Chain pages by the private ptr. */
  77         struct page *pages;
  78
  79         /* RX: fragments + linear part + virtio header */
  80         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  81
  82         /* Name of this receive queue: input.$index */
  83         char name[40];
  84 };
  85
  86 struct virtnet_info {
  87         struct virtio_device *vdev;
  88         struct virtqueue *cvq;
  89         struct net_device *dev;
  90         struct send_queue *sq;
  91         struct receive_queue *rq;
  92         unsigned int status;
  93
  94         /* Max # of queue pairs supported by the device */
  95         u16 max_queue_pairs;
  96
  97         /* # of queue pairs currently used by the driver */
  98         u16 curr_queue_pairs;
  99
 100         /* I like... big packets and I cannot lie! */
 101         bool big_packets;
 102
 103         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 104         bool mergeable_rx_bufs;
 105
 106         /* Has control virtqueue */
 107         bool has_cvq;
 108
 109         /* enable config space updates */
 110         bool config_enable;
 111
 112         /* Active statistics */
 113         struct virtnet_stats __percpu *stats;
 114
 115         /* Work struct for refilling if we run low on memory. */
 116         struct delayed_work refill;
 117
 118         /* Work struct for config space updates */
 119         struct work_struct config_work;
 120
 121         /* Lock for config space updates */
 122         struct mutex config_lock;
 123
 124         /* Does the affinity hint is set for virtqueues? */
 125         bool affinity_hint_set;
 126
 127         /* Per-cpu variable to show the mapping from CPU to virtqueue */
 128         int __percpu *vq_index;
 129 };
 130
 131 struct skb_vnet_hdr {
 132         union {
 133                 struct virtio_net_hdr hdr;
 134                 struct virtio_net_hdr_mrg_rxbuf mhdr;
 135         };
 136 };
 137
 138 struct padded_vnet_hdr {
 139         struct virtio_net_hdr hdr;
 140         /*
 141          * virtio_net_hdr should be in a separated sg buffer because of a
 142          * QEMU bug, and data sg buffer shares same page with this header sg.
 143          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
 144          */
 145         char padding[6];
 146 };
 147
 148 /* Converting between virtqueue no. and kernel tx/rx queue no.
 149  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 150  */
 151 static int vq2txq(struct virtqueue *vq)
 152 {
 153         return (virtqueue_get_queue_index(vq) - 1) / 2;
 154 }
 155
 156 static int txq2vq(int txq)
 157 {
 158         return txq * 2 + 1;
 159 }
 160
 161 static int vq2rxq(struct virtqueue *vq)
 162 {
 163         return virtqueue_get_queue_index(vq) / 2;
 164 }
 165
 166 static int rxq2vq(int rxq)
 167 {
 168         return rxq * 2;
 169 }
 170
 171 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 172 {
 173         return (struct skb_vnet_hdr *)skb->cb;
 174 }
 175
 176 /*
 177  * private is used to chain pages for big packets, put the whole
 178  * most recent used list in the beginning for reuse
 179  */
 180 static void give_pages(struct receive_queue *rq, struct page *page)
 181 {
 182         struct page *end;
 183
 184         /* Find end of list, sew whole thing into vi->rq.pages. */
 185         for (end = page; end->private; end = (struct page *)end->private);
 186         end->private = (unsigned long)rq->pages;
 187         rq->pages = page;
 188 }
 189
 190 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 191 {
 192         struct page *p = rq->pages;
 193
 194         if (p) {
 195                 rq->pages = (struct page *)p->private;
 196                 /* clear private here, it is used to chain pages */
 197                 p->private = 0;
 198         } else
 199                 p = alloc_page(gfp_mask);
 200         return p;
 201 }
 202
 203 static void skb_xmit_done(struct virtqueue *vq)
 204 {
 205         struct virtnet_info *vi = vq->vdev->priv;
 206
 207         /* Suppress further interrupts. */
 208         virtqueue_disable_cb(vq);
 209
 210         /* We were probably waiting for more output buffers. */
 211         netif_wake_subqueue(vi->dev, vq2txq(vq));
 212 }
 213
 214 static void set_skb_frag(struct sk_buff *skb, struct page *page,
 215                          unsigned int offset, unsigned int *len)
 216 {
 217         int size = min((unsigned)PAGE_SIZE - offset, *len);
 218         int i = skb_shinfo(skb)->nr_frags;
 219
 220         __skb_fill_page_desc(skb, i, page, offset, size);
 221
 222         skb->data_len += size;
 223         skb->len += size;
 224         skb->truesize += PAGE_SIZE;
 225         skb_shinfo(skb)->nr_frags++;
 226         *len -= size;
 227 }
 228
 229 /* Called from bottom half context */
 230 static struct sk_buff *page_to_skb(struct receive_queue *rq,
 231                                    struct page *page, unsigned int len)
 232 {
 233         struct virtnet_info *vi = rq->vq->vdev->priv;
 234         struct sk_buff *skb;
 235         struct skb_vnet_hdr *hdr;
 236         unsigned int copy, hdr_len, offset;
 237         char *p;
 238
 239         p = page_address(page);
 240
 241         /* copy small packet so we can reuse these pages for small data */
 242         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 243         if (unlikely(!skb))
 244                 return NULL;
 245
 246         hdr = skb_vnet_hdr(skb);
 247
 248         if (vi->mergeable_rx_bufs) {
 249                 hdr_len = sizeof hdr->mhdr;
 250                 offset = hdr_len;
 251         } else {
 252                 hdr_len = sizeof hdr->hdr;
 253                 offset = sizeof(struct padded_vnet_hdr);
 254         }
 255
 256         memcpy(hdr, p, hdr_len);
 257
 258         len -= hdr_len;
 259         p += offset;
 260
 261         copy = len;
 262         if (copy > skb_tailroom(skb))
 263                 copy = skb_tailroom(skb);
 264         memcpy(skb_put(skb, copy), p, copy);
 265
 266         len -= copy;
 267         offset += copy;
 268
 269         /*
 270          * Verify that we can indeed put this data into a skb.
 271          * This is here to handle cases when the device erroneously
 272          * tries to receive more than is possible. This is usually
 273          * the case of a broken device.
 274          */
 275         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 276                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 277                 dev_kfree_skb(skb);
 278                 return NULL;
 279         }
 280
 281         while (len) {
 282                 set_skb_frag(skb, page, offset, &len);
 283                 page = (struct page *)page->private;
 284                 offset = 0;
 285         }
 286
 287         if (page)
 288                 give_pages(rq, page);
 289
 290         return skb;
 291 }
 292
 293 static int receive_mergeable(struct receive_queue *rq, struct sk_buff *skb)
 294 {
 295         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 296         struct page *page;
 297         int num_buf, i, len;
 298
 299         num_buf = hdr->mhdr.num_buffers;
 300         while (--num_buf) {
 301                 i = skb_shinfo(skb)->nr_frags;
 302                 if (i >= MAX_SKB_FRAGS) {
 303                         pr_debug("%s: packet too long\n", skb->dev->name);
 304                         skb->dev->stats.rx_length_errors++;
 305                         return -EINVAL;
 306                 }
 307                 page = virtqueue_get_buf(rq->vq, &len);
 308                 if (!page) {
 309                         pr_debug("%s: rx error: %d buffers missing\n",
 310                                  skb->dev->name, hdr->mhdr.num_buffers);
 311                         skb->dev->stats.rx_length_errors++;
 312                         return -EINVAL;
 313                 }
 314
 315                 if (len > PAGE_SIZE)
 316                         len = PAGE_SIZE;
 317
 318                 set_skb_frag(skb, page, 0, &len);
 319
 320                 --rq->num;
 321         }
 322         return 0;
 323 }
 324
 325 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
 326 {
 327         struct virtnet_info *vi = rq->vq->vdev->priv;
 328         struct net_device *dev = vi->dev;
 329         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 330         struct sk_buff *skb;
 331         struct page *page;
 332         struct skb_vnet_hdr *hdr;
 333
 334         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 335                 pr_debug("%s: short packet %i\n", dev->name, len);
 336                 dev->stats.rx_length_errors++;
 337                 if (vi->mergeable_rx_bufs || vi->big_packets)
 338                         give_pages(rq, buf);
 339                 else
 340                         dev_kfree_skb(buf);
 341                 return;
 342         }
 343
 344         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
 345                 skb = buf;
 346                 len -= sizeof(struct virtio_net_hdr);
 347                 skb_trim(skb, len);
 348         } else {
 349                 page = buf;
 350                 skb = page_to_skb(rq, page, len);
 351                 if (unlikely(!skb)) {
 352                         dev->stats.rx_dropped++;
 353                         give_pages(rq, page);
 354                         return;
 355                 }
 356                 if (vi->mergeable_rx_bufs)
 357                         if (receive_mergeable(rq, skb)) {
 358                                 dev_kfree_skb(skb);
 359                                 return;
 360                         }
 361         }
 362
 363         hdr = skb_vnet_hdr(skb);
 364
 365         u64_stats_update_begin(&stats->rx_syncp);
 366         stats->rx_bytes += skb->len;
 367         stats->rx_packets++;
 368         u64_stats_update_end(&stats->rx_syncp);
 369
 370         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 371                 pr_debug("Needs csum!\n");
 372                 if (!skb_partial_csum_set(skb,
 373                                           hdr->hdr.csum_start,
 374                                           hdr->hdr.csum_offset))
 375                         goto frame_err;
 376         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 377                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 378         }
 379
 380         skb->protocol = eth_type_trans(skb, dev);
 381         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 382                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 383
 384         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 385                 pr_debug("GSO!\n");
 386                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 387                 case VIRTIO_NET_HDR_GSO_TCPV4:
 388                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 389                         break;
 390                 case VIRTIO_NET_HDR_GSO_UDP:
 391                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 392                         break;
 393                 case VIRTIO_NET_HDR_GSO_TCPV6:
 394                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 395                         break;
 396                 default:
 397                         net_warn_ratelimited("%s: bad gso type %u.\n",
 398                                              dev->name, hdr->hdr.gso_type);
 399                         goto frame_err;
 400                 }
 401
 402                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 403                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 404
 405                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 406                 if (skb_shinfo(skb)->gso_size == 0) {
 407                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 408                         goto frame_err;
 409                 }
 410
 411                 /* Header must be checked, and gso_segs computed. */
 412                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 413                 skb_shinfo(skb)->gso_segs = 0;
 414         }
 415
 416         netif_receive_skb(skb);
 417         return;
 418
 419 frame_err:
 420         dev->stats.rx_frame_errors++;
 421         dev_kfree_skb(skb);
 422 }
 423
 424 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
 425 {
 426         struct virtnet_info *vi = rq->vq->vdev->priv;
 427         struct sk_buff *skb;
 428         struct skb_vnet_hdr *hdr;
 429         int err;
 430
 431         skb = __netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN, gfp);
 432         if (unlikely(!skb))
 433                 return -ENOMEM;
 434
 435         skb_put(skb, MAX_PACKET_LEN);
 436
 437         hdr = skb_vnet_hdr(skb);
 438         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
 439
 440         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 441
 442         err = virtqueue_add_buf(rq->vq, rq->sg, 0, 2, skb, gfp);
 443         if (err < 0)
 444                 dev_kfree_skb(skb);
 445
 446         return err;
 447 }
 448
 449 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
 450 {
 451         struct page *first, *list = NULL;
 452         char *p;
 453         int i, err, offset;
 454
 455         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 456         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 457                 first = get_a_page(rq, gfp);
 458                 if (!first) {
 459                         if (list)
 460                                 give_pages(rq, list);
 461                         return -ENOMEM;
 462                 }
 463                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 464
 465                 /* chain new page in list head to match sg */
 466                 first->private = (unsigned long)list;
 467                 list = first;
 468         }
 469
 470         first = get_a_page(rq, gfp);
 471         if (!first) {
 472                 give_pages(rq, list);
 473                 return -ENOMEM;
 474         }
 475         p = page_address(first);
 476
 477         /* rq->sg[0], rq->sg[1] share the same page */
 478         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
 479         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
 480
 481         /* rq->sg[1] for data packet, from offset */
 482         offset = sizeof(struct padded_vnet_hdr);
 483         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 484
 485         /* chain first in list head */
 486         first->private = (unsigned long)list;
 487         err = virtqueue_add_buf(rq->vq, rq->sg, 0, MAX_SKB_FRAGS + 2,
 488                                 first, gfp);
 489         if (err < 0)
 490                 give_pages(rq, first);
 491
 492         return err;
 493 }
 494
 495 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 496 {
 497         struct page *page;
 498         int err;
 499
 500         page = get_a_page(rq, gfp);
 501         if (!page)
 502                 return -ENOMEM;
 503
 504         sg_init_one(rq->sg, page_address(page), PAGE_SIZE);
 505
 506         err = virtqueue_add_buf(rq->vq, rq->sg, 0, 1, page, gfp);
 507         if (err < 0)
 508                 give_pages(rq, page);
 509
 510         return err;
 511 }
 512
 513 /*
 514  * Returns false if we couldn't fill entirely (OOM).
 515  *
 516  * Normally run in the receive path, but can also be run from ndo_open
 517  * before we're receiving packets, or from refill_work which is
 518  * careful to disable receiving (using napi_disable).
 519  */
 520 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
 521 {
 522         struct virtnet_info *vi = rq->vq->vdev->priv;
 523         int err;
 524         bool oom;
 525
 526         do {
 527                 if (vi->mergeable_rx_bufs)
 528                         err = add_recvbuf_mergeable(rq, gfp);
 529                 else if (vi->big_packets)
 530                         err = add_recvbuf_big(rq, gfp);
 531                 else
 532                         err = add_recvbuf_small(rq, gfp);
 533
 534                 oom = err == -ENOMEM;
 535                 if (err)
 536                         break;
 537                 ++rq->num;
 538         } while (rq->vq->num_free);
 539         if (unlikely(rq->num > rq->max))
 540                 rq->max = rq->num;
 541         virtqueue_kick(rq->vq);
 542         return !oom;
 543 }
 544
 545 static void skb_recv_done(struct virtqueue *rvq)
 546 {
 547         struct virtnet_info *vi = rvq->vdev->priv;
 548         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 549
 550         /* Schedule NAPI, Suppress further interrupts if successful. */
 551         if (napi_schedule_prep(&rq->napi)) {
 552                 virtqueue_disable_cb(rvq);
 553                 __napi_schedule(&rq->napi);
 554         }
 555 }
 556
 557 static void virtnet_napi_enable(struct receive_queue *rq)
 558 {
 559         napi_enable(&rq->napi);
 560
 561         /* If all buffers were filled by other side before we napi_enabled, we
 562          * won't get another interrupt, so process any outstanding packets
 563          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 564          * We synchronize against interrupts via NAPI_STATE_SCHED */
 565         if (napi_schedule_prep(&rq->napi)) {
 566                 virtqueue_disable_cb(rq->vq);
 567                 local_bh_disable();
 568                 __napi_schedule(&rq->napi);
 569                 local_bh_enable();
 570         }
 571 }
 572
 573 static void refill_work(struct work_struct *work)
 574 {
 575         struct virtnet_info *vi =
 576                 container_of(work, struct virtnet_info, refill.work);
 577         bool still_empty;
 578         int i;
 579
 580         for (i = 0; i < vi->max_queue_pairs; i++) {
 581                 struct receive_queue *rq = &vi->rq[i];
 582
 583                 napi_disable(&rq->napi);
 584                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 585                 virtnet_napi_enable(rq);
 586
 587                 /* In theory, this can happen: if we don't get any buffers in
 588                  * we will *never* try to fill again.
 589                  */
 590                 if (still_empty)
 591                         schedule_delayed_work(&vi->refill, HZ/2);
 592         }
 593 }
 594
 595 static int virtnet_poll(struct napi_struct *napi, int budget)
 596 {
 597         struct receive_queue *rq =
 598                 container_of(napi, struct receive_queue, napi);
 599         struct virtnet_info *vi = rq->vq->vdev->priv;
 600         void *buf;
 601         unsigned int len, received = 0;
 602
 603 again:
 604         while (received < budget &&
 605                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 606                 receive_buf(rq, buf, len);
 607                 --rq->num;
 608                 received++;
 609         }
 610
 611         if (rq->num < rq->max / 2) {
 612                 if (!try_fill_recv(rq, GFP_ATOMIC))
 613                         schedule_delayed_work(&vi->refill, 0);
 614         }
 615
 616         /* Out of packets? */
 617         if (received < budget) {
 618                 napi_complete(napi);
 619                 if (unlikely(!virtqueue_enable_cb(rq->vq)) &&
 620                     napi_schedule_prep(napi)) {
 621                         virtqueue_disable_cb(rq->vq);
 622                         __napi_schedule(napi);
 623                         goto again;
 624                 }
 625         }
 626
 627         return received;
 628 }
 629
 630 static int virtnet_open(struct net_device *dev)
 631 {
 632         struct virtnet_info *vi = netdev_priv(dev);
 633         int i;
 634
 635         for (i = 0; i < vi->max_queue_pairs; i++) {
 636                 /* Make sure we have some buffers: if oom use wq. */
 637                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 638                         schedule_delayed_work(&vi->refill, 0);
 639                 virtnet_napi_enable(&vi->rq[i]);
 640         }
 641
 642         return 0;
 643 }
 644
 645 static void free_old_xmit_skbs(struct send_queue *sq)
 646 {
 647         struct sk_buff *skb;
 648         unsigned int len;
 649         struct virtnet_info *vi = sq->vq->vdev->priv;
 650         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 651
 652         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 653                 pr_debug("Sent skb %p\n", skb);
 654
 655                 u64_stats_update_begin(&stats->tx_syncp);
 656                 stats->tx_bytes += skb->len;
 657                 stats->tx_packets++;
 658                 u64_stats_update_end(&stats->tx_syncp);
 659
 660                 dev_kfree_skb_any(skb);
 661         }
 662 }
 663
 664 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 665 {
 666         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 667         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 668         struct virtnet_info *vi = sq->vq->vdev->priv;
 669         unsigned num_sg;
 670
 671         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 672
 673         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 674                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 675                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 676                 hdr->hdr.csum_offset = skb->csum_offset;
 677         } else {
 678                 hdr->hdr.flags = 0;
 679                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 680         }
 681
 682         if (skb_is_gso(skb)) {
 683                 hdr->hdr.hdr_len = skb_headlen(skb);
 684                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 685                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 686                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 687                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 688                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 689                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 690                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 691                 else
 692                         BUG();
 693                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 694                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 695         } else {
 696                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 697                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 698         }
 699
 700         hdr->mhdr.num_buffers = 0;
 701
 702         /* Encode metadata header at front. */
 703         if (vi->mergeable_rx_bufs)
 704                 sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
 705         else
 706                 sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
 707
 708         num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 709         return virtqueue_add_buf(sq->vq, sq->sg, num_sg,
 710                                  0, skb, GFP_ATOMIC);
 711 }
 712
 713 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 714 {
 715         struct virtnet_info *vi = netdev_priv(dev);
 716         int qnum = skb_get_queue_mapping(skb);
 717         struct send_queue *sq = &vi->sq[qnum];
 718         int err;
 719
 720         /* Free up any pending old buffers before queueing new ones. */
 721         free_old_xmit_skbs(sq);
 722
 723         /* Try to transmit */
 724         err = xmit_skb(sq, skb);
 725
 726         /* This should not happen! */
 727         if (unlikely(err)) {
 728                 dev->stats.tx_fifo_errors++;
 729                 if (net_ratelimit())
 730                         dev_warn(&dev->dev,
 731                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 732                 dev->stats.tx_dropped++;
 733                 kfree_skb(skb);
 734                 return NETDEV_TX_OK;
 735         }
 736         virtqueue_kick(sq->vq);
 737
 738         /* Don't wait up for transmitted skbs to be freed. */
 739         skb_orphan(skb);
 740         nf_reset(skb);
 741
 742         /* Apparently nice girls don't return TX_BUSY; stop the queue
 743          * before it gets out of hand.  Naturally, this wastes entries. */
 744         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 745                 netif_stop_subqueue(dev, qnum);
 746                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 747                         /* More just got used, free them then recheck. */
 748                         free_old_xmit_skbs(sq);
 749                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 750                                 netif_start_subqueue(dev, qnum);
 751                                 virtqueue_disable_cb(sq->vq);
 752                         }
 753                 }
 754         }
 755
 756         return NETDEV_TX_OK;
 757 }
 758
 759 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 760 {
 761         struct virtnet_info *vi = netdev_priv(dev);
 762         struct virtio_device *vdev = vi->vdev;
 763         int ret;
 764
 765         ret = eth_mac_addr(dev, p);
 766         if (ret)
 767                 return ret;
 768
 769         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
 770                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 771                                   dev->dev_addr, dev->addr_len);
 772
 773         return 0;
 774 }
 775
 776 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 777                                                struct rtnl_link_stats64 *tot)
 778 {
 779         struct virtnet_info *vi = netdev_priv(dev);
 780         int cpu;
 781         unsigned int start;
 782
 783         for_each_possible_cpu(cpu) {
 784                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 785                 u64 tpackets, tbytes, rpackets, rbytes;
 786
 787                 do {
 788                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 789                         tpackets = stats->tx_packets;
 790                         tbytes   = stats->tx_bytes;
 791                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 792
 793                 do {
 794                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 795                         rpackets = stats->rx_packets;
 796                         rbytes   = stats->rx_bytes;
 797                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 798
 799                 tot->rx_packets += rpackets;
 800                 tot->tx_packets += tpackets;
 801                 tot->rx_bytes   += rbytes;
 802                 tot->tx_bytes   += tbytes;
 803         }
 804
 805         tot->tx_dropped = dev->stats.tx_dropped;
 806         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 807         tot->rx_dropped = dev->stats.rx_dropped;
 808         tot->rx_length_errors = dev->stats.rx_length_errors;
 809         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 810
 811         return tot;
 812 }
 813
 814 #ifdef CONFIG_NET_POLL_CONTROLLER
 815 static void virtnet_netpoll(struct net_device *dev)
 816 {
 817         struct virtnet_info *vi = netdev_priv(dev);
 818         int i;
 819
 820         for (i = 0; i < vi->curr_queue_pairs; i++)
 821                 napi_schedule(&vi->rq[i].napi);
 822 }
 823 #endif
 824
 825 /*
 826  * Send command via the control virtqueue and check status.  Commands
 827  * supported by the hypervisor, as indicated by feature bits, should
 828  * never fail unless improperly formated.
 829  */
 830 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 831                                  struct scatterlist *data, int out, int in)
 832 {
 833         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
 834         struct virtio_net_ctrl_hdr ctrl;
 835         virtio_net_ctrl_ack status = ~0;
 836         unsigned int tmp;
 837         int i;
 838
 839         /* Caller should know better */
 840         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
 841                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
 842
 843         out++; /* Add header */
 844         in++; /* Add return status */
 845
 846         ctrl.class = class;
 847         ctrl.cmd = cmd;
 848
 849         sg_init_table(sg, out + in);
 850
 851         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
 852         for_each_sg(data, s, out + in - 2, i)
 853                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
 854         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
 855
 856         BUG_ON(virtqueue_add_buf(vi->cvq, sg, out, in, vi, GFP_ATOMIC) < 0);
 857
 858         virtqueue_kick(vi->cvq);
 859
 860         /*
 861          * Spin for a response, the kick causes an ioport write, trapping
 862          * into the hypervisor, so the request should be handled immediately.
 863          */
 864         while (!virtqueue_get_buf(vi->cvq, &tmp))
 865                 cpu_relax();
 866
 867         return status == VIRTIO_NET_OK;
 868 }
 869
 870 static void virtnet_ack_link_announce(struct virtnet_info *vi)
 871 {
 872         rtnl_lock();
 873         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
 874                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL,
 875                                   0, 0))
 876                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
 877         rtnl_unlock();
 878 }
 879
 880 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
 881 {
 882         struct scatterlist sg;
 883         struct virtio_net_ctrl_mq s;
 884         struct net_device *dev = vi->dev;
 885
 886         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
 887                 return 0;
 888
 889         s.virtqueue_pairs = queue_pairs;
 890         sg_init_one(&sg, &s, sizeof(s));
 891
 892         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
 893                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, 1, 0)){
 894                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
 895                          queue_pairs);
 896                 return -EINVAL;
 897         } else
 898                 vi->curr_queue_pairs = queue_pairs;
 899
 900         return 0;
 901 }
 902
 903 static int virtnet_close(struct net_device *dev)
 904 {
 905         struct virtnet_info *vi = netdev_priv(dev);
 906         int i;
 907
 908         /* Make sure refill_work doesn't re-enable napi! */
 909         cancel_delayed_work_sync(&vi->refill);
 910
 911         for (i = 0; i < vi->max_queue_pairs; i++)
 912                 napi_disable(&vi->rq[i].napi);
 913
 914         return 0;
 915 }
 916
 917 static void virtnet_set_rx_mode(struct net_device *dev)
 918 {
 919         struct virtnet_info *vi = netdev_priv(dev);
 920         struct scatterlist sg[2];
 921         u8 promisc, allmulti;
 922         struct virtio_net_ctrl_mac *mac_data;
 923         struct netdev_hw_addr *ha;
 924         int uc_count;
 925         int mc_count;
 926         void *buf;
 927         int i;
 928
 929         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 930         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 931                 return;
 932
 933         promisc = ((dev->flags & IFF_PROMISC) != 0);
 934         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 935
 936         sg_init_one(sg, &promisc, sizeof(promisc));
 937
 938         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 939                                   VIRTIO_NET_CTRL_RX_PROMISC,
 940                                   sg, 1, 0))
 941                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 942                          promisc ? "en" : "dis");
 943
 944         sg_init_one(sg, &allmulti, sizeof(allmulti));
 945
 946         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 947                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 948                                   sg, 1, 0))
 949                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
 950                          allmulti ? "en" : "dis");
 951
 952         uc_count = netdev_uc_count(dev);
 953         mc_count = netdev_mc_count(dev);
 954         /* MAC filter - use one buffer for both lists */
 955         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
 956                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
 957         mac_data = buf;
 958         if (!buf) {
 959                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
 960                 return;
 961         }
 962
 963         sg_init_table(sg, 2);
 964
 965         /* Store the unicast list and count in the front of the buffer */
 966         mac_data->entries = uc_count;
 967         i = 0;
 968         netdev_for_each_uc_addr(ha, dev)
 969                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 970
 971         sg_set_buf(&sg[0], mac_data,
 972                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
 973
 974         /* multicast list and count fill the end */
 975         mac_data = (void *)&mac_data->macs[uc_count][0];
 976
 977         mac_data->entries = mc_count;
 978         i = 0;
 979         netdev_for_each_mc_addr(ha, dev)
 980                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 981
 982         sg_set_buf(&sg[1], mac_data,
 983                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
 984
 985         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 986                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
 987                                   sg, 2, 0))
 988                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
 989
 990         kfree(buf);
 991 }
 992
 993 static int virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
 994 {
 995         struct virtnet_info *vi = netdev_priv(dev);
 996         struct scatterlist sg;
 997
 998         sg_init_one(&sg, &vid, sizeof(vid));
 999
1000         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1001                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
1002                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1003         return 0;
1004 }
1005
1006 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
1007 {
1008         struct virtnet_info *vi = netdev_priv(dev);
1009         struct scatterlist sg;
1010
1011         sg_init_one(&sg, &vid, sizeof(vid));
1012
1013         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1014                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
1015                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1016         return 0;
1017 }
1018
1019 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1020 {
1021         int i;
1022         int cpu;
1023
1024         if (vi->affinity_hint_set) {
1025                 for (i = 0; i < vi->max_queue_pairs; i++) {
1026                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1027                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1028                 }
1029
1030                 vi->affinity_hint_set = false;
1031         }
1032
1033         i = 0;
1034         for_each_online_cpu(cpu) {
1035                 if (cpu == hcpu) {
1036                         *per_cpu_ptr(vi->vq_index, cpu) = -1;
1037                 } else {
1038                         *per_cpu_ptr(vi->vq_index, cpu) =
1039                                 ++i % vi->curr_queue_pairs;
1040                 }
1041         }
1042 }
1043
1044 static void virtnet_set_affinity(struct virtnet_info *vi)
1045 {
1046         int i;
1047         int cpu;
1048
1049         /* In multiqueue mode, when the number of cpu is equal to the number of
1050          * queue pairs, we let the queue pairs to be private to one cpu by
1051          * setting the affinity hint to eliminate the contention.
1052          */
1053         if (vi->curr_queue_pairs == 1 ||
1054             vi->max_queue_pairs != num_online_cpus()) {
1055                 virtnet_clean_affinity(vi, -1);
1056                 return;
1057         }
1058
1059         i = 0;
1060         for_each_online_cpu(cpu) {
1061                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1062                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1063                 *per_cpu_ptr(vi->vq_index, cpu) = i;
1064                 i++;
1065         }
1066
1067         vi->affinity_hint_set = true;
1068 }
1069
1070 static void virtnet_get_ringparam(struct net_device *dev,
1071                                 struct ethtool_ringparam *ring)
1072 {
1073         struct virtnet_info *vi = netdev_priv(dev);
1074
1075         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1076         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1077         ring->rx_pending = ring->rx_max_pending;
1078         ring->tx_pending = ring->tx_max_pending;
1079 }
1080
1081
1082 static void virtnet_get_drvinfo(struct net_device *dev,
1083                                 struct ethtool_drvinfo *info)
1084 {
1085         struct virtnet_info *vi = netdev_priv(dev);
1086         struct virtio_device *vdev = vi->vdev;
1087
1088         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1089         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1090         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1091
1092 }
1093
1094 /* TODO: Eliminate OOO packets during switching */
1095 static int virtnet_set_channels(struct net_device *dev,
1096                                 struct ethtool_channels *channels)
1097 {
1098         struct virtnet_info *vi = netdev_priv(dev);
1099         u16 queue_pairs = channels->combined_count;
1100         int err;
1101
1102         /* We don't support separate rx/tx channels.
1103          * We don't allow setting 'other' channels.
1104          */
1105         if (channels->rx_count || channels->tx_count || channels->other_count)
1106                 return -EINVAL;
1107
1108         if (queue_pairs > vi->max_queue_pairs)
1109                 return -EINVAL;
1110
1111         get_online_cpus();
1112         err = virtnet_set_queues(vi, queue_pairs);
1113         if (!err) {
1114                 netif_set_real_num_tx_queues(dev, queue_pairs);
1115                 netif_set_real_num_rx_queues(dev, queue_pairs);
1116
1117                 virtnet_set_affinity(vi);
1118         }
1119         put_online_cpus();
1120
1121         return err;
1122 }
1123
1124 static void virtnet_get_channels(struct net_device *dev,
1125                                  struct ethtool_channels *channels)
1126 {
1127         struct virtnet_info *vi = netdev_priv(dev);
1128
1129         channels->combined_count = vi->curr_queue_pairs;
1130         channels->max_combined = vi->max_queue_pairs;
1131         channels->max_other = 0;
1132         channels->rx_count = 0;
1133         channels->tx_count = 0;
1134         channels->other_count = 0;
1135 }
1136
1137 static const struct ethtool_ops virtnet_ethtool_ops = {
1138         .get_drvinfo = virtnet_get_drvinfo,
1139         .get_link = ethtool_op_get_link,
1140         .get_ringparam = virtnet_get_ringparam,
1141         .set_channels = virtnet_set_channels,
1142         .get_channels = virtnet_get_channels,
1143 };
1144
1145 #define MIN_MTU 68
1146 #define MAX_MTU 65535
1147
1148 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1149 {
1150         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1151                 return -EINVAL;
1152         dev->mtu = new_mtu;
1153         return 0;
1154 }
1155
1156 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
1157  * txq based on the processor id.
1158  */
1159 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
1160 {
1161         int txq;
1162         struct virtnet_info *vi = netdev_priv(dev);
1163
1164         if (skb_rx_queue_recorded(skb)) {
1165                 txq = skb_get_rx_queue(skb);
1166         } else {
1167                 txq = *__this_cpu_ptr(vi->vq_index);
1168                 if (txq == -1)
1169                         txq = 0;
1170         }
1171
1172         while (unlikely(txq >= dev->real_num_tx_queues))
1173                 txq -= dev->real_num_tx_queues;
1174
1175         return txq;
1176 }
1177
1178 static const struct net_device_ops virtnet_netdev = {
1179         .ndo_open            = virtnet_open,
1180         .ndo_stop            = virtnet_close,
1181         .ndo_start_xmit      = start_xmit,
1182         .ndo_validate_addr   = eth_validate_addr,
1183         .ndo_set_mac_address = virtnet_set_mac_address,
1184         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1185         .ndo_change_mtu      = virtnet_change_mtu,
1186         .ndo_get_stats64     = virtnet_stats,
1187         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1188         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1189         .ndo_select_queue     = virtnet_select_queue,
1190 #ifdef CONFIG_NET_POLL_CONTROLLER
1191         .ndo_poll_controller = virtnet_netpoll,
1192 #endif
1193 };
1194
1195 static void virtnet_config_changed_work(struct work_struct *work)
1196 {
1197         struct virtnet_info *vi =
1198                 container_of(work, struct virtnet_info, config_work);
1199         u16 v;
1200
1201         mutex_lock(&vi->config_lock);
1202         if (!vi->config_enable)
1203                 goto done;
1204
1205         if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1206                               offsetof(struct virtio_net_config, status),
1207                               &v) < 0)
1208                 goto done;
1209
1210         if (v & VIRTIO_NET_S_ANNOUNCE) {
1211                 netdev_notify_peers(vi->dev);
1212                 virtnet_ack_link_announce(vi);
1213         }
1214
1215         /* Ignore unknown (future) status bits */
1216         v &= VIRTIO_NET_S_LINK_UP;
1217
1218         if (vi->status == v)
1219                 goto done;
1220
1221         vi->status = v;
1222
1223         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1224                 netif_carrier_on(vi->dev);
1225                 netif_tx_wake_all_queues(vi->dev);
1226         } else {
1227                 netif_carrier_off(vi->dev);
1228                 netif_tx_stop_all_queues(vi->dev);
1229         }
1230 done:
1231         mutex_unlock(&vi->config_lock);
1232 }
1233
1234 static void virtnet_config_changed(struct virtio_device *vdev)
1235 {
1236         struct virtnet_info *vi = vdev->priv;
1237
1238         schedule_work(&vi->config_work);
1239 }
1240
1241 static void virtnet_free_queues(struct virtnet_info *vi)
1242 {
1243         kfree(vi->rq);
1244         kfree(vi->sq);
1245 }
1246
1247 static void free_receive_bufs(struct virtnet_info *vi)
1248 {
1249         int i;
1250
1251         for (i = 0; i < vi->max_queue_pairs; i++) {
1252                 while (vi->rq[i].pages)
1253                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1254         }
1255 }
1256
1257 static void free_unused_bufs(struct virtnet_info *vi)
1258 {
1259         void *buf;
1260         int i;
1261
1262         for (i = 0; i < vi->max_queue_pairs; i++) {
1263                 struct virtqueue *vq = vi->sq[i].vq;
1264                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1265                         dev_kfree_skb(buf);
1266         }
1267
1268         for (i = 0; i < vi->max_queue_pairs; i++) {
1269                 struct virtqueue *vq = vi->rq[i].vq;
1270
1271                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1272                         if (vi->mergeable_rx_bufs || vi->big_packets)
1273                                 give_pages(&vi->rq[i], buf);
1274                         else
1275                                 dev_kfree_skb(buf);
1276                         --vi->rq[i].num;
1277                 }
1278                 BUG_ON(vi->rq[i].num != 0);
1279         }
1280 }
1281
1282 static void virtnet_del_vqs(struct virtnet_info *vi)
1283 {
1284         struct virtio_device *vdev = vi->vdev;
1285
1286         virtnet_clean_affinity(vi, -1);
1287
1288         vdev->config->del_vqs(vdev);
1289
1290         virtnet_free_queues(vi);
1291 }
1292
1293 static int virtnet_find_vqs(struct virtnet_info *vi)
1294 {
1295         vq_callback_t **callbacks;
1296         struct virtqueue **vqs;
1297         int ret = -ENOMEM;
1298         int i, total_vqs;
1299         const char **names;
1300
1301         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1302          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1303          * possible control vq.
1304          */
1305         total_vqs = vi->max_queue_pairs * 2 +
1306                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1307
1308         /* Allocate space for find_vqs parameters */
1309         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1310         if (!vqs)
1311                 goto err_vq;
1312         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1313         if (!callbacks)
1314                 goto err_callback;
1315         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1316         if (!names)
1317                 goto err_names;
1318
1319         /* Parameters for control virtqueue, if any */
1320         if (vi->has_cvq) {
1321                 callbacks[total_vqs - 1] = NULL;
1322                 names[total_vqs - 1] = "control";
1323         }
1324
1325         /* Allocate/initialize parameters for send/receive virtqueues */
1326         for (i = 0; i < vi->max_queue_pairs; i++) {
1327                 callbacks[rxq2vq(i)] = skb_recv_done;
1328                 callbacks[txq2vq(i)] = skb_xmit_done;
1329                 sprintf(vi->rq[i].name, "input.%d", i);
1330                 sprintf(vi->sq[i].name, "output.%d", i);
1331                 names[rxq2vq(i)] = vi->rq[i].name;
1332                 names[txq2vq(i)] = vi->sq[i].name;
1333         }
1334
1335         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1336                                          names);
1337         if (ret)
1338                 goto err_find;
1339
1340         if (vi->has_cvq) {
1341                 vi->cvq = vqs[total_vqs - 1];
1342                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1343                         vi->dev->features |= NETIF_F_HW_VLAN_FILTER;
1344         }
1345
1346         for (i = 0; i < vi->max_queue_pairs; i++) {
1347                 vi->rq[i].vq = vqs[rxq2vq(i)];
1348                 vi->sq[i].vq = vqs[txq2vq(i)];
1349         }
1350
1351         kfree(names);
1352         kfree(callbacks);
1353         kfree(vqs);
1354
1355         return 0;
1356
1357 err_find:
1358         kfree(names);
1359 err_names:
1360         kfree(callbacks);
1361 err_callback:
1362         kfree(vqs);
1363 err_vq:
1364         return ret;
1365 }
1366
1367 static int virtnet_alloc_queues(struct virtnet_info *vi)
1368 {
1369         int i;
1370
1371         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1372         if (!vi->sq)
1373                 goto err_sq;
1374         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1375         if (!vi->rq)
1376                 goto err_rq;
1377
1378         INIT_DELAYED_WORK(&vi->refill, refill_work);
1379         for (i = 0; i < vi->max_queue_pairs; i++) {
1380                 vi->rq[i].pages = NULL;
1381                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1382                                napi_weight);
1383
1384                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1385                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1386         }
1387
1388         return 0;
1389
1390 err_rq:
1391         kfree(vi->sq);
1392 err_sq:
1393         return -ENOMEM;
1394 }
1395
1396 static int init_vqs(struct virtnet_info *vi)
1397 {
1398         int ret;
1399
1400         /* Allocate send & receive queues */
1401         ret = virtnet_alloc_queues(vi);
1402         if (ret)
1403                 goto err;
1404
1405         ret = virtnet_find_vqs(vi);
1406         if (ret)
1407                 goto err_free;
1408
1409         get_online_cpus();
1410         virtnet_set_affinity(vi);
1411         put_online_cpus();
1412
1413         return 0;
1414
1415 err_free:
1416         virtnet_free_queues(vi);
1417 err:
1418         return ret;
1419 }
1420
1421 static int virtnet_probe(struct virtio_device *vdev)
1422 {
1423         int i, err;
1424         struct net_device *dev;
1425         struct virtnet_info *vi;
1426         u16 max_queue_pairs;
1427
1428         /* Find if host supports multiqueue virtio_net device */
1429         err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1430                                 offsetof(struct virtio_net_config,
1431                                 max_virtqueue_pairs), &max_queue_pairs);
1432
1433         /* We need at least 2 queue's */
1434         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1435             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1436             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1437                 max_queue_pairs = 1;
1438
1439         /* Allocate ourselves a network device with room for our info */
1440         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1441         if (!dev)
1442                 return -ENOMEM;
1443
1444         /* Set up network device as normal. */
1445         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1446         dev->netdev_ops = &virtnet_netdev;
1447         dev->features = NETIF_F_HIGHDMA;
1448
1449         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1450         SET_NETDEV_DEV(dev, &vdev->dev);
1451
1452         /* Do we support "hardware" checksums? */
1453         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1454                 /* This opens up the world of extra features. */
1455                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1456                 if (csum)
1457                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1458
1459                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1460                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1461                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1462                 }
1463                 /* Individual feature bits: what can host handle? */
1464                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1465                         dev->hw_features |= NETIF_F_TSO;
1466                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1467                         dev->hw_features |= NETIF_F_TSO6;
1468                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1469                         dev->hw_features |= NETIF_F_TSO_ECN;
1470                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1471                         dev->hw_features |= NETIF_F_UFO;
1472
1473                 if (gso)
1474                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1475                 /* (!csum && gso) case will be fixed by register_netdev() */
1476         }
1477
1478         /* Configuration may specify what MAC to use.  Otherwise random. */
1479         if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1480                                   offsetof(struct virtio_net_config, mac),
1481                                   dev->dev_addr, dev->addr_len) < 0)
1482                 eth_hw_addr_random(dev);
1483
1484         /* Set up our device-specific information */
1485         vi = netdev_priv(dev);
1486         vi->dev = dev;
1487         vi->vdev = vdev;
1488         vdev->priv = vi;
1489         vi->stats = alloc_percpu(struct virtnet_stats);
1490         err = -ENOMEM;
1491         if (vi->stats == NULL)
1492                 goto free;
1493
1494         vi->vq_index = alloc_percpu(int);
1495         if (vi->vq_index == NULL)
1496                 goto free_stats;
1497
1498         mutex_init(&vi->config_lock);
1499         vi->config_enable = true;
1500         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1501
1502         /* If we can receive ANY GSO packets, we must allocate large ones. */
1503         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1504             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1505             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1506                 vi->big_packets = true;
1507
1508         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1509                 vi->mergeable_rx_bufs = true;
1510
1511         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1512                 vi->has_cvq = true;
1513
1514         /* Use single tx/rx queue pair as default */
1515         vi->curr_queue_pairs = 1;
1516         vi->max_queue_pairs = max_queue_pairs;
1517
1518         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1519         err = init_vqs(vi);
1520         if (err)
1521                 goto free_index;
1522
1523         netif_set_real_num_tx_queues(dev, 1);
1524         netif_set_real_num_rx_queues(dev, 1);
1525
1526         err = register_netdev(dev);
1527         if (err) {
1528                 pr_debug("virtio_net: registering device failed\n");
1529                 goto free_vqs;
1530         }
1531
1532         /* Last of all, set up some receive buffers. */
1533         for (i = 0; i < vi->max_queue_pairs; i++) {
1534                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1535
1536                 /* If we didn't even get one input buffer, we're useless. */
1537                 if (vi->rq[i].num == 0) {
1538                         free_unused_bufs(vi);
1539                         err = -ENOMEM;
1540                         goto free_recv_bufs;
1541                 }
1542         }
1543
1544         /* Assume link up if device can't report link status,
1545            otherwise get link status from config. */
1546         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1547                 netif_carrier_off(dev);
1548                 schedule_work(&vi->config_work);
1549         } else {
1550                 vi->status = VIRTIO_NET_S_LINK_UP;
1551                 netif_carrier_on(dev);
1552         }
1553
1554         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1555                  dev->name, max_queue_pairs);
1556
1557         return 0;
1558
1559 free_recv_bufs:
1560         free_receive_bufs(vi);
1561         unregister_netdev(dev);
1562 free_vqs:
1563         cancel_delayed_work_sync(&vi->refill);
1564         virtnet_del_vqs(vi);
1565 free_index:
1566         free_percpu(vi->vq_index);
1567 free_stats:
1568         free_percpu(vi->stats);
1569 free:
1570         free_netdev(dev);
1571         return err;
1572 }
1573
1574 static void remove_vq_common(struct virtnet_info *vi)
1575 {
1576         vi->vdev->config->reset(vi->vdev);
1577
1578         /* Free unused buffers in both send and recv, if any. */
1579         free_unused_bufs(vi);
1580
1581         free_receive_bufs(vi);
1582
1583         virtnet_del_vqs(vi);
1584 }
1585
1586 static void virtnet_remove(struct virtio_device *vdev)
1587 {
1588         struct virtnet_info *vi = vdev->priv;
1589
1590         /* Prevent config work handler from accessing the device. */
1591         mutex_lock(&vi->config_lock);
1592         vi->config_enable = false;
1593         mutex_unlock(&vi->config_lock);
1594
1595         unregister_netdev(vi->dev);
1596
1597         remove_vq_common(vi);
1598
1599         flush_work(&vi->config_work);
1600
1601         free_percpu(vi->vq_index);
1602         free_percpu(vi->stats);
1603         free_netdev(vi->dev);
1604 }
1605
1606 #ifdef CONFIG_PM
1607 static int virtnet_freeze(struct virtio_device *vdev)
1608 {
1609         struct virtnet_info *vi = vdev->priv;
1610         int i;
1611
1612         /* Prevent config work handler from accessing the device */
1613         mutex_lock(&vi->config_lock);
1614         vi->config_enable = false;
1615         mutex_unlock(&vi->config_lock);
1616
1617         netif_device_detach(vi->dev);
1618         cancel_delayed_work_sync(&vi->refill);
1619
1620         if (netif_running(vi->dev))
1621                 for (i = 0; i < vi->max_queue_pairs; i++) {
1622                         napi_disable(&vi->rq[i].napi);
1623                         netif_napi_del(&vi->rq[i].napi);
1624                 }
1625
1626         remove_vq_common(vi);
1627
1628         flush_work(&vi->config_work);
1629
1630         return 0;
1631 }
1632
1633 static int virtnet_restore(struct virtio_device *vdev)
1634 {
1635         struct virtnet_info *vi = vdev->priv;
1636         int err, i;
1637
1638         err = init_vqs(vi);
1639         if (err)
1640                 return err;
1641
1642         if (netif_running(vi->dev))
1643                 for (i = 0; i < vi->max_queue_pairs; i++)
1644                         virtnet_napi_enable(&vi->rq[i]);
1645
1646         netif_device_attach(vi->dev);
1647
1648         for (i = 0; i < vi->max_queue_pairs; i++)
1649                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1650                         schedule_delayed_work(&vi->refill, 0);
1651
1652         mutex_lock(&vi->config_lock);
1653         vi->config_enable = true;
1654         mutex_unlock(&vi->config_lock);
1655
1656         virtnet_set_queues(vi, vi->curr_queue_pairs);
1657
1658         return 0;
1659 }
1660 #endif
1661
1662 static struct virtio_device_id id_table[] = {
1663         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1664         { 0 },
1665 };
1666
1667 static unsigned int features[] = {
1668         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1669         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1670         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1671         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1672         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1673         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1674         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1675         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1676 };
1677
1678 static struct virtio_driver virtio_net_driver = {
1679         .feature_table = features,
1680         .feature_table_size = ARRAY_SIZE(features),
1681         .driver.name =  KBUILD_MODNAME,
1682         .driver.owner = THIS_MODULE,
1683         .id_table =     id_table,
1684         .probe =        virtnet_probe,
1685         .remove =       virtnet_remove,
1686         .config_changed = virtnet_config_changed,
1687 #ifdef CONFIG_PM
1688         .freeze =       virtnet_freeze,
1689         .restore =      virtnet_restore,
1690 #endif
1691 };
1692
1693 static int __init init(void)
1694 {
1695         return register_virtio_driver(&virtio_net_driver);
1696 }
1697
1698 static void __exit fini(void)
1699 {
1700         unregister_virtio_driver(&virtio_net_driver);
1701 }
1702 module_init(init);
1703 module_exit(fini);
1704
1705 MODULE_DEVICE_TABLE(virtio, id_table);
1706 MODULE_DESCRIPTION("Virtio network driver");
1707 MODULE_LICENSE("GPL");