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, see <http://www.gnu.org/licenses/>.
  17  */
  18 //#define DEBUG
  19 #include <linux/netdevice.h>
  20 #include <linux/etherdevice.h>
  21 #include <linux/ethtool.h>
  22 #include <linux/module.h>
  23 #include <linux/virtio.h>
  24 #include <linux/virtio_net.h>
  25 #include <linux/scatterlist.h>
  26 #include <linux/if_vlan.h>
  27 #include <linux/slab.h>
  28 #include <linux/cpu.h>
  29 #include <linux/average.h>
  30 #include <net/busy_poll.h>
  31
  32 static int napi_weight = NAPI_POLL_WEIGHT;
  33 module_param(napi_weight, int, 0444);
  34
  35 static bool csum = true, gso = true;
  36 module_param(csum, bool, 0444);
  37 module_param(gso, bool, 0444);
  38
  39 /* FIXME: MTU in config. */
  40 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  41 #define GOOD_COPY_LEN   128
  42
  43 /* Weight used for the RX packet size EWMA. The average packet size is used to
  44  * determine the packet buffer size when refilling RX rings. As the entire RX
  45  * ring may be refilled at once, the weight is chosen so that the EWMA will be
  46  * insensitive to short-term, transient changes in packet size.
  47  */
  48 #define RECEIVE_AVG_WEIGHT 64
  49
  50 /* Minimum alignment for mergeable packet buffers. */
  51 #define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, 256)
  52
  53 #define VIRTNET_DRIVER_VERSION "1.0.0"
  54
  55 struct virtnet_stats {
  56         struct u64_stats_sync tx_syncp;
  57         struct u64_stats_sync rx_syncp;
  58         u64 tx_bytes;
  59         u64 tx_packets;
  60
  61         u64 rx_bytes;
  62         u64 rx_packets;
  63 };
  64
  65 /* Internal representation of a send virtqueue */
  66 struct send_queue {
  67         /* Virtqueue associated with this send _queue */
  68         struct virtqueue *vq;
  69
  70         /* TX: fragments + linear part + virtio header */
  71         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  72
  73         /* Name of the send queue: output.$index */
  74         char name[40];
  75 };
  76
  77 /* Internal representation of a receive virtqueue */
  78 struct receive_queue {
  79         /* Virtqueue associated with this receive_queue */
  80         struct virtqueue *vq;
  81
  82         struct napi_struct napi;
  83
  84         /* Chain pages by the private ptr. */
  85         struct page *pages;
  86
  87         /* Average packet length for mergeable receive buffers. */
  88         struct ewma mrg_avg_pkt_len;
  89
  90         /* Page frag for packet buffer allocation. */
  91         struct page_frag alloc_frag;
  92
  93         /* RX: fragments + linear part + virtio header */
  94         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  95
  96         /* Name of this receive queue: input.$index */
  97         char name[40];
  98 };
  99
 100 struct virtnet_info {
 101         struct virtio_device *vdev;
 102         struct virtqueue *cvq;
 103         struct net_device *dev;
 104         struct send_queue *sq;
 105         struct receive_queue *rq;
 106         unsigned int status;
 107
 108         /* Max # of queue pairs supported by the device */
 109         u16 max_queue_pairs;
 110
 111         /* # of queue pairs currently used by the driver */
 112         u16 curr_queue_pairs;
 113
 114         /* I like... big packets and I cannot lie! */
 115         bool big_packets;
 116
 117         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 118         bool mergeable_rx_bufs;
 119
 120         /* Has control virtqueue */
 121         bool has_cvq;
 122
 123         /* Host can handle any s/g split between our header and packet data */
 124         bool any_header_sg;
 125
 126         /* Packet virtio header size */
 127         u8 hdr_len;
 128
 129         /* Active statistics */
 130         struct virtnet_stats __percpu *stats;
 131
 132         /* Work struct for refilling if we run low on memory. */
 133         struct delayed_work refill;
 134
 135         /* Work struct for config space updates */
 136         struct work_struct config_work;
 137
 138         /* Does the affinity hint is set for virtqueues? */
 139         bool affinity_hint_set;
 140
 141         /* CPU hot plug notifier */
 142         struct notifier_block nb;
 143 };
 144
 145 struct padded_vnet_hdr {
 146         struct virtio_net_hdr_mrg_rxbuf hdr;
 147         /*
 148          * hdr is in a separate sg buffer, and data sg buffer shares same page
 149          * with this header sg. This padding makes next sg 16 byte aligned
 150          * after the header.
 151          */
 152         char padding[4];
 153 };
 154
 155 /* Converting between virtqueue no. and kernel tx/rx queue no.
 156  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 157  */
 158 static int vq2txq(struct virtqueue *vq)
 159 {
 160         return (vq->index - 1) / 2;
 161 }
 162
 163 static int txq2vq(int txq)
 164 {
 165         return txq * 2 + 1;
 166 }
 167
 168 static int vq2rxq(struct virtqueue *vq)
 169 {
 170         return vq->index / 2;
 171 }
 172
 173 static int rxq2vq(int rxq)
 174 {
 175         return rxq * 2;
 176 }
 177
 178 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 179 {
 180         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 181 }
 182
 183 /*
 184  * private is used to chain pages for big packets, put the whole
 185  * most recent used list in the beginning for reuse
 186  */
 187 static void give_pages(struct receive_queue *rq, struct page *page)
 188 {
 189         struct page *end;
 190
 191         /* Find end of list, sew whole thing into vi->rq.pages. */
 192         for (end = page; end->private; end = (struct page *)end->private);
 193         end->private = (unsigned long)rq->pages;
 194         rq->pages = page;
 195 }
 196
 197 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 198 {
 199         struct page *p = rq->pages;
 200
 201         if (p) {
 202                 rq->pages = (struct page *)p->private;
 203                 /* clear private here, it is used to chain pages */
 204                 p->private = 0;
 205         } else
 206                 p = alloc_page(gfp_mask);
 207         return p;
 208 }
 209
 210 static void skb_xmit_done(struct virtqueue *vq)
 211 {
 212         struct virtnet_info *vi = vq->vdev->priv;
 213
 214         /* Suppress further interrupts. */
 215         virtqueue_disable_cb(vq);
 216
 217         /* We were probably waiting for more output buffers. */
 218         netif_wake_subqueue(vi->dev, vq2txq(vq));
 219 }
 220
 221 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
 222 {
 223         unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
 224         return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
 225 }
 226
 227 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
 228 {
 229         return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
 230
 231 }
 232
 233 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
 234 {
 235         unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
 236         return (unsigned long)buf | (size - 1);
 237 }
 238
 239 /* Called from bottom half context */
 240 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 241                                    struct receive_queue *rq,
 242                                    struct page *page, unsigned int offset,
 243                                    unsigned int len, unsigned int truesize)
 244 {
 245         struct sk_buff *skb;
 246         struct virtio_net_hdr_mrg_rxbuf *hdr;
 247         unsigned int copy, hdr_len, hdr_padded_len;
 248         char *p;
 249
 250         p = page_address(page) + offset;
 251
 252         /* copy small packet so we can reuse these pages for small data */
 253         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 254         if (unlikely(!skb))
 255                 return NULL;
 256
 257         hdr = skb_vnet_hdr(skb);
 258
 259         hdr_len = vi->hdr_len;
 260         if (vi->mergeable_rx_bufs)
 261                 hdr_padded_len = sizeof *hdr;
 262         else
 263                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 264
 265         memcpy(hdr, p, hdr_len);
 266
 267         len -= hdr_len;
 268         offset += hdr_padded_len;
 269         p += hdr_padded_len;
 270
 271         copy = len;
 272         if (copy > skb_tailroom(skb))
 273                 copy = skb_tailroom(skb);
 274         memcpy(skb_put(skb, copy), p, copy);
 275
 276         len -= copy;
 277         offset += copy;
 278
 279         if (vi->mergeable_rx_bufs) {
 280                 if (len)
 281                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 282                 else
 283                         put_page(page);
 284                 return skb;
 285         }
 286
 287         /*
 288          * Verify that we can indeed put this data into a skb.
 289          * This is here to handle cases when the device erroneously
 290          * tries to receive more than is possible. This is usually
 291          * the case of a broken device.
 292          */
 293         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 294                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 295                 dev_kfree_skb(skb);
 296                 return NULL;
 297         }
 298         BUG_ON(offset >= PAGE_SIZE);
 299         while (len) {
 300                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 301                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 302                                 frag_size, truesize);
 303                 len -= frag_size;
 304                 page = (struct page *)page->private;
 305                 offset = 0;
 306         }
 307
 308         if (page)
 309                 give_pages(rq, page);
 310
 311         return skb;
 312 }
 313
 314 static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
 315 {
 316         struct sk_buff * skb = buf;
 317
 318         len -= vi->hdr_len;
 319         skb_trim(skb, len);
 320
 321         return skb;
 322 }
 323
 324 static struct sk_buff *receive_big(struct net_device *dev,
 325                                    struct virtnet_info *vi,
 326                                    struct receive_queue *rq,
 327                                    void *buf,
 328                                    unsigned int len)
 329 {
 330         struct page *page = buf;
 331         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 332
 333         if (unlikely(!skb))
 334                 goto err;
 335
 336         return skb;
 337
 338 err:
 339         dev->stats.rx_dropped++;
 340         give_pages(rq, page);
 341         return NULL;
 342 }
 343
 344 static struct sk_buff *receive_mergeable(struct net_device *dev,
 345                                          struct virtnet_info *vi,
 346                                          struct receive_queue *rq,
 347                                          unsigned long ctx,
 348                                          unsigned int len)
 349 {
 350         void *buf = mergeable_ctx_to_buf_address(ctx);
 351         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 352         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 353         struct page *page = virt_to_head_page(buf);
 354         int offset = buf - page_address(page);
 355         unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 356
 357         struct sk_buff *head_skb = page_to_skb(vi, rq, page, offset, len,
 358                                                truesize);
 359         struct sk_buff *curr_skb = head_skb;
 360
 361         if (unlikely(!curr_skb))
 362                 goto err_skb;
 363         while (--num_buf) {
 364                 int num_skb_frags;
 365
 366                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 367                 if (unlikely(!ctx)) {
 368                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 369                                  dev->name, num_buf,
 370                                  virtio16_to_cpu(vi->vdev,
 371                                                  hdr->num_buffers));
 372                         dev->stats.rx_length_errors++;
 373                         goto err_buf;
 374                 }
 375
 376                 buf = mergeable_ctx_to_buf_address(ctx);
 377                 page = virt_to_head_page(buf);
 378
 379                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 380                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 381                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 382
 383                         if (unlikely(!nskb))
 384                                 goto err_skb;
 385                         if (curr_skb == head_skb)
 386                                 skb_shinfo(curr_skb)->frag_list = nskb;
 387                         else
 388                                 curr_skb->next = nskb;
 389                         curr_skb = nskb;
 390                         head_skb->truesize += nskb->truesize;
 391                         num_skb_frags = 0;
 392                 }
 393                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 394                 if (curr_skb != head_skb) {
 395                         head_skb->data_len += len;
 396                         head_skb->len += len;
 397                         head_skb->truesize += truesize;
 398                 }
 399                 offset = buf - page_address(page);
 400                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 401                         put_page(page);
 402                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 403                                              len, truesize);
 404                 } else {
 405                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 406                                         offset, len, truesize);
 407                 }
 408         }
 409
 410         ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
 411         return head_skb;
 412
 413 err_skb:
 414         put_page(page);
 415         while (--num_buf) {
 416                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 417                 if (unlikely(!ctx)) {
 418                         pr_debug("%s: rx error: %d buffers missing\n",
 419                                  dev->name, num_buf);
 420                         dev->stats.rx_length_errors++;
 421                         break;
 422                 }
 423                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 424                 put_page(page);
 425         }
 426 err_buf:
 427         dev->stats.rx_dropped++;
 428         dev_kfree_skb(head_skb);
 429         return NULL;
 430 }
 431
 432 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 433                         void *buf, unsigned int len)
 434 {
 435         struct net_device *dev = vi->dev;
 436         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 437         struct sk_buff *skb;
 438         struct virtio_net_hdr_mrg_rxbuf *hdr;
 439
 440         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 441                 pr_debug("%s: short packet %i\n", dev->name, len);
 442                 dev->stats.rx_length_errors++;
 443                 if (vi->mergeable_rx_bufs) {
 444                         unsigned long ctx = (unsigned long)buf;
 445                         void *base = mergeable_ctx_to_buf_address(ctx);
 446                         put_page(virt_to_head_page(base));
 447                 } else if (vi->big_packets) {
 448                         give_pages(rq, buf);
 449                 } else {
 450                         dev_kfree_skb(buf);
 451                 }
 452                 return;
 453         }
 454
 455         if (vi->mergeable_rx_bufs)
 456                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
 457         else if (vi->big_packets)
 458                 skb = receive_big(dev, vi, rq, buf, len);
 459         else
 460                 skb = receive_small(vi, buf, len);
 461
 462         if (unlikely(!skb))
 463                 return;
 464
 465         hdr = skb_vnet_hdr(skb);
 466
 467         u64_stats_update_begin(&stats->rx_syncp);
 468         stats->rx_bytes += skb->len;
 469         stats->rx_packets++;
 470         u64_stats_update_end(&stats->rx_syncp);
 471
 472         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 473                 pr_debug("Needs csum!\n");
 474                 if (!skb_partial_csum_set(skb,
 475                           virtio16_to_cpu(vi->vdev, hdr->hdr.csum_start),
 476                           virtio16_to_cpu(vi->vdev, hdr->hdr.csum_offset)))
 477                         goto frame_err;
 478         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 479                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 480         }
 481
 482         skb->protocol = eth_type_trans(skb, dev);
 483         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 484                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 485
 486         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 487                 pr_debug("GSO!\n");
 488                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 489                 case VIRTIO_NET_HDR_GSO_TCPV4:
 490                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 491                         break;
 492                 case VIRTIO_NET_HDR_GSO_UDP:
 493                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 494                         break;
 495                 case VIRTIO_NET_HDR_GSO_TCPV6:
 496                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 497                         break;
 498                 default:
 499                         net_warn_ratelimited("%s: bad gso type %u.\n",
 500                                              dev->name, hdr->hdr.gso_type);
 501                         goto frame_err;
 502                 }
 503
 504                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 505                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 506
 507                 skb_shinfo(skb)->gso_size = virtio16_to_cpu(vi->vdev,
 508                                                             hdr->hdr.gso_size);
 509                 if (skb_shinfo(skb)->gso_size == 0) {
 510                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 511                         goto frame_err;
 512                 }
 513
 514                 /* Header must be checked, and gso_segs computed. */
 515                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 516                 skb_shinfo(skb)->gso_segs = 0;
 517         }
 518
 519         skb_mark_napi_id(skb, &rq->napi);
 520
 521         netif_receive_skb(skb);
 522         return;
 523
 524 frame_err:
 525         dev->stats.rx_frame_errors++;
 526         dev_kfree_skb(skb);
 527 }
 528
 529 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 530                              gfp_t gfp)
 531 {
 532         struct sk_buff *skb;
 533         struct virtio_net_hdr_mrg_rxbuf *hdr;
 534         int err;
 535
 536         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 537         if (unlikely(!skb))
 538                 return -ENOMEM;
 539
 540         skb_put(skb, GOOD_PACKET_LEN);
 541
 542         hdr = skb_vnet_hdr(skb);
 543         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 544         sg_set_buf(rq->sg, hdr, vi->hdr_len);
 545         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 546
 547         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 548         if (err < 0)
 549                 dev_kfree_skb(skb);
 550
 551         return err;
 552 }
 553
 554 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 555                            gfp_t gfp)
 556 {
 557         struct page *first, *list = NULL;
 558         char *p;
 559         int i, err, offset;
 560
 561         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 562
 563         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 564         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 565                 first = get_a_page(rq, gfp);
 566                 if (!first) {
 567                         if (list)
 568                                 give_pages(rq, list);
 569                         return -ENOMEM;
 570                 }
 571                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 572
 573                 /* chain new page in list head to match sg */
 574                 first->private = (unsigned long)list;
 575                 list = first;
 576         }
 577
 578         first = get_a_page(rq, gfp);
 579         if (!first) {
 580                 give_pages(rq, list);
 581                 return -ENOMEM;
 582         }
 583         p = page_address(first);
 584
 585         /* rq->sg[0], rq->sg[1] share the same page */
 586         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 587         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 588
 589         /* rq->sg[1] for data packet, from offset */
 590         offset = sizeof(struct padded_vnet_hdr);
 591         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 592
 593         /* chain first in list head */
 594         first->private = (unsigned long)list;
 595         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 596                                   first, gfp);
 597         if (err < 0)
 598                 give_pages(rq, first);
 599
 600         return err;
 601 }
 602
 603 static unsigned int get_mergeable_buf_len(struct ewma *avg_pkt_len)
 604 {
 605         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 606         unsigned int len;
 607
 608         len = hdr_len + clamp_t(unsigned int, ewma_read(avg_pkt_len),
 609                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 610         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 611 }
 612
 613 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 614 {
 615         struct page_frag *alloc_frag = &rq->alloc_frag;
 616         char *buf;
 617         unsigned long ctx;
 618         int err;
 619         unsigned int len, hole;
 620
 621         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 622         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 623                 return -ENOMEM;
 624
 625         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 626         ctx = mergeable_buf_to_ctx(buf, len);
 627         get_page(alloc_frag->page);
 628         alloc_frag->offset += len;
 629         hole = alloc_frag->size - alloc_frag->offset;
 630         if (hole < len) {
 631                 /* To avoid internal fragmentation, if there is very likely not
 632                  * enough space for another buffer, add the remaining space to
 633                  * the current buffer. This extra space is not included in
 634                  * the truesize stored in ctx.
 635                  */
 636                 len += hole;
 637                 alloc_frag->offset += hole;
 638         }
 639
 640         sg_init_one(rq->sg, buf, len);
 641         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 642         if (err < 0)
 643                 put_page(virt_to_head_page(buf));
 644
 645         return err;
 646 }
 647
 648 /*
 649  * Returns false if we couldn't fill entirely (OOM).
 650  *
 651  * Normally run in the receive path, but can also be run from ndo_open
 652  * before we're receiving packets, or from refill_work which is
 653  * careful to disable receiving (using napi_disable).
 654  */
 655 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 656                           gfp_t gfp)
 657 {
 658         int err;
 659         bool oom;
 660
 661         gfp |= __GFP_COLD;
 662         do {
 663                 if (vi->mergeable_rx_bufs)
 664                         err = add_recvbuf_mergeable(rq, gfp);
 665                 else if (vi->big_packets)
 666                         err = add_recvbuf_big(vi, rq, gfp);
 667                 else
 668                         err = add_recvbuf_small(vi, rq, gfp);
 669
 670                 oom = err == -ENOMEM;
 671                 if (err)
 672                         break;
 673         } while (rq->vq->num_free);
 674         virtqueue_kick(rq->vq);
 675         return !oom;
 676 }
 677
 678 static void skb_recv_done(struct virtqueue *rvq)
 679 {
 680         struct virtnet_info *vi = rvq->vdev->priv;
 681         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 682
 683         /* Schedule NAPI, Suppress further interrupts if successful. */
 684         if (napi_schedule_prep(&rq->napi)) {
 685                 virtqueue_disable_cb(rvq);
 686                 __napi_schedule(&rq->napi);
 687         }
 688 }
 689
 690 static void virtnet_napi_enable(struct receive_queue *rq)
 691 {
 692         napi_enable(&rq->napi);
 693
 694         /* If all buffers were filled by other side before we napi_enabled, we
 695          * won't get another interrupt, so process any outstanding packets
 696          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 697          * We synchronize against interrupts via NAPI_STATE_SCHED */
 698         if (napi_schedule_prep(&rq->napi)) {
 699                 virtqueue_disable_cb(rq->vq);
 700                 local_bh_disable();
 701                 __napi_schedule(&rq->napi);
 702                 local_bh_enable();
 703         }
 704 }
 705
 706 static void refill_work(struct work_struct *work)
 707 {
 708         struct virtnet_info *vi =
 709                 container_of(work, struct virtnet_info, refill.work);
 710         bool still_empty;
 711         int i;
 712
 713         for (i = 0; i < vi->curr_queue_pairs; i++) {
 714                 struct receive_queue *rq = &vi->rq[i];
 715
 716                 napi_disable(&rq->napi);
 717                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
 718                 virtnet_napi_enable(rq);
 719
 720                 /* In theory, this can happen: if we don't get any buffers in
 721                  * we will *never* try to fill again.
 722                  */
 723                 if (still_empty)
 724                         schedule_delayed_work(&vi->refill, HZ/2);
 725         }
 726 }
 727
 728 static int virtnet_receive(struct receive_queue *rq, int budget)
 729 {
 730         struct virtnet_info *vi = rq->vq->vdev->priv;
 731         unsigned int len, received = 0;
 732         void *buf;
 733
 734         while (received < budget &&
 735                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 736                 receive_buf(vi, rq, buf, len);
 737                 received++;
 738         }
 739
 740         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 741                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 742                         schedule_delayed_work(&vi->refill, 0);
 743         }
 744
 745         return received;
 746 }
 747
 748 static int virtnet_poll(struct napi_struct *napi, int budget)
 749 {
 750         struct receive_queue *rq =
 751                 container_of(napi, struct receive_queue, napi);
 752         unsigned int r, received = 0;
 753
 754         received += virtnet_receive(rq, budget - received);
 755
 756         /* Out of packets? */
 757         if (received < budget) {
 758                 r = virtqueue_enable_cb_prepare(rq->vq);
 759                 napi_complete(napi);
 760                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 761                     napi_schedule_prep(napi)) {
 762                         virtqueue_disable_cb(rq->vq);
 763                         __napi_schedule(napi);
 764                 }
 765         }
 766
 767         return received;
 768 }
 769
 770 #ifdef CONFIG_NET_RX_BUSY_POLL
 771 /* must be called with local_bh_disable()d */
 772 static int virtnet_busy_poll(struct napi_struct *napi)
 773 {
 774         struct receive_queue *rq =
 775                 container_of(napi, struct receive_queue, napi);
 776         struct virtnet_info *vi = rq->vq->vdev->priv;
 777         int r, received = 0, budget = 4;
 778
 779         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
 780                 return LL_FLUSH_FAILED;
 781
 782         if (!napi_schedule_prep(napi))
 783                 return LL_FLUSH_BUSY;
 784
 785         virtqueue_disable_cb(rq->vq);
 786
 787 again:
 788         received += virtnet_receive(rq, budget);
 789
 790         r = virtqueue_enable_cb_prepare(rq->vq);
 791         clear_bit(NAPI_STATE_SCHED, &napi->state);
 792         if (unlikely(virtqueue_poll(rq->vq, r)) &&
 793             napi_schedule_prep(napi)) {
 794                 virtqueue_disable_cb(rq->vq);
 795                 if (received < budget) {
 796                         budget -= received;
 797                         goto again;
 798                 } else {
 799                         __napi_schedule(napi);
 800                 }
 801         }
 802
 803         return received;
 804 }
 805 #endif  /* CONFIG_NET_RX_BUSY_POLL */
 806
 807 static int virtnet_open(struct net_device *dev)
 808 {
 809         struct virtnet_info *vi = netdev_priv(dev);
 810         int i;
 811
 812         for (i = 0; i < vi->max_queue_pairs; i++) {
 813                 if (i < vi->curr_queue_pairs)
 814                         /* Make sure we have some buffers: if oom use wq. */
 815                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
 816                                 schedule_delayed_work(&vi->refill, 0);
 817                 virtnet_napi_enable(&vi->rq[i]);
 818         }
 819
 820         return 0;
 821 }
 822
 823 static void free_old_xmit_skbs(struct send_queue *sq)
 824 {
 825         struct sk_buff *skb;
 826         unsigned int len;
 827         struct virtnet_info *vi = sq->vq->vdev->priv;
 828         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 829
 830         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 831                 pr_debug("Sent skb %p\n", skb);
 832
 833                 u64_stats_update_begin(&stats->tx_syncp);
 834                 stats->tx_bytes += skb->len;
 835                 stats->tx_packets++;
 836                 u64_stats_update_end(&stats->tx_syncp);
 837
 838                 dev_kfree_skb_any(skb);
 839         }
 840 }
 841
 842 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 843 {
 844         struct virtio_net_hdr_mrg_rxbuf *hdr;
 845         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 846         struct virtnet_info *vi = sq->vq->vdev->priv;
 847         unsigned num_sg;
 848         unsigned hdr_len = vi->hdr_len;
 849         bool can_push;
 850
 851         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 852
 853         can_push = vi->any_header_sg &&
 854                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 855                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 856         /* Even if we can, don't push here yet as this would skew
 857          * csum_start offset below. */
 858         if (can_push)
 859                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
 860         else
 861                 hdr = skb_vnet_hdr(skb);
 862
 863         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 864                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 865                 hdr->hdr.csum_start = cpu_to_virtio16(vi->vdev,
 866                                                 skb_checksum_start_offset(skb));
 867                 hdr->hdr.csum_offset = cpu_to_virtio16(vi->vdev,
 868                                                          skb->csum_offset);
 869         } else {
 870                 hdr->hdr.flags = 0;
 871                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 872         }
 873
 874         if (skb_is_gso(skb)) {
 875                 hdr->hdr.hdr_len = cpu_to_virtio16(vi->vdev, skb_headlen(skb));
 876                 hdr->hdr.gso_size = cpu_to_virtio16(vi->vdev,
 877                                                     skb_shinfo(skb)->gso_size);
 878                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 879                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 880                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 881                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 882                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 883                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 884                 else
 885                         BUG();
 886                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 887                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 888         } else {
 889                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 890                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 891         }
 892
 893         if (vi->mergeable_rx_bufs)
 894                 hdr->num_buffers = 0;
 895
 896         sg_init_table(sq->sg, MAX_SKB_FRAGS + 2);
 897         if (can_push) {
 898                 __skb_push(skb, hdr_len);
 899                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 900                 /* Pull header back to avoid skew in tx bytes calculations. */
 901                 __skb_pull(skb, hdr_len);
 902         } else {
 903                 sg_set_buf(sq->sg, hdr, hdr_len);
 904                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 905         }
 906         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 907 }
 908
 909 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 910 {
 911         struct virtnet_info *vi = netdev_priv(dev);
 912         int qnum = skb_get_queue_mapping(skb);
 913         struct send_queue *sq = &vi->sq[qnum];
 914         int err;
 915         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
 916         bool kick = !skb->xmit_more;
 917
 918         /* Free up any pending old buffers before queueing new ones. */
 919         free_old_xmit_skbs(sq);
 920
 921         /* Try to transmit */
 922         err = xmit_skb(sq, skb);
 923
 924         /* This should not happen! */
 925         if (unlikely(err)) {
 926                 dev->stats.tx_fifo_errors++;
 927                 if (net_ratelimit())
 928                         dev_warn(&dev->dev,
 929                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 930                 dev->stats.tx_dropped++;
 931                 dev_kfree_skb_any(skb);
 932                 return NETDEV_TX_OK;
 933         }
 934
 935         /* Don't wait up for transmitted skbs to be freed. */
 936         skb_orphan(skb);
 937         nf_reset(skb);
 938
 939         /* Apparently nice girls don't return TX_BUSY; stop the queue
 940          * before it gets out of hand.  Naturally, this wastes entries. */
 941         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 942                 netif_stop_subqueue(dev, qnum);
 943                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 944                         /* More just got used, free them then recheck. */
 945                         free_old_xmit_skbs(sq);
 946                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 947                                 netif_start_subqueue(dev, qnum);
 948                                 virtqueue_disable_cb(sq->vq);
 949                         }
 950                 }
 951         }
 952
 953         if (kick || netif_xmit_stopped(txq))
 954                 virtqueue_kick(sq->vq);
 955
 956         return NETDEV_TX_OK;
 957 }
 958
 959 /*
 960  * Send command via the control virtqueue and check status.  Commands
 961  * supported by the hypervisor, as indicated by feature bits, should
 962  * never fail unless improperly formatted.
 963  */
 964 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 965                                  struct scatterlist *out)
 966 {
 967         struct scatterlist *sgs[4], hdr, stat;
 968         struct virtio_net_ctrl_hdr ctrl;
 969         virtio_net_ctrl_ack status = ~0;
 970         unsigned out_num = 0, tmp;
 971
 972         /* Caller should know better */
 973         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 974
 975         ctrl.class = class;
 976         ctrl.cmd = cmd;
 977         /* Add header */
 978         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 979         sgs[out_num++] = &hdr;
 980
 981         if (out)
 982                 sgs[out_num++] = out;
 983
 984         /* Add return status. */
 985         sg_init_one(&stat, &status, sizeof(status));
 986         sgs[out_num] = &stat;
 987
 988         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
 989         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
 990
 991         if (unlikely(!virtqueue_kick(vi->cvq)))
 992                 return status == VIRTIO_NET_OK;
 993
 994         /* Spin for a response, the kick causes an ioport write, trapping
 995          * into the hypervisor, so the request should be handled immediately.
 996          */
 997         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
 998                !virtqueue_is_broken(vi->cvq))
 999                 cpu_relax();
1000
1001         return status == VIRTIO_NET_OK;
1002 }
1003
1004 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1005 {
1006         struct virtnet_info *vi = netdev_priv(dev);
1007         struct virtio_device *vdev = vi->vdev;
1008         int ret;
1009         struct sockaddr *addr = p;
1010         struct scatterlist sg;
1011
1012         ret = eth_prepare_mac_addr_change(dev, p);
1013         if (ret)
1014                 return ret;
1015
1016         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1017                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1018                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1019                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1020                         dev_warn(&vdev->dev,
1021                                  "Failed to set mac address by vq command.\n");
1022                         return -EINVAL;
1023                 }
1024         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1025                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1026                 unsigned int i;
1027
1028                 /* Naturally, this has an atomicity problem. */
1029                 for (i = 0; i < dev->addr_len; i++)
1030                         virtio_cwrite8(vdev,
1031                                        offsetof(struct virtio_net_config, mac) +
1032                                        i, addr->sa_data[i]);
1033         }
1034
1035         eth_commit_mac_addr_change(dev, p);
1036
1037         return 0;
1038 }
1039
1040 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1041                                                struct rtnl_link_stats64 *tot)
1042 {
1043         struct virtnet_info *vi = netdev_priv(dev);
1044         int cpu;
1045         unsigned int start;
1046
1047         for_each_possible_cpu(cpu) {
1048                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1049                 u64 tpackets, tbytes, rpackets, rbytes;
1050
1051                 do {
1052                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1053                         tpackets = stats->tx_packets;
1054                         tbytes   = stats->tx_bytes;
1055                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1056
1057                 do {
1058                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1059                         rpackets = stats->rx_packets;
1060                         rbytes   = stats->rx_bytes;
1061                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1062
1063                 tot->rx_packets += rpackets;
1064                 tot->tx_packets += tpackets;
1065                 tot->rx_bytes   += rbytes;
1066                 tot->tx_bytes   += tbytes;
1067         }
1068
1069         tot->tx_dropped = dev->stats.tx_dropped;
1070         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1071         tot->rx_dropped = dev->stats.rx_dropped;
1072         tot->rx_length_errors = dev->stats.rx_length_errors;
1073         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1074
1075         return tot;
1076 }
1077
1078 #ifdef CONFIG_NET_POLL_CONTROLLER
1079 static void virtnet_netpoll(struct net_device *dev)
1080 {
1081         struct virtnet_info *vi = netdev_priv(dev);
1082         int i;
1083
1084         for (i = 0; i < vi->curr_queue_pairs; i++)
1085                 napi_schedule(&vi->rq[i].napi);
1086 }
1087 #endif
1088
1089 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1090 {
1091         rtnl_lock();
1092         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1093                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1094                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1095         rtnl_unlock();
1096 }
1097
1098 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1099 {
1100         struct scatterlist sg;
1101         struct virtio_net_ctrl_mq s;
1102         struct net_device *dev = vi->dev;
1103
1104         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1105                 return 0;
1106
1107         s.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1108         sg_init_one(&sg, &s, sizeof(s));
1109
1110         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1111                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1112                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1113                          queue_pairs);
1114                 return -EINVAL;
1115         } else {
1116                 vi->curr_queue_pairs = queue_pairs;
1117                 /* virtnet_open() will refill when device is going to up. */
1118                 if (dev->flags & IFF_UP)
1119                         schedule_delayed_work(&vi->refill, 0);
1120         }
1121
1122         return 0;
1123 }
1124
1125 static int virtnet_close(struct net_device *dev)
1126 {
1127         struct virtnet_info *vi = netdev_priv(dev);
1128         int i;
1129
1130         /* Make sure refill_work doesn't re-enable napi! */
1131         cancel_delayed_work_sync(&vi->refill);
1132
1133         for (i = 0; i < vi->max_queue_pairs; i++)
1134                 napi_disable(&vi->rq[i].napi);
1135
1136         return 0;
1137 }
1138
1139 static void virtnet_set_rx_mode(struct net_device *dev)
1140 {
1141         struct virtnet_info *vi = netdev_priv(dev);
1142         struct scatterlist sg[2];
1143         u8 promisc, allmulti;
1144         struct virtio_net_ctrl_mac *mac_data;
1145         struct netdev_hw_addr *ha;
1146         int uc_count;
1147         int mc_count;
1148         void *buf;
1149         int i;
1150
1151         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1152         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1153                 return;
1154
1155         promisc = ((dev->flags & IFF_PROMISC) != 0);
1156         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1157
1158         sg_init_one(sg, &promisc, sizeof(promisc));
1159
1160         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1161                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1162                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1163                          promisc ? "en" : "dis");
1164
1165         sg_init_one(sg, &allmulti, sizeof(allmulti));
1166
1167         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1168                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1169                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1170                          allmulti ? "en" : "dis");
1171
1172         uc_count = netdev_uc_count(dev);
1173         mc_count = netdev_mc_count(dev);
1174         /* MAC filter - use one buffer for both lists */
1175         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1176                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1177         mac_data = buf;
1178         if (!buf)
1179                 return;
1180
1181         sg_init_table(sg, 2);
1182
1183         /* Store the unicast list and count in the front of the buffer */
1184         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1185         i = 0;
1186         netdev_for_each_uc_addr(ha, dev)
1187                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1188
1189         sg_set_buf(&sg[0], mac_data,
1190                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1191
1192         /* multicast list and count fill the end */
1193         mac_data = (void *)&mac_data->macs[uc_count][0];
1194
1195         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1196         i = 0;
1197         netdev_for_each_mc_addr(ha, dev)
1198                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1199
1200         sg_set_buf(&sg[1], mac_data,
1201                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1202
1203         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1204                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1205                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1206
1207         kfree(buf);
1208 }
1209
1210 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1211                                    __be16 proto, u16 vid)
1212 {
1213         struct virtnet_info *vi = netdev_priv(dev);
1214         struct scatterlist sg;
1215
1216         sg_init_one(&sg, &vid, sizeof(vid));
1217
1218         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1219                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1220                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1221         return 0;
1222 }
1223
1224 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1225                                     __be16 proto, u16 vid)
1226 {
1227         struct virtnet_info *vi = netdev_priv(dev);
1228         struct scatterlist sg;
1229
1230         sg_init_one(&sg, &vid, sizeof(vid));
1231
1232         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1233                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1234                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1235         return 0;
1236 }
1237
1238 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1239 {
1240         int i;
1241
1242         if (vi->affinity_hint_set) {
1243                 for (i = 0; i < vi->max_queue_pairs; i++) {
1244                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1245                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1246                 }
1247
1248                 vi->affinity_hint_set = false;
1249         }
1250 }
1251
1252 static void virtnet_set_affinity(struct virtnet_info *vi)
1253 {
1254         int i;
1255         int cpu;
1256
1257         /* In multiqueue mode, when the number of cpu is equal to the number of
1258          * queue pairs, we let the queue pairs to be private to one cpu by
1259          * setting the affinity hint to eliminate the contention.
1260          */
1261         if (vi->curr_queue_pairs == 1 ||
1262             vi->max_queue_pairs != num_online_cpus()) {
1263                 virtnet_clean_affinity(vi, -1);
1264                 return;
1265         }
1266
1267         i = 0;
1268         for_each_online_cpu(cpu) {
1269                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1270                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1271                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1272                 i++;
1273         }
1274
1275         vi->affinity_hint_set = true;
1276 }
1277
1278 static int virtnet_cpu_callback(struct notifier_block *nfb,
1279                                 unsigned long action, void *hcpu)
1280 {
1281         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1282
1283         switch(action & ~CPU_TASKS_FROZEN) {
1284         case CPU_ONLINE:
1285         case CPU_DOWN_FAILED:
1286         case CPU_DEAD:
1287                 virtnet_set_affinity(vi);
1288                 break;
1289         case CPU_DOWN_PREPARE:
1290                 virtnet_clean_affinity(vi, (long)hcpu);
1291                 break;
1292         default:
1293                 break;
1294         }
1295
1296         return NOTIFY_OK;
1297 }
1298
1299 static void virtnet_get_ringparam(struct net_device *dev,
1300                                 struct ethtool_ringparam *ring)
1301 {
1302         struct virtnet_info *vi = netdev_priv(dev);
1303
1304         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1305         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1306         ring->rx_pending = ring->rx_max_pending;
1307         ring->tx_pending = ring->tx_max_pending;
1308 }
1309
1310
1311 static void virtnet_get_drvinfo(struct net_device *dev,
1312                                 struct ethtool_drvinfo *info)
1313 {
1314         struct virtnet_info *vi = netdev_priv(dev);
1315         struct virtio_device *vdev = vi->vdev;
1316
1317         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1318         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1319         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1320
1321 }
1322
1323 /* TODO: Eliminate OOO packets during switching */
1324 static int virtnet_set_channels(struct net_device *dev,
1325                                 struct ethtool_channels *channels)
1326 {
1327         struct virtnet_info *vi = netdev_priv(dev);
1328         u16 queue_pairs = channels->combined_count;
1329         int err;
1330
1331         /* We don't support separate rx/tx channels.
1332          * We don't allow setting 'other' channels.
1333          */
1334         if (channels->rx_count || channels->tx_count || channels->other_count)
1335                 return -EINVAL;
1336
1337         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1338                 return -EINVAL;
1339
1340         get_online_cpus();
1341         err = virtnet_set_queues(vi, queue_pairs);
1342         if (!err) {
1343                 netif_set_real_num_tx_queues(dev, queue_pairs);
1344                 netif_set_real_num_rx_queues(dev, queue_pairs);
1345
1346                 virtnet_set_affinity(vi);
1347         }
1348         put_online_cpus();
1349
1350         return err;
1351 }
1352
1353 static void virtnet_get_channels(struct net_device *dev,
1354                                  struct ethtool_channels *channels)
1355 {
1356         struct virtnet_info *vi = netdev_priv(dev);
1357
1358         channels->combined_count = vi->curr_queue_pairs;
1359         channels->max_combined = vi->max_queue_pairs;
1360         channels->max_other = 0;
1361         channels->rx_count = 0;
1362         channels->tx_count = 0;
1363         channels->other_count = 0;
1364 }
1365
1366 static const struct ethtool_ops virtnet_ethtool_ops = {
1367         .get_drvinfo = virtnet_get_drvinfo,
1368         .get_link = ethtool_op_get_link,
1369         .get_ringparam = virtnet_get_ringparam,
1370         .set_channels = virtnet_set_channels,
1371         .get_channels = virtnet_get_channels,
1372 };
1373
1374 #define MIN_MTU 68
1375 #define MAX_MTU 65535
1376
1377 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1378 {
1379         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1380                 return -EINVAL;
1381         dev->mtu = new_mtu;
1382         return 0;
1383 }
1384
1385 static const struct net_device_ops virtnet_netdev = {
1386         .ndo_open            = virtnet_open,
1387         .ndo_stop            = virtnet_close,
1388         .ndo_start_xmit      = start_xmit,
1389         .ndo_validate_addr   = eth_validate_addr,
1390         .ndo_set_mac_address = virtnet_set_mac_address,
1391         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1392         .ndo_change_mtu      = virtnet_change_mtu,
1393         .ndo_get_stats64     = virtnet_stats,
1394         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1395         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1396 #ifdef CONFIG_NET_POLL_CONTROLLER
1397         .ndo_poll_controller = virtnet_netpoll,
1398 #endif
1399 #ifdef CONFIG_NET_RX_BUSY_POLL
1400         .ndo_busy_poll          = virtnet_busy_poll,
1401 #endif
1402 };
1403
1404 static void virtnet_config_changed_work(struct work_struct *work)
1405 {
1406         struct virtnet_info *vi =
1407                 container_of(work, struct virtnet_info, config_work);
1408         u16 v;
1409
1410         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1411                                  struct virtio_net_config, status, &v) < 0)
1412                 return;
1413
1414         if (v & VIRTIO_NET_S_ANNOUNCE) {
1415                 netdev_notify_peers(vi->dev);
1416                 virtnet_ack_link_announce(vi);
1417         }
1418
1419         /* Ignore unknown (future) status bits */
1420         v &= VIRTIO_NET_S_LINK_UP;
1421
1422         if (vi->status == v)
1423                 return;
1424
1425         vi->status = v;
1426
1427         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1428                 netif_carrier_on(vi->dev);
1429                 netif_tx_wake_all_queues(vi->dev);
1430         } else {
1431                 netif_carrier_off(vi->dev);
1432                 netif_tx_stop_all_queues(vi->dev);
1433         }
1434 }
1435
1436 static void virtnet_config_changed(struct virtio_device *vdev)
1437 {
1438         struct virtnet_info *vi = vdev->priv;
1439
1440         schedule_work(&vi->config_work);
1441 }
1442
1443 static void virtnet_free_queues(struct virtnet_info *vi)
1444 {
1445         int i;
1446
1447         for (i = 0; i < vi->max_queue_pairs; i++)
1448                 netif_napi_del(&vi->rq[i].napi);
1449
1450         kfree(vi->rq);
1451         kfree(vi->sq);
1452 }
1453
1454 static void free_receive_bufs(struct virtnet_info *vi)
1455 {
1456         int i;
1457
1458         for (i = 0; i < vi->max_queue_pairs; i++) {
1459                 while (vi->rq[i].pages)
1460                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1461         }
1462 }
1463
1464 static void free_receive_page_frags(struct virtnet_info *vi)
1465 {
1466         int i;
1467         for (i = 0; i < vi->max_queue_pairs; i++)
1468                 if (vi->rq[i].alloc_frag.page)
1469                         put_page(vi->rq[i].alloc_frag.page);
1470 }
1471
1472 static void free_unused_bufs(struct virtnet_info *vi)
1473 {
1474         void *buf;
1475         int i;
1476
1477         for (i = 0; i < vi->max_queue_pairs; i++) {
1478                 struct virtqueue *vq = vi->sq[i].vq;
1479                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1480                         dev_kfree_skb(buf);
1481         }
1482
1483         for (i = 0; i < vi->max_queue_pairs; i++) {
1484                 struct virtqueue *vq = vi->rq[i].vq;
1485
1486                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1487                         if (vi->mergeable_rx_bufs) {
1488                                 unsigned long ctx = (unsigned long)buf;
1489                                 void *base = mergeable_ctx_to_buf_address(ctx);
1490                                 put_page(virt_to_head_page(base));
1491                         } else if (vi->big_packets) {
1492                                 give_pages(&vi->rq[i], buf);
1493                         } else {
1494                                 dev_kfree_skb(buf);
1495                         }
1496                 }
1497         }
1498 }
1499
1500 static void virtnet_del_vqs(struct virtnet_info *vi)
1501 {
1502         struct virtio_device *vdev = vi->vdev;
1503
1504         virtnet_clean_affinity(vi, -1);
1505
1506         vdev->config->del_vqs(vdev);
1507
1508         virtnet_free_queues(vi);
1509 }
1510
1511 static int virtnet_find_vqs(struct virtnet_info *vi)
1512 {
1513         vq_callback_t **callbacks;
1514         struct virtqueue **vqs;
1515         int ret = -ENOMEM;
1516         int i, total_vqs;
1517         const char **names;
1518
1519         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1520          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1521          * possible control vq.
1522          */
1523         total_vqs = vi->max_queue_pairs * 2 +
1524                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1525
1526         /* Allocate space for find_vqs parameters */
1527         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1528         if (!vqs)
1529                 goto err_vq;
1530         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1531         if (!callbacks)
1532                 goto err_callback;
1533         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1534         if (!names)
1535                 goto err_names;
1536
1537         /* Parameters for control virtqueue, if any */
1538         if (vi->has_cvq) {
1539                 callbacks[total_vqs - 1] = NULL;
1540                 names[total_vqs - 1] = "control";
1541         }
1542
1543         /* Allocate/initialize parameters for send/receive virtqueues */
1544         for (i = 0; i < vi->max_queue_pairs; i++) {
1545                 callbacks[rxq2vq(i)] = skb_recv_done;
1546                 callbacks[txq2vq(i)] = skb_xmit_done;
1547                 sprintf(vi->rq[i].name, "input.%d", i);
1548                 sprintf(vi->sq[i].name, "output.%d", i);
1549                 names[rxq2vq(i)] = vi->rq[i].name;
1550                 names[txq2vq(i)] = vi->sq[i].name;
1551         }
1552
1553         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1554                                          names);
1555         if (ret)
1556                 goto err_find;
1557
1558         if (vi->has_cvq) {
1559                 vi->cvq = vqs[total_vqs - 1];
1560                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1561                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1562         }
1563
1564         for (i = 0; i < vi->max_queue_pairs; i++) {
1565                 vi->rq[i].vq = vqs[rxq2vq(i)];
1566                 vi->sq[i].vq = vqs[txq2vq(i)];
1567         }
1568
1569         kfree(names);
1570         kfree(callbacks);
1571         kfree(vqs);
1572
1573         return 0;
1574
1575 err_find:
1576         kfree(names);
1577 err_names:
1578         kfree(callbacks);
1579 err_callback:
1580         kfree(vqs);
1581 err_vq:
1582         return ret;
1583 }
1584
1585 static int virtnet_alloc_queues(struct virtnet_info *vi)
1586 {
1587         int i;
1588
1589         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1590         if (!vi->sq)
1591                 goto err_sq;
1592         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1593         if (!vi->rq)
1594                 goto err_rq;
1595
1596         INIT_DELAYED_WORK(&vi->refill, refill_work);
1597         for (i = 0; i < vi->max_queue_pairs; i++) {
1598                 vi->rq[i].pages = NULL;
1599                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1600                                napi_weight);
1601                 napi_hash_add(&vi->rq[i].napi);
1602
1603                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1604                 ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
1605                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1606         }
1607
1608         return 0;
1609
1610 err_rq:
1611         kfree(vi->sq);
1612 err_sq:
1613         return -ENOMEM;
1614 }
1615
1616 static int init_vqs(struct virtnet_info *vi)
1617 {
1618         int ret;
1619
1620         /* Allocate send & receive queues */
1621         ret = virtnet_alloc_queues(vi);
1622         if (ret)
1623                 goto err;
1624
1625         ret = virtnet_find_vqs(vi);
1626         if (ret)
1627                 goto err_free;
1628
1629         get_online_cpus();
1630         virtnet_set_affinity(vi);
1631         put_online_cpus();
1632
1633         return 0;
1634
1635 err_free:
1636         virtnet_free_queues(vi);
1637 err:
1638         return ret;
1639 }
1640
1641 #ifdef CONFIG_SYSFS
1642 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1643                 struct rx_queue_attribute *attribute, char *buf)
1644 {
1645         struct virtnet_info *vi = netdev_priv(queue->dev);
1646         unsigned int queue_index = get_netdev_rx_queue_index(queue);
1647         struct ewma *avg;
1648
1649         BUG_ON(queue_index >= vi->max_queue_pairs);
1650         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1651         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1652 }
1653
1654 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1655         __ATTR_RO(mergeable_rx_buffer_size);
1656
1657 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1658         &mergeable_rx_buffer_size_attribute.attr,
1659         NULL
1660 };
1661
1662 static const struct attribute_group virtio_net_mrg_rx_group = {
1663         .name = "virtio_net",
1664         .attrs = virtio_net_mrg_rx_attrs
1665 };
1666 #endif
1667
1668 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1669                                     unsigned int fbit,
1670                                     const char *fname, const char *dname)
1671 {
1672         if (!virtio_has_feature(vdev, fbit))
1673                 return false;
1674
1675         dev_err(&vdev->dev, "device advertises feature %s but not %s",
1676                 fname, dname);
1677
1678         return true;
1679 }
1680
1681 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
1682         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1683
1684 static bool virtnet_validate_features(struct virtio_device *vdev)
1685 {
1686         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1687             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1688                              "VIRTIO_NET_F_CTRL_VQ") ||
1689              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1690                              "VIRTIO_NET_F_CTRL_VQ") ||
1691              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1692                              "VIRTIO_NET_F_CTRL_VQ") ||
1693              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1694              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1695                              "VIRTIO_NET_F_CTRL_VQ"))) {
1696                 return false;
1697         }
1698
1699         return true;
1700 }
1701
1702 static int virtnet_probe(struct virtio_device *vdev)
1703 {
1704         int i, err;
1705         struct net_device *dev;
1706         struct virtnet_info *vi;
1707         u16 max_queue_pairs;
1708
1709         if (!virtnet_validate_features(vdev))
1710                 return -EINVAL;
1711
1712         /* Find if host supports multiqueue virtio_net device */
1713         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1714                                    struct virtio_net_config,
1715                                    max_virtqueue_pairs, &max_queue_pairs);
1716
1717         /* We need at least 2 queue's */
1718         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1719             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1720             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1721                 max_queue_pairs = 1;
1722
1723         /* Allocate ourselves a network device with room for our info */
1724         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1725         if (!dev)
1726                 return -ENOMEM;
1727
1728         /* Set up network device as normal. */
1729         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1730         dev->netdev_ops = &virtnet_netdev;
1731         dev->features = NETIF_F_HIGHDMA;
1732
1733         dev->ethtool_ops = &virtnet_ethtool_ops;
1734         SET_NETDEV_DEV(dev, &vdev->dev);
1735
1736         /* Do we support "hardware" checksums? */
1737         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1738                 /* This opens up the world of extra features. */
1739                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1740                 if (csum)
1741                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1742
1743                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1744                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1745                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1746                 }
1747                 /* Individual feature bits: what can host handle? */
1748                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1749                         dev->hw_features |= NETIF_F_TSO;
1750                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1751                         dev->hw_features |= NETIF_F_TSO6;
1752                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1753                         dev->hw_features |= NETIF_F_TSO_ECN;
1754                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1755                         dev->hw_features |= NETIF_F_UFO;
1756
1757                 if (gso)
1758                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1759                 /* (!csum && gso) case will be fixed by register_netdev() */
1760         }
1761         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1762                 dev->features |= NETIF_F_RXCSUM;
1763
1764         dev->vlan_features = dev->features;
1765
1766         /* Configuration may specify what MAC to use.  Otherwise random. */
1767         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1768                 virtio_cread_bytes(vdev,
1769                                    offsetof(struct virtio_net_config, mac),
1770                                    dev->dev_addr, dev->addr_len);
1771         else
1772                 eth_hw_addr_random(dev);
1773
1774         /* Set up our device-specific information */
1775         vi = netdev_priv(dev);
1776         vi->dev = dev;
1777         vi->vdev = vdev;
1778         vdev->priv = vi;
1779         vi->stats = alloc_percpu(struct virtnet_stats);
1780         err = -ENOMEM;
1781         if (vi->stats == NULL)
1782                 goto free;
1783
1784         for_each_possible_cpu(i) {
1785                 struct virtnet_stats *virtnet_stats;
1786                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1787                 u64_stats_init(&virtnet_stats->tx_syncp);
1788                 u64_stats_init(&virtnet_stats->rx_syncp);
1789         }
1790
1791         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1792
1793         /* If we can receive ANY GSO packets, we must allocate large ones. */
1794         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1795             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1796             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1797             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1798                 vi->big_packets = true;
1799
1800         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1801                 vi->mergeable_rx_bufs = true;
1802
1803         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
1804             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1805                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1806         else
1807                 vi->hdr_len = sizeof(struct virtio_net_hdr);
1808
1809         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1810                 vi->any_header_sg = true;
1811
1812         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1813                 vi->has_cvq = true;
1814
1815         if (vi->any_header_sg)
1816                 dev->needed_headroom = vi->hdr_len;
1817
1818         /* Use single tx/rx queue pair as default */
1819         vi->curr_queue_pairs = 1;
1820         vi->max_queue_pairs = max_queue_pairs;
1821
1822         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1823         err = init_vqs(vi);
1824         if (err)
1825                 goto free_stats;
1826
1827 #ifdef CONFIG_SYSFS
1828         if (vi->mergeable_rx_bufs)
1829                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1830 #endif
1831         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1832         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1833
1834         err = register_netdev(dev);
1835         if (err) {
1836                 pr_debug("virtio_net: registering device failed\n");
1837                 goto free_vqs;
1838         }
1839
1840         virtio_device_ready(vdev);
1841
1842         /* Last of all, set up some receive buffers. */
1843         for (i = 0; i < vi->curr_queue_pairs; i++) {
1844                 try_fill_recv(vi, &vi->rq[i], GFP_KERNEL);
1845
1846                 /* If we didn't even get one input buffer, we're useless. */
1847                 if (vi->rq[i].vq->num_free ==
1848                     virtqueue_get_vring_size(vi->rq[i].vq)) {
1849                         free_unused_bufs(vi);
1850                         err = -ENOMEM;
1851                         goto free_recv_bufs;
1852                 }
1853         }
1854
1855         vi->nb.notifier_call = &virtnet_cpu_callback;
1856         err = register_hotcpu_notifier(&vi->nb);
1857         if (err) {
1858                 pr_debug("virtio_net: registering cpu notifier failed\n");
1859                 goto free_recv_bufs;
1860         }
1861
1862         /* Assume link up if device can't report link status,
1863            otherwise get link status from config. */
1864         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1865                 netif_carrier_off(dev);
1866                 schedule_work(&vi->config_work);
1867         } else {
1868                 vi->status = VIRTIO_NET_S_LINK_UP;
1869                 netif_carrier_on(dev);
1870         }
1871
1872         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1873                  dev->name, max_queue_pairs);
1874
1875         return 0;
1876
1877 free_recv_bufs:
1878         vi->vdev->config->reset(vdev);
1879
1880         free_receive_bufs(vi);
1881         unregister_netdev(dev);
1882 free_vqs:
1883         cancel_delayed_work_sync(&vi->refill);
1884         free_receive_page_frags(vi);
1885         virtnet_del_vqs(vi);
1886 free_stats:
1887         free_percpu(vi->stats);
1888 free:
1889         free_netdev(dev);
1890         return err;
1891 }
1892
1893 static void remove_vq_common(struct virtnet_info *vi)
1894 {
1895         vi->vdev->config->reset(vi->vdev);
1896
1897         /* Free unused buffers in both send and recv, if any. */
1898         free_unused_bufs(vi);
1899
1900         free_receive_bufs(vi);
1901
1902         free_receive_page_frags(vi);
1903
1904         virtnet_del_vqs(vi);
1905 }
1906
1907 static void virtnet_remove(struct virtio_device *vdev)
1908 {
1909         struct virtnet_info *vi = vdev->priv;
1910
1911         unregister_hotcpu_notifier(&vi->nb);
1912
1913         /* Make sure no work handler is accessing the device. */
1914         flush_work(&vi->config_work);
1915
1916         unregister_netdev(vi->dev);
1917
1918         remove_vq_common(vi);
1919
1920         free_percpu(vi->stats);
1921         free_netdev(vi->dev);
1922 }
1923
1924 #ifdef CONFIG_PM_SLEEP
1925 static int virtnet_freeze(struct virtio_device *vdev)
1926 {
1927         struct virtnet_info *vi = vdev->priv;
1928         int i;
1929
1930         unregister_hotcpu_notifier(&vi->nb);
1931
1932         /* Make sure no work handler is accessing the device */
1933         flush_work(&vi->config_work);
1934
1935         netif_device_detach(vi->dev);
1936         cancel_delayed_work_sync(&vi->refill);
1937
1938         if (netif_running(vi->dev)) {
1939                 for (i = 0; i < vi->max_queue_pairs; i++) {
1940                         napi_disable(&vi->rq[i].napi);
1941                         napi_hash_del(&vi->rq[i].napi);
1942                         netif_napi_del(&vi->rq[i].napi);
1943                 }
1944         }
1945
1946         remove_vq_common(vi);
1947
1948         return 0;
1949 }
1950
1951 static int virtnet_restore(struct virtio_device *vdev)
1952 {
1953         struct virtnet_info *vi = vdev->priv;
1954         int err, i;
1955
1956         err = init_vqs(vi);
1957         if (err)
1958                 return err;
1959
1960         virtio_device_ready(vdev);
1961
1962         if (netif_running(vi->dev)) {
1963                 for (i = 0; i < vi->curr_queue_pairs; i++)
1964                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1965                                 schedule_delayed_work(&vi->refill, 0);
1966
1967                 for (i = 0; i < vi->max_queue_pairs; i++)
1968                         virtnet_napi_enable(&vi->rq[i]);
1969         }
1970
1971         netif_device_attach(vi->dev);
1972
1973         rtnl_lock();
1974         virtnet_set_queues(vi, vi->curr_queue_pairs);
1975         rtnl_unlock();
1976
1977         err = register_hotcpu_notifier(&vi->nb);
1978         if (err)
1979                 return err;
1980
1981         return 0;
1982 }
1983 #endif
1984
1985 static struct virtio_device_id id_table[] = {
1986         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1987         { 0 },
1988 };
1989
1990 static unsigned int features[] = {
1991         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1992         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1993         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1994         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1995         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1996         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1997         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1998         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1999         VIRTIO_NET_F_CTRL_MAC_ADDR,
2000         VIRTIO_F_ANY_LAYOUT,
2001 };
2002
2003 static struct virtio_driver virtio_net_driver = {
2004         .feature_table = features,
2005         .feature_table_size = ARRAY_SIZE(features),
2006         .driver.name =  KBUILD_MODNAME,
2007         .driver.owner = THIS_MODULE,
2008         .id_table =     id_table,
2009         .probe =        virtnet_probe,
2010         .remove =       virtnet_remove,
2011         .config_changed = virtnet_config_changed,
2012 #ifdef CONFIG_PM_SLEEP
2013         .freeze =       virtnet_freeze,
2014         .restore =      virtnet_restore,
2015 #endif
2016 };
2017
2018 module_virtio_driver(virtio_net_driver);
2019
2020 MODULE_DEVICE_TABLE(virtio, id_table);
2021 MODULE_DESCRIPTION("Virtio network driver");
2022 MODULE_LICENSE("GPL");