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 #include <linux/cpu.h>
  30
  31 static int napi_weight = NAPI_POLL_WEIGHT;
  32 module_param(napi_weight, int, 0444);
  33
  34 static bool csum = true, gso = true;
  35 module_param(csum, bool, 0444);
  36 module_param(gso, bool, 0444);
  37
  38 /* FIXME: MTU in config. */
  39 #define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  40 #define GOOD_COPY_LEN   128
  41
  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         /* CPU hot plug notifier */
 131         struct notifier_block nb;
 132 };
 133
 134 struct skb_vnet_hdr {
 135         union {
 136                 struct virtio_net_hdr hdr;
 137                 struct virtio_net_hdr_mrg_rxbuf mhdr;
 138         };
 139 };
 140
 141 struct padded_vnet_hdr {
 142         struct virtio_net_hdr hdr;
 143         /*
 144          * virtio_net_hdr should be in a separated sg buffer because of a
 145          * QEMU bug, and data sg buffer shares same page with this header sg.
 146          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
 147          */
 148         char padding[6];
 149 };
 150
 151 /* Converting between virtqueue no. and kernel tx/rx queue no.
 152  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 153  */
 154 static int vq2txq(struct virtqueue *vq)
 155 {
 156         return (vq->index - 1) / 2;
 157 }
 158
 159 static int txq2vq(int txq)
 160 {
 161         return txq * 2 + 1;
 162 }
 163
 164 static int vq2rxq(struct virtqueue *vq)
 165 {
 166         return vq->index / 2;
 167 }
 168
 169 static int rxq2vq(int rxq)
 170 {
 171         return rxq * 2;
 172 }
 173
 174 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 175 {
 176         return (struct skb_vnet_hdr *)skb->cb;
 177 }
 178
 179 /*
 180  * private is used to chain pages for big packets, put the whole
 181  * most recent used list in the beginning for reuse
 182  */
 183 static void give_pages(struct receive_queue *rq, struct page *page)
 184 {
 185         struct page *end;
 186
 187         /* Find end of list, sew whole thing into vi->rq.pages. */
 188         for (end = page; end->private; end = (struct page *)end->private);
 189         end->private = (unsigned long)rq->pages;
 190         rq->pages = page;
 191 }
 192
 193 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 194 {
 195         struct page *p = rq->pages;
 196
 197         if (p) {
 198                 rq->pages = (struct page *)p->private;
 199                 /* clear private here, it is used to chain pages */
 200                 p->private = 0;
 201         } else
 202                 p = alloc_page(gfp_mask);
 203         return p;
 204 }
 205
 206 static void skb_xmit_done(struct virtqueue *vq)
 207 {
 208         struct virtnet_info *vi = vq->vdev->priv;
 209
 210         /* Suppress further interrupts. */
 211         virtqueue_disable_cb(vq);
 212
 213         /* We were probably waiting for more output buffers. */
 214         netif_wake_subqueue(vi->dev, vq2txq(vq));
 215 }
 216
 217 static void set_skb_frag(struct sk_buff *skb, struct page *page,
 218                          unsigned int offset, unsigned int *len)
 219 {
 220         int size = min((unsigned)PAGE_SIZE - offset, *len);
 221         int i = skb_shinfo(skb)->nr_frags;
 222
 223         __skb_fill_page_desc(skb, i, page, offset, size);
 224
 225         skb->data_len += size;
 226         skb->len += size;
 227         skb->truesize += PAGE_SIZE;
 228         skb_shinfo(skb)->nr_frags++;
 229         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 230         *len -= size;
 231 }
 232
 233 /* Called from bottom half context */
 234 static struct sk_buff *page_to_skb(struct receive_queue *rq,
 235                                    struct page *page, unsigned int len)
 236 {
 237         struct virtnet_info *vi = rq->vq->vdev->priv;
 238         struct sk_buff *skb;
 239         struct skb_vnet_hdr *hdr;
 240         unsigned int copy, hdr_len, offset;
 241         char *p;
 242
 243         p = page_address(page);
 244
 245         /* copy small packet so we can reuse these pages for small data */
 246         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 247         if (unlikely(!skb))
 248                 return NULL;
 249
 250         hdr = skb_vnet_hdr(skb);
 251
 252         if (vi->mergeable_rx_bufs) {
 253                 hdr_len = sizeof hdr->mhdr;
 254                 offset = hdr_len;
 255         } else {
 256                 hdr_len = sizeof hdr->hdr;
 257                 offset = sizeof(struct padded_vnet_hdr);
 258         }
 259
 260         memcpy(hdr, p, hdr_len);
 261
 262         len -= hdr_len;
 263         p += offset;
 264
 265         copy = len;
 266         if (copy > skb_tailroom(skb))
 267                 copy = skb_tailroom(skb);
 268         memcpy(skb_put(skb, copy), p, copy);
 269
 270         len -= copy;
 271         offset += copy;
 272
 273         /*
 274          * Verify that we can indeed put this data into a skb.
 275          * This is here to handle cases when the device erroneously
 276          * tries to receive more than is possible. This is usually
 277          * the case of a broken device.
 278          */
 279         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 280                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 281                 dev_kfree_skb(skb);
 282                 return NULL;
 283         }
 284
 285         while (len) {
 286                 set_skb_frag(skb, page, offset, &len);
 287                 page = (struct page *)page->private;
 288                 offset = 0;
 289         }
 290
 291         if (page)
 292                 give_pages(rq, page);
 293
 294         return skb;
 295 }
 296
 297 static struct sk_buff *receive_small(void *buf, unsigned int len)
 298 {
 299         struct sk_buff * skb = buf;
 300
 301         len -= sizeof(struct virtio_net_hdr);
 302         skb_trim(skb, len);
 303
 304         return skb;
 305 }
 306
 307 static struct sk_buff *receive_big(struct net_device *dev,
 308                                    struct receive_queue *rq,
 309                                    void *buf)
 310 {
 311         struct page *page = buf;
 312         struct sk_buff *skb = page_to_skb(rq, page, 0);
 313
 314         if (unlikely(!skb))
 315                 goto err;
 316
 317         return skb;
 318
 319 err:
 320         dev->stats.rx_dropped++;
 321         give_pages(rq, page);
 322         return NULL;
 323 }
 324
 325 static struct sk_buff *receive_mergeable(struct net_device *dev,
 326                                          struct receive_queue *rq,
 327                                          void *buf,
 328                                          unsigned int len)
 329 {
 330         struct skb_vnet_hdr *hdr = page_address(buf);
 331         int num_buf = hdr->mhdr.num_buffers;
 332         struct page *page = buf;
 333         struct sk_buff *skb = page_to_skb(rq, page, len);
 334         int i;
 335
 336         if (unlikely(!skb))
 337                 goto err_skb;
 338
 339         while (--num_buf) {
 340                 i = skb_shinfo(skb)->nr_frags;
 341                 if (i >= MAX_SKB_FRAGS) {
 342                         pr_debug("%s: packet too long\n", skb->dev->name);
 343                         skb->dev->stats.rx_length_errors++;
 344                         goto err_frags;
 345                 }
 346                 page = virtqueue_get_buf(rq->vq, &len);
 347                 if (!page) {
 348                         pr_debug("%s: rx error: %d buffers %d missing\n",
 349                                  dev->name, hdr->mhdr.num_buffers, num_buf);
 350                         dev->stats.rx_length_errors++;
 351                         goto err_buf;
 352                 }
 353
 354                 if (len > PAGE_SIZE)
 355                         len = PAGE_SIZE;
 356
 357                 set_skb_frag(skb, page, 0, &len);
 358
 359                 --rq->num;
 360         }
 361         return skb;
 362 err_skb:
 363         give_pages(rq, page);
 364         while (--num_buf) {
 365 err_frags:
 366                 buf = virtqueue_get_buf(rq->vq, &len);
 367                 if (unlikely(!buf)) {
 368                         pr_debug("%s: rx error: %d buffers missing\n",
 369                                  dev->name, num_buf);
 370                         dev->stats.rx_length_errors++;
 371                         break;
 372                 }
 373                 page = buf;
 374                 give_pages(rq, page);
 375                 --rq->num;
 376         }
 377 err_buf:
 378         dev->stats.rx_dropped++;
 379         dev_kfree_skb(skb);
 380         return NULL;
 381 }
 382
 383 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
 384 {
 385         struct virtnet_info *vi = rq->vq->vdev->priv;
 386         struct net_device *dev = vi->dev;
 387         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 388         struct sk_buff *skb;
 389         struct skb_vnet_hdr *hdr;
 390
 391         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 392                 pr_debug("%s: short packet %i\n", dev->name, len);
 393                 dev->stats.rx_length_errors++;
 394                 if (vi->mergeable_rx_bufs || vi->big_packets)
 395                         give_pages(rq, buf);
 396                 else
 397                         dev_kfree_skb(buf);
 398                 return;
 399         }
 400         if (vi->mergeable_rx_bufs)
 401                 skb = receive_mergeable(dev, rq, buf, len);
 402         else if (vi->big_packets)
 403                 skb = receive_big(dev, rq, buf);
 404         else
 405                 skb = receive_small(buf, len);
 406
 407         if (unlikely(!skb))
 408                 return;
 409
 410         hdr = skb_vnet_hdr(skb);
 411
 412         u64_stats_update_begin(&stats->rx_syncp);
 413         stats->rx_bytes += skb->len;
 414         stats->rx_packets++;
 415         u64_stats_update_end(&stats->rx_syncp);
 416
 417         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 418                 pr_debug("Needs csum!\n");
 419                 if (!skb_partial_csum_set(skb,
 420                                           hdr->hdr.csum_start,
 421                                           hdr->hdr.csum_offset))
 422                         goto frame_err;
 423         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 424                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 425         }
 426
 427         skb->protocol = eth_type_trans(skb, dev);
 428         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 429                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 430
 431         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 432                 pr_debug("GSO!\n");
 433                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 434                 case VIRTIO_NET_HDR_GSO_TCPV4:
 435                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 436                         break;
 437                 case VIRTIO_NET_HDR_GSO_UDP:
 438                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 439                         break;
 440                 case VIRTIO_NET_HDR_GSO_TCPV6:
 441                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 442                         break;
 443                 default:
 444                         net_warn_ratelimited("%s: bad gso type %u.\n",
 445                                              dev->name, hdr->hdr.gso_type);
 446                         goto frame_err;
 447                 }
 448
 449                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 450                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 451
 452                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 453                 if (skb_shinfo(skb)->gso_size == 0) {
 454                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 455                         goto frame_err;
 456                 }
 457
 458                 /* Header must be checked, and gso_segs computed. */
 459                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 460                 skb_shinfo(skb)->gso_segs = 0;
 461         }
 462
 463         netif_receive_skb(skb);
 464         return;
 465
 466 frame_err:
 467         dev->stats.rx_frame_errors++;
 468         dev_kfree_skb(skb);
 469 }
 470
 471 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
 472 {
 473         struct virtnet_info *vi = rq->vq->vdev->priv;
 474         struct sk_buff *skb;
 475         struct skb_vnet_hdr *hdr;
 476         int err;
 477
 478         skb = __netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN, gfp);
 479         if (unlikely(!skb))
 480                 return -ENOMEM;
 481
 482         skb_put(skb, MAX_PACKET_LEN);
 483
 484         hdr = skb_vnet_hdr(skb);
 485         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
 486
 487         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 488
 489         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 490         if (err < 0)
 491                 dev_kfree_skb(skb);
 492
 493         return err;
 494 }
 495
 496 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
 497 {
 498         struct page *first, *list = NULL;
 499         char *p;
 500         int i, err, offset;
 501
 502         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 503         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 504                 first = get_a_page(rq, gfp);
 505                 if (!first) {
 506                         if (list)
 507                                 give_pages(rq, list);
 508                         return -ENOMEM;
 509                 }
 510                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 511
 512                 /* chain new page in list head to match sg */
 513                 first->private = (unsigned long)list;
 514                 list = first;
 515         }
 516
 517         first = get_a_page(rq, gfp);
 518         if (!first) {
 519                 give_pages(rq, list);
 520                 return -ENOMEM;
 521         }
 522         p = page_address(first);
 523
 524         /* rq->sg[0], rq->sg[1] share the same page */
 525         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
 526         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
 527
 528         /* rq->sg[1] for data packet, from offset */
 529         offset = sizeof(struct padded_vnet_hdr);
 530         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 531
 532         /* chain first in list head */
 533         first->private = (unsigned long)list;
 534         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 535                                   first, gfp);
 536         if (err < 0)
 537                 give_pages(rq, first);
 538
 539         return err;
 540 }
 541
 542 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 543 {
 544         struct page *page;
 545         int err;
 546
 547         page = get_a_page(rq, gfp);
 548         if (!page)
 549                 return -ENOMEM;
 550
 551         sg_init_one(rq->sg, page_address(page), PAGE_SIZE);
 552
 553         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, page, gfp);
 554         if (err < 0)
 555                 give_pages(rq, page);
 556
 557         return err;
 558 }
 559
 560 /*
 561  * Returns false if we couldn't fill entirely (OOM).
 562  *
 563  * Normally run in the receive path, but can also be run from ndo_open
 564  * before we're receiving packets, or from refill_work which is
 565  * careful to disable receiving (using napi_disable).
 566  */
 567 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
 568 {
 569         struct virtnet_info *vi = rq->vq->vdev->priv;
 570         int err;
 571         bool oom;
 572
 573         do {
 574                 if (vi->mergeable_rx_bufs)
 575                         err = add_recvbuf_mergeable(rq, gfp);
 576                 else if (vi->big_packets)
 577                         err = add_recvbuf_big(rq, gfp);
 578                 else
 579                         err = add_recvbuf_small(rq, gfp);
 580
 581                 oom = err == -ENOMEM;
 582                 if (err)
 583                         break;
 584                 ++rq->num;
 585         } while (rq->vq->num_free);
 586         if (unlikely(rq->num > rq->max))
 587                 rq->max = rq->num;
 588         virtqueue_kick(rq->vq);
 589         return !oom;
 590 }
 591
 592 static void skb_recv_done(struct virtqueue *rvq)
 593 {
 594         struct virtnet_info *vi = rvq->vdev->priv;
 595         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 596
 597         /* Schedule NAPI, Suppress further interrupts if successful. */
 598         if (napi_schedule_prep(&rq->napi)) {
 599                 virtqueue_disable_cb(rvq);
 600                 __napi_schedule(&rq->napi);
 601         }
 602 }
 603
 604 static void virtnet_napi_enable(struct receive_queue *rq)
 605 {
 606         napi_enable(&rq->napi);
 607
 608         /* If all buffers were filled by other side before we napi_enabled, we
 609          * won't get another interrupt, so process any outstanding packets
 610          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 611          * We synchronize against interrupts via NAPI_STATE_SCHED */
 612         if (napi_schedule_prep(&rq->napi)) {
 613                 virtqueue_disable_cb(rq->vq);
 614                 local_bh_disable();
 615                 __napi_schedule(&rq->napi);
 616                 local_bh_enable();
 617         }
 618 }
 619
 620 static void refill_work(struct work_struct *work)
 621 {
 622         struct virtnet_info *vi =
 623                 container_of(work, struct virtnet_info, refill.work);
 624         bool still_empty;
 625         int i;
 626
 627         for (i = 0; i < vi->curr_queue_pairs; i++) {
 628                 struct receive_queue *rq = &vi->rq[i];
 629
 630                 napi_disable(&rq->napi);
 631                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 632                 virtnet_napi_enable(rq);
 633
 634                 /* In theory, this can happen: if we don't get any buffers in
 635                  * we will *never* try to fill again.
 636                  */
 637                 if (still_empty)
 638                         schedule_delayed_work(&vi->refill, HZ/2);
 639         }
 640 }
 641
 642 static int virtnet_poll(struct napi_struct *napi, int budget)
 643 {
 644         struct receive_queue *rq =
 645                 container_of(napi, struct receive_queue, napi);
 646         struct virtnet_info *vi = rq->vq->vdev->priv;
 647         void *buf;
 648         unsigned int r, len, received = 0;
 649
 650 again:
 651         while (received < budget &&
 652                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 653                 receive_buf(rq, buf, len);
 654                 --rq->num;
 655                 received++;
 656         }
 657
 658         if (rq->num < rq->max / 2) {
 659                 if (!try_fill_recv(rq, GFP_ATOMIC))
 660                         schedule_delayed_work(&vi->refill, 0);
 661         }
 662
 663         /* Out of packets? */
 664         if (received < budget) {
 665                 r = virtqueue_enable_cb_prepare(rq->vq);
 666                 napi_complete(napi);
 667                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 668                     napi_schedule_prep(napi)) {
 669                         virtqueue_disable_cb(rq->vq);
 670                         __napi_schedule(napi);
 671                         goto again;
 672                 }
 673         }
 674
 675         return received;
 676 }
 677
 678 static int virtnet_open(struct net_device *dev)
 679 {
 680         struct virtnet_info *vi = netdev_priv(dev);
 681         int i;
 682
 683         for (i = 0; i < vi->max_queue_pairs; i++) {
 684                 if (i < vi->curr_queue_pairs)
 685                         /* Make sure we have some buffers: if oom use wq. */
 686                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 687                                 schedule_delayed_work(&vi->refill, 0);
 688                 virtnet_napi_enable(&vi->rq[i]);
 689         }
 690
 691         return 0;
 692 }
 693
 694 static void free_old_xmit_skbs(struct send_queue *sq)
 695 {
 696         struct sk_buff *skb;
 697         unsigned int len;
 698         struct virtnet_info *vi = sq->vq->vdev->priv;
 699         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 700
 701         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 702                 pr_debug("Sent skb %p\n", skb);
 703
 704                 u64_stats_update_begin(&stats->tx_syncp);
 705                 stats->tx_bytes += skb->len;
 706                 stats->tx_packets++;
 707                 u64_stats_update_end(&stats->tx_syncp);
 708
 709                 dev_kfree_skb_any(skb);
 710         }
 711 }
 712
 713 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 714 {
 715         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 716         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 717         struct virtnet_info *vi = sq->vq->vdev->priv;
 718         unsigned num_sg;
 719
 720         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 721
 722         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 723                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 724                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 725                 hdr->hdr.csum_offset = skb->csum_offset;
 726         } else {
 727                 hdr->hdr.flags = 0;
 728                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 729         }
 730
 731         if (skb_is_gso(skb)) {
 732                 hdr->hdr.hdr_len = skb_headlen(skb);
 733                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 734                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 735                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 736                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 737                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 738                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 739                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 740                 else
 741                         BUG();
 742                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 743                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 744         } else {
 745                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 746                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 747         }
 748
 749         hdr->mhdr.num_buffers = 0;
 750
 751         /* Encode metadata header at front. */
 752         if (vi->mergeable_rx_bufs)
 753                 sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
 754         else
 755                 sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
 756
 757         num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 758         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 759 }
 760
 761 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 762 {
 763         struct virtnet_info *vi = netdev_priv(dev);
 764         int qnum = skb_get_queue_mapping(skb);
 765         struct send_queue *sq = &vi->sq[qnum];
 766         int err;
 767
 768         /* Free up any pending old buffers before queueing new ones. */
 769         free_old_xmit_skbs(sq);
 770
 771         /* Try to transmit */
 772         err = xmit_skb(sq, skb);
 773
 774         /* This should not happen! */
 775         if (unlikely(err)) {
 776                 dev->stats.tx_fifo_errors++;
 777                 if (net_ratelimit())
 778                         dev_warn(&dev->dev,
 779                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 780                 dev->stats.tx_dropped++;
 781                 kfree_skb(skb);
 782                 return NETDEV_TX_OK;
 783         }
 784         virtqueue_kick(sq->vq);
 785
 786         /* Don't wait up for transmitted skbs to be freed. */
 787         skb_orphan(skb);
 788         nf_reset(skb);
 789
 790         /* Apparently nice girls don't return TX_BUSY; stop the queue
 791          * before it gets out of hand.  Naturally, this wastes entries. */
 792         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 793                 netif_stop_subqueue(dev, qnum);
 794                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 795                         /* More just got used, free them then recheck. */
 796                         free_old_xmit_skbs(sq);
 797                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 798                                 netif_start_subqueue(dev, qnum);
 799                                 virtqueue_disable_cb(sq->vq);
 800                         }
 801                 }
 802         }
 803
 804         return NETDEV_TX_OK;
 805 }
 806
 807 /*
 808  * Send command via the control virtqueue and check status.  Commands
 809  * supported by the hypervisor, as indicated by feature bits, should
 810  * never fail unless improperly formated.
 811  */
 812 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 813                                  struct scatterlist *out,
 814                                  struct scatterlist *in)
 815 {
 816         struct scatterlist *sgs[4], hdr, stat;
 817         struct virtio_net_ctrl_hdr ctrl;
 818         virtio_net_ctrl_ack status = ~0;
 819         unsigned out_num = 0, in_num = 0, tmp;
 820
 821         /* Caller should know better */
 822         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 823
 824         ctrl.class = class;
 825         ctrl.cmd = cmd;
 826         /* Add header */
 827         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 828         sgs[out_num++] = &hdr;
 829
 830         if (out)
 831                 sgs[out_num++] = out;
 832         if (in)
 833                 sgs[out_num + in_num++] = in;
 834
 835         /* Add return status. */
 836         sg_init_one(&stat, &status, sizeof(status));
 837         sgs[out_num + in_num++] = &stat;
 838
 839         BUG_ON(out_num + in_num > ARRAY_SIZE(sgs));
 840         BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, in_num, vi, GFP_ATOMIC)
 841                < 0);
 842
 843         virtqueue_kick(vi->cvq);
 844
 845         /* Spin for a response, the kick causes an ioport write, trapping
 846          * into the hypervisor, so the request should be handled immediately.
 847          */
 848         while (!virtqueue_get_buf(vi->cvq, &tmp))
 849                 cpu_relax();
 850
 851         return status == VIRTIO_NET_OK;
 852 }
 853
 854 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 855 {
 856         struct virtnet_info *vi = netdev_priv(dev);
 857         struct virtio_device *vdev = vi->vdev;
 858         int ret;
 859         struct sockaddr *addr = p;
 860         struct scatterlist sg;
 861
 862         ret = eth_prepare_mac_addr_change(dev, p);
 863         if (ret)
 864                 return ret;
 865
 866         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 867                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 868                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 869                                           VIRTIO_NET_CTRL_MAC_ADDR_SET,
 870                                           &sg, NULL)) {
 871                         dev_warn(&vdev->dev,
 872                                  "Failed to set mac address by vq command.\n");
 873                         return -EINVAL;
 874                 }
 875         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 876                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 877                                   addr->sa_data, dev->addr_len);
 878         }
 879
 880         eth_commit_mac_addr_change(dev, p);
 881
 882         return 0;
 883 }
 884
 885 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 886                                                struct rtnl_link_stats64 *tot)
 887 {
 888         struct virtnet_info *vi = netdev_priv(dev);
 889         int cpu;
 890         unsigned int start;
 891
 892         for_each_possible_cpu(cpu) {
 893                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 894                 u64 tpackets, tbytes, rpackets, rbytes;
 895
 896                 do {
 897                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 898                         tpackets = stats->tx_packets;
 899                         tbytes   = stats->tx_bytes;
 900                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 901
 902                 do {
 903                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 904                         rpackets = stats->rx_packets;
 905                         rbytes   = stats->rx_bytes;
 906                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 907
 908                 tot->rx_packets += rpackets;
 909                 tot->tx_packets += tpackets;
 910                 tot->rx_bytes   += rbytes;
 911                 tot->tx_bytes   += tbytes;
 912         }
 913
 914         tot->tx_dropped = dev->stats.tx_dropped;
 915         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 916         tot->rx_dropped = dev->stats.rx_dropped;
 917         tot->rx_length_errors = dev->stats.rx_length_errors;
 918         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 919
 920         return tot;
 921 }
 922
 923 #ifdef CONFIG_NET_POLL_CONTROLLER
 924 static void virtnet_netpoll(struct net_device *dev)
 925 {
 926         struct virtnet_info *vi = netdev_priv(dev);
 927         int i;
 928
 929         for (i = 0; i < vi->curr_queue_pairs; i++)
 930                 napi_schedule(&vi->rq[i].napi);
 931 }
 932 #endif
 933
 934 static void virtnet_ack_link_announce(struct virtnet_info *vi)
 935 {
 936         rtnl_lock();
 937         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
 938                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
 939                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
 940         rtnl_unlock();
 941 }
 942
 943 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
 944 {
 945         struct scatterlist sg;
 946         struct virtio_net_ctrl_mq s;
 947         struct net_device *dev = vi->dev;
 948
 949         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
 950                 return 0;
 951
 952         s.virtqueue_pairs = queue_pairs;
 953         sg_init_one(&sg, &s, sizeof(s));
 954
 955         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
 956                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
 957                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
 958                          queue_pairs);
 959                 return -EINVAL;
 960         } else {
 961                 vi->curr_queue_pairs = queue_pairs;
 962                 /* virtnet_open() will refill when device is going to up. */
 963                 if (dev->flags & IFF_UP)
 964                         schedule_delayed_work(&vi->refill, 0);
 965         }
 966
 967         return 0;
 968 }
 969
 970 static int virtnet_close(struct net_device *dev)
 971 {
 972         struct virtnet_info *vi = netdev_priv(dev);
 973         int i;
 974
 975         /* Make sure refill_work doesn't re-enable napi! */
 976         cancel_delayed_work_sync(&vi->refill);
 977
 978         for (i = 0; i < vi->max_queue_pairs; i++)
 979                 napi_disable(&vi->rq[i].napi);
 980
 981         return 0;
 982 }
 983
 984 static void virtnet_set_rx_mode(struct net_device *dev)
 985 {
 986         struct virtnet_info *vi = netdev_priv(dev);
 987         struct scatterlist sg[2];
 988         u8 promisc, allmulti;
 989         struct virtio_net_ctrl_mac *mac_data;
 990         struct netdev_hw_addr *ha;
 991         int uc_count;
 992         int mc_count;
 993         void *buf;
 994         int i;
 995
 996         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 997         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 998                 return;
 999
1000         promisc = ((dev->flags & IFF_PROMISC) != 0);
1001         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1002
1003         sg_init_one(sg, &promisc, sizeof(promisc));
1004
1005         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1006                                   VIRTIO_NET_CTRL_RX_PROMISC,
1007                                   sg, NULL))
1008                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1009                          promisc ? "en" : "dis");
1010
1011         sg_init_one(sg, &allmulti, sizeof(allmulti));
1012
1013         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1014                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
1015                                   sg, NULL))
1016                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1017                          allmulti ? "en" : "dis");
1018
1019         uc_count = netdev_uc_count(dev);
1020         mc_count = netdev_mc_count(dev);
1021         /* MAC filter - use one buffer for both lists */
1022         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1023                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1024         mac_data = buf;
1025         if (!buf)
1026                 return;
1027
1028         sg_init_table(sg, 2);
1029
1030         /* Store the unicast list and count in the front of the buffer */
1031         mac_data->entries = uc_count;
1032         i = 0;
1033         netdev_for_each_uc_addr(ha, dev)
1034                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1035
1036         sg_set_buf(&sg[0], mac_data,
1037                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1038
1039         /* multicast list and count fill the end */
1040         mac_data = (void *)&mac_data->macs[uc_count][0];
1041
1042         mac_data->entries = mc_count;
1043         i = 0;
1044         netdev_for_each_mc_addr(ha, dev)
1045                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1046
1047         sg_set_buf(&sg[1], mac_data,
1048                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1049
1050         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1051                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
1052                                   sg, NULL))
1053                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1054
1055         kfree(buf);
1056 }
1057
1058 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1059                                    __be16 proto, u16 vid)
1060 {
1061         struct virtnet_info *vi = netdev_priv(dev);
1062         struct scatterlist sg;
1063
1064         sg_init_one(&sg, &vid, sizeof(vid));
1065
1066         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1067                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1068                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1069         return 0;
1070 }
1071
1072 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1073                                     __be16 proto, u16 vid)
1074 {
1075         struct virtnet_info *vi = netdev_priv(dev);
1076         struct scatterlist sg;
1077
1078         sg_init_one(&sg, &vid, sizeof(vid));
1079
1080         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1081                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1082                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1083         return 0;
1084 }
1085
1086 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1087 {
1088         int i;
1089         int cpu;
1090
1091         if (vi->affinity_hint_set) {
1092                 for (i = 0; i < vi->max_queue_pairs; i++) {
1093                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1094                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1095                 }
1096
1097                 vi->affinity_hint_set = false;
1098         }
1099
1100         i = 0;
1101         for_each_online_cpu(cpu) {
1102                 if (cpu == hcpu) {
1103                         *per_cpu_ptr(vi->vq_index, cpu) = -1;
1104                 } else {
1105                         *per_cpu_ptr(vi->vq_index, cpu) =
1106                                 ++i % vi->curr_queue_pairs;
1107                 }
1108         }
1109 }
1110
1111 static void virtnet_set_affinity(struct virtnet_info *vi)
1112 {
1113         int i;
1114         int cpu;
1115
1116         /* In multiqueue mode, when the number of cpu is equal to the number of
1117          * queue pairs, we let the queue pairs to be private to one cpu by
1118          * setting the affinity hint to eliminate the contention.
1119          */
1120         if (vi->curr_queue_pairs == 1 ||
1121             vi->max_queue_pairs != num_online_cpus()) {
1122                 virtnet_clean_affinity(vi, -1);
1123                 return;
1124         }
1125
1126         i = 0;
1127         for_each_online_cpu(cpu) {
1128                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1129                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1130                 *per_cpu_ptr(vi->vq_index, cpu) = i;
1131                 i++;
1132         }
1133
1134         vi->affinity_hint_set = true;
1135 }
1136
1137 static int virtnet_cpu_callback(struct notifier_block *nfb,
1138                                 unsigned long action, void *hcpu)
1139 {
1140         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1141
1142         switch(action & ~CPU_TASKS_FROZEN) {
1143         case CPU_ONLINE:
1144         case CPU_DOWN_FAILED:
1145         case CPU_DEAD:
1146                 virtnet_set_affinity(vi);
1147                 break;
1148         case CPU_DOWN_PREPARE:
1149                 virtnet_clean_affinity(vi, (long)hcpu);
1150                 break;
1151         default:
1152                 break;
1153         }
1154
1155         return NOTIFY_OK;
1156 }
1157
1158 static void virtnet_get_ringparam(struct net_device *dev,
1159                                 struct ethtool_ringparam *ring)
1160 {
1161         struct virtnet_info *vi = netdev_priv(dev);
1162
1163         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1164         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1165         ring->rx_pending = ring->rx_max_pending;
1166         ring->tx_pending = ring->tx_max_pending;
1167 }
1168
1169
1170 static void virtnet_get_drvinfo(struct net_device *dev,
1171                                 struct ethtool_drvinfo *info)
1172 {
1173         struct virtnet_info *vi = netdev_priv(dev);
1174         struct virtio_device *vdev = vi->vdev;
1175
1176         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1177         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1178         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1179
1180 }
1181
1182 /* TODO: Eliminate OOO packets during switching */
1183 static int virtnet_set_channels(struct net_device *dev,
1184                                 struct ethtool_channels *channels)
1185 {
1186         struct virtnet_info *vi = netdev_priv(dev);
1187         u16 queue_pairs = channels->combined_count;
1188         int err;
1189
1190         /* We don't support separate rx/tx channels.
1191          * We don't allow setting 'other' channels.
1192          */
1193         if (channels->rx_count || channels->tx_count || channels->other_count)
1194                 return -EINVAL;
1195
1196         if (queue_pairs > vi->max_queue_pairs)
1197                 return -EINVAL;
1198
1199         get_online_cpus();
1200         err = virtnet_set_queues(vi, queue_pairs);
1201         if (!err) {
1202                 netif_set_real_num_tx_queues(dev, queue_pairs);
1203                 netif_set_real_num_rx_queues(dev, queue_pairs);
1204
1205                 virtnet_set_affinity(vi);
1206         }
1207         put_online_cpus();
1208
1209         return err;
1210 }
1211
1212 static void virtnet_get_channels(struct net_device *dev,
1213                                  struct ethtool_channels *channels)
1214 {
1215         struct virtnet_info *vi = netdev_priv(dev);
1216
1217         channels->combined_count = vi->curr_queue_pairs;
1218         channels->max_combined = vi->max_queue_pairs;
1219         channels->max_other = 0;
1220         channels->rx_count = 0;
1221         channels->tx_count = 0;
1222         channels->other_count = 0;
1223 }
1224
1225 static const struct ethtool_ops virtnet_ethtool_ops = {
1226         .get_drvinfo = virtnet_get_drvinfo,
1227         .get_link = ethtool_op_get_link,
1228         .get_ringparam = virtnet_get_ringparam,
1229         .set_channels = virtnet_set_channels,
1230         .get_channels = virtnet_get_channels,
1231 };
1232
1233 #define MIN_MTU 68
1234 #define MAX_MTU 65535
1235
1236 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1237 {
1238         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1239                 return -EINVAL;
1240         dev->mtu = new_mtu;
1241         return 0;
1242 }
1243
1244 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
1245  * txq based on the processor id.
1246  */
1247 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
1248 {
1249         int txq;
1250         struct virtnet_info *vi = netdev_priv(dev);
1251
1252         if (skb_rx_queue_recorded(skb)) {
1253                 txq = skb_get_rx_queue(skb);
1254         } else {
1255                 txq = *__this_cpu_ptr(vi->vq_index);
1256                 if (txq == -1)
1257                         txq = 0;
1258         }
1259
1260         while (unlikely(txq >= dev->real_num_tx_queues))
1261                 txq -= dev->real_num_tx_queues;
1262
1263         return txq;
1264 }
1265
1266 static const struct net_device_ops virtnet_netdev = {
1267         .ndo_open            = virtnet_open,
1268         .ndo_stop            = virtnet_close,
1269         .ndo_start_xmit      = start_xmit,
1270         .ndo_validate_addr   = eth_validate_addr,
1271         .ndo_set_mac_address = virtnet_set_mac_address,
1272         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1273         .ndo_change_mtu      = virtnet_change_mtu,
1274         .ndo_get_stats64     = virtnet_stats,
1275         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1276         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1277         .ndo_select_queue     = virtnet_select_queue,
1278 #ifdef CONFIG_NET_POLL_CONTROLLER
1279         .ndo_poll_controller = virtnet_netpoll,
1280 #endif
1281 };
1282
1283 static void virtnet_config_changed_work(struct work_struct *work)
1284 {
1285         struct virtnet_info *vi =
1286                 container_of(work, struct virtnet_info, config_work);
1287         u16 v;
1288
1289         mutex_lock(&vi->config_lock);
1290         if (!vi->config_enable)
1291                 goto done;
1292
1293         if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1294                               offsetof(struct virtio_net_config, status),
1295                               &v) < 0)
1296                 goto done;
1297
1298         if (v & VIRTIO_NET_S_ANNOUNCE) {
1299                 netdev_notify_peers(vi->dev);
1300                 virtnet_ack_link_announce(vi);
1301         }
1302
1303         /* Ignore unknown (future) status bits */
1304         v &= VIRTIO_NET_S_LINK_UP;
1305
1306         if (vi->status == v)
1307                 goto done;
1308
1309         vi->status = v;
1310
1311         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1312                 netif_carrier_on(vi->dev);
1313                 netif_tx_wake_all_queues(vi->dev);
1314         } else {
1315                 netif_carrier_off(vi->dev);
1316                 netif_tx_stop_all_queues(vi->dev);
1317         }
1318 done:
1319         mutex_unlock(&vi->config_lock);
1320 }
1321
1322 static void virtnet_config_changed(struct virtio_device *vdev)
1323 {
1324         struct virtnet_info *vi = vdev->priv;
1325
1326         schedule_work(&vi->config_work);
1327 }
1328
1329 static void virtnet_free_queues(struct virtnet_info *vi)
1330 {
1331         int i;
1332
1333         for (i = 0; i < vi->max_queue_pairs; i++)
1334                 netif_napi_del(&vi->rq[i].napi);
1335
1336         kfree(vi->rq);
1337         kfree(vi->sq);
1338 }
1339
1340 static void free_receive_bufs(struct virtnet_info *vi)
1341 {
1342         int i;
1343
1344         for (i = 0; i < vi->max_queue_pairs; i++) {
1345                 while (vi->rq[i].pages)
1346                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1347         }
1348 }
1349
1350 static void free_unused_bufs(struct virtnet_info *vi)
1351 {
1352         void *buf;
1353         int i;
1354
1355         for (i = 0; i < vi->max_queue_pairs; i++) {
1356                 struct virtqueue *vq = vi->sq[i].vq;
1357                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1358                         dev_kfree_skb(buf);
1359         }
1360
1361         for (i = 0; i < vi->max_queue_pairs; i++) {
1362                 struct virtqueue *vq = vi->rq[i].vq;
1363
1364                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1365                         if (vi->mergeable_rx_bufs || vi->big_packets)
1366                                 give_pages(&vi->rq[i], buf);
1367                         else
1368                                 dev_kfree_skb(buf);
1369                         --vi->rq[i].num;
1370                 }
1371                 BUG_ON(vi->rq[i].num != 0);
1372         }
1373 }
1374
1375 static void virtnet_del_vqs(struct virtnet_info *vi)
1376 {
1377         struct virtio_device *vdev = vi->vdev;
1378
1379         virtnet_clean_affinity(vi, -1);
1380
1381         vdev->config->del_vqs(vdev);
1382
1383         virtnet_free_queues(vi);
1384 }
1385
1386 static int virtnet_find_vqs(struct virtnet_info *vi)
1387 {
1388         vq_callback_t **callbacks;
1389         struct virtqueue **vqs;
1390         int ret = -ENOMEM;
1391         int i, total_vqs;
1392         const char **names;
1393
1394         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1395          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1396          * possible control vq.
1397          */
1398         total_vqs = vi->max_queue_pairs * 2 +
1399                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1400
1401         /* Allocate space for find_vqs parameters */
1402         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1403         if (!vqs)
1404                 goto err_vq;
1405         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1406         if (!callbacks)
1407                 goto err_callback;
1408         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1409         if (!names)
1410                 goto err_names;
1411
1412         /* Parameters for control virtqueue, if any */
1413         if (vi->has_cvq) {
1414                 callbacks[total_vqs - 1] = NULL;
1415                 names[total_vqs - 1] = "control";
1416         }
1417
1418         /* Allocate/initialize parameters for send/receive virtqueues */
1419         for (i = 0; i < vi->max_queue_pairs; i++) {
1420                 callbacks[rxq2vq(i)] = skb_recv_done;
1421                 callbacks[txq2vq(i)] = skb_xmit_done;
1422                 sprintf(vi->rq[i].name, "input.%d", i);
1423                 sprintf(vi->sq[i].name, "output.%d", i);
1424                 names[rxq2vq(i)] = vi->rq[i].name;
1425                 names[txq2vq(i)] = vi->sq[i].name;
1426         }
1427
1428         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1429                                          names);
1430         if (ret)
1431                 goto err_find;
1432
1433         if (vi->has_cvq) {
1434                 vi->cvq = vqs[total_vqs - 1];
1435                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1436                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1437         }
1438
1439         for (i = 0; i < vi->max_queue_pairs; i++) {
1440                 vi->rq[i].vq = vqs[rxq2vq(i)];
1441                 vi->sq[i].vq = vqs[txq2vq(i)];
1442         }
1443
1444         kfree(names);
1445         kfree(callbacks);
1446         kfree(vqs);
1447
1448         return 0;
1449
1450 err_find:
1451         kfree(names);
1452 err_names:
1453         kfree(callbacks);
1454 err_callback:
1455         kfree(vqs);
1456 err_vq:
1457         return ret;
1458 }
1459
1460 static int virtnet_alloc_queues(struct virtnet_info *vi)
1461 {
1462         int i;
1463
1464         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1465         if (!vi->sq)
1466                 goto err_sq;
1467         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1468         if (!vi->rq)
1469                 goto err_rq;
1470
1471         INIT_DELAYED_WORK(&vi->refill, refill_work);
1472         for (i = 0; i < vi->max_queue_pairs; i++) {
1473                 vi->rq[i].pages = NULL;
1474                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1475                                napi_weight);
1476
1477                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1478                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1479         }
1480
1481         return 0;
1482
1483 err_rq:
1484         kfree(vi->sq);
1485 err_sq:
1486         return -ENOMEM;
1487 }
1488
1489 static int init_vqs(struct virtnet_info *vi)
1490 {
1491         int ret;
1492
1493         /* Allocate send & receive queues */
1494         ret = virtnet_alloc_queues(vi);
1495         if (ret)
1496                 goto err;
1497
1498         ret = virtnet_find_vqs(vi);
1499         if (ret)
1500                 goto err_free;
1501
1502         get_online_cpus();
1503         virtnet_set_affinity(vi);
1504         put_online_cpus();
1505
1506         return 0;
1507
1508 err_free:
1509         virtnet_free_queues(vi);
1510 err:
1511         return ret;
1512 }
1513
1514 static int virtnet_probe(struct virtio_device *vdev)
1515 {
1516         int i, err;
1517         struct net_device *dev;
1518         struct virtnet_info *vi;
1519         u16 max_queue_pairs;
1520
1521         /* Find if host supports multiqueue virtio_net device */
1522         err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1523                                 offsetof(struct virtio_net_config,
1524                                 max_virtqueue_pairs), &max_queue_pairs);
1525
1526         /* We need at least 2 queue's */
1527         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1528             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1529             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1530                 max_queue_pairs = 1;
1531
1532         /* Allocate ourselves a network device with room for our info */
1533         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1534         if (!dev)
1535                 return -ENOMEM;
1536
1537         /* Set up network device as normal. */
1538         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1539         dev->netdev_ops = &virtnet_netdev;
1540         dev->features = NETIF_F_HIGHDMA;
1541
1542         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1543         SET_NETDEV_DEV(dev, &vdev->dev);
1544
1545         /* Do we support "hardware" checksums? */
1546         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1547                 /* This opens up the world of extra features. */
1548                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1549                 if (csum)
1550                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1551
1552                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1553                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1554                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1555                 }
1556                 /* Individual feature bits: what can host handle? */
1557                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1558                         dev->hw_features |= NETIF_F_TSO;
1559                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1560                         dev->hw_features |= NETIF_F_TSO6;
1561                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1562                         dev->hw_features |= NETIF_F_TSO_ECN;
1563                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1564                         dev->hw_features |= NETIF_F_UFO;
1565
1566                 if (gso)
1567                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1568                 /* (!csum && gso) case will be fixed by register_netdev() */
1569         }
1570
1571         dev->vlan_features = dev->features;
1572
1573         /* Configuration may specify what MAC to use.  Otherwise random. */
1574         if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1575                                   offsetof(struct virtio_net_config, mac),
1576                                   dev->dev_addr, dev->addr_len) < 0)
1577                 eth_hw_addr_random(dev);
1578
1579         /* Set up our device-specific information */
1580         vi = netdev_priv(dev);
1581         vi->dev = dev;
1582         vi->vdev = vdev;
1583         vdev->priv = vi;
1584         vi->stats = alloc_percpu(struct virtnet_stats);
1585         err = -ENOMEM;
1586         if (vi->stats == NULL)
1587                 goto free;
1588
1589         vi->vq_index = alloc_percpu(int);
1590         if (vi->vq_index == NULL)
1591                 goto free_stats;
1592
1593         mutex_init(&vi->config_lock);
1594         vi->config_enable = true;
1595         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1596
1597         /* If we can receive ANY GSO packets, we must allocate large ones. */
1598         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1599             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1600             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1601             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1602                 vi->big_packets = true;
1603
1604         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1605                 vi->mergeable_rx_bufs = true;
1606
1607         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1608                 vi->has_cvq = true;
1609
1610         /* Use single tx/rx queue pair as default */
1611         vi->curr_queue_pairs = 1;
1612         vi->max_queue_pairs = max_queue_pairs;
1613
1614         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1615         err = init_vqs(vi);
1616         if (err)
1617                 goto free_index;
1618
1619         netif_set_real_num_tx_queues(dev, 1);
1620         netif_set_real_num_rx_queues(dev, 1);
1621
1622         err = register_netdev(dev);
1623         if (err) {
1624                 pr_debug("virtio_net: registering device failed\n");
1625                 goto free_vqs;
1626         }
1627
1628         /* Last of all, set up some receive buffers. */
1629         for (i = 0; i < vi->curr_queue_pairs; i++) {
1630                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1631
1632                 /* If we didn't even get one input buffer, we're useless. */
1633                 if (vi->rq[i].num == 0) {
1634                         free_unused_bufs(vi);
1635                         err = -ENOMEM;
1636                         goto free_recv_bufs;
1637                 }
1638         }
1639
1640         vi->nb.notifier_call = &virtnet_cpu_callback;
1641         err = register_hotcpu_notifier(&vi->nb);
1642         if (err) {
1643                 pr_debug("virtio_net: registering cpu notifier failed\n");
1644                 goto free_recv_bufs;
1645         }
1646
1647         /* Assume link up if device can't report link status,
1648            otherwise get link status from config. */
1649         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1650                 netif_carrier_off(dev);
1651                 schedule_work(&vi->config_work);
1652         } else {
1653                 vi->status = VIRTIO_NET_S_LINK_UP;
1654                 netif_carrier_on(dev);
1655         }
1656
1657         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1658                  dev->name, max_queue_pairs);
1659
1660         return 0;
1661
1662 free_recv_bufs:
1663         free_receive_bufs(vi);
1664         unregister_netdev(dev);
1665 free_vqs:
1666         cancel_delayed_work_sync(&vi->refill);
1667         virtnet_del_vqs(vi);
1668 free_index:
1669         free_percpu(vi->vq_index);
1670 free_stats:
1671         free_percpu(vi->stats);
1672 free:
1673         free_netdev(dev);
1674         return err;
1675 }
1676
1677 static void remove_vq_common(struct virtnet_info *vi)
1678 {
1679         vi->vdev->config->reset(vi->vdev);
1680
1681         /* Free unused buffers in both send and recv, if any. */
1682         free_unused_bufs(vi);
1683
1684         free_receive_bufs(vi);
1685
1686         virtnet_del_vqs(vi);
1687 }
1688
1689 static void virtnet_remove(struct virtio_device *vdev)
1690 {
1691         struct virtnet_info *vi = vdev->priv;
1692
1693         unregister_hotcpu_notifier(&vi->nb);
1694
1695         /* Prevent config work handler from accessing the device. */
1696         mutex_lock(&vi->config_lock);
1697         vi->config_enable = false;
1698         mutex_unlock(&vi->config_lock);
1699
1700         unregister_netdev(vi->dev);
1701
1702         remove_vq_common(vi);
1703
1704         flush_work(&vi->config_work);
1705
1706         free_percpu(vi->vq_index);
1707         free_percpu(vi->stats);
1708         free_netdev(vi->dev);
1709 }
1710
1711 #ifdef CONFIG_PM
1712 static int virtnet_freeze(struct virtio_device *vdev)
1713 {
1714         struct virtnet_info *vi = vdev->priv;
1715         int i;
1716
1717         unregister_hotcpu_notifier(&vi->nb);
1718
1719         /* Prevent config work handler from accessing the device */
1720         mutex_lock(&vi->config_lock);
1721         vi->config_enable = false;
1722         mutex_unlock(&vi->config_lock);
1723
1724         netif_device_detach(vi->dev);
1725         cancel_delayed_work_sync(&vi->refill);
1726
1727         if (netif_running(vi->dev))
1728                 for (i = 0; i < vi->max_queue_pairs; i++) {
1729                         napi_disable(&vi->rq[i].napi);
1730                         netif_napi_del(&vi->rq[i].napi);
1731                 }
1732
1733         remove_vq_common(vi);
1734
1735         flush_work(&vi->config_work);
1736
1737         return 0;
1738 }
1739
1740 static int virtnet_restore(struct virtio_device *vdev)
1741 {
1742         struct virtnet_info *vi = vdev->priv;
1743         int err, i;
1744
1745         err = init_vqs(vi);
1746         if (err)
1747                 return err;
1748
1749         if (netif_running(vi->dev)) {
1750                 for (i = 0; i < vi->curr_queue_pairs; i++)
1751                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1752                                 schedule_delayed_work(&vi->refill, 0);
1753
1754                 for (i = 0; i < vi->max_queue_pairs; i++)
1755                         virtnet_napi_enable(&vi->rq[i]);
1756         }
1757
1758         netif_device_attach(vi->dev);
1759
1760         mutex_lock(&vi->config_lock);
1761         vi->config_enable = true;
1762         mutex_unlock(&vi->config_lock);
1763
1764         rtnl_lock();
1765         virtnet_set_queues(vi, vi->curr_queue_pairs);
1766         rtnl_unlock();
1767
1768         err = register_hotcpu_notifier(&vi->nb);
1769         if (err)
1770                 return err;
1771
1772         return 0;
1773 }
1774 #endif
1775
1776 static struct virtio_device_id id_table[] = {
1777         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1778         { 0 },
1779 };
1780
1781 static unsigned int features[] = {
1782         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1783         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1784         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1785         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1786         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1787         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1788         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1789         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1790         VIRTIO_NET_F_CTRL_MAC_ADDR,
1791 };
1792
1793 static struct virtio_driver virtio_net_driver = {
1794         .feature_table = features,
1795         .feature_table_size = ARRAY_SIZE(features),
1796         .driver.name =  KBUILD_MODNAME,
1797         .driver.owner = THIS_MODULE,
1798         .id_table =     id_table,
1799         .probe =        virtnet_probe,
1800         .remove =       virtnet_remove,
1801         .config_changed = virtnet_config_changed,
1802 #ifdef CONFIG_PM
1803         .freeze =       virtnet_freeze,
1804         .restore =      virtnet_restore,
1805 #endif
1806 };
1807
1808 module_virtio_driver(virtio_net_driver);
1809
1810 MODULE_DEVICE_TABLE(virtio, id_table);
1811 MODULE_DESCRIPTION("Virtio network driver");
1812 MODULE_LICENSE("GPL");