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         /* timestamp packet in software */
 922         skb_tx_timestamp(skb);
 923
 924         /* Try to transmit */
 925         err = xmit_skb(sq, skb);
 926
 927         /* This should not happen! */
 928         if (unlikely(err)) {
 929                 dev->stats.tx_fifo_errors++;
 930                 if (net_ratelimit())
 931                         dev_warn(&dev->dev,
 932                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 933                 dev->stats.tx_dropped++;
 934                 dev_kfree_skb_any(skb);
 935                 return NETDEV_TX_OK;
 936         }
 937
 938         /* Don't wait up for transmitted skbs to be freed. */
 939         skb_orphan(skb);
 940         nf_reset(skb);
 941
 942         /* Apparently nice girls don't return TX_BUSY; stop the queue
 943          * before it gets out of hand.  Naturally, this wastes entries. */
 944         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 945                 netif_stop_subqueue(dev, qnum);
 946                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 947                         /* More just got used, free them then recheck. */
 948                         free_old_xmit_skbs(sq);
 949                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 950                                 netif_start_subqueue(dev, qnum);
 951                                 virtqueue_disable_cb(sq->vq);
 952                         }
 953                 }
 954         }
 955
 956         if (kick || netif_xmit_stopped(txq))
 957                 virtqueue_kick(sq->vq);
 958
 959         return NETDEV_TX_OK;
 960 }
 961
 962 /*
 963  * Send command via the control virtqueue and check status.  Commands
 964  * supported by the hypervisor, as indicated by feature bits, should
 965  * never fail unless improperly formatted.
 966  */
 967 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 968                                  struct scatterlist *out)
 969 {
 970         struct scatterlist *sgs[4], hdr, stat;
 971         struct virtio_net_ctrl_hdr ctrl;
 972         virtio_net_ctrl_ack status = ~0;
 973         unsigned out_num = 0, tmp;
 974
 975         /* Caller should know better */
 976         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 977
 978         ctrl.class = class;
 979         ctrl.cmd = cmd;
 980         /* Add header */
 981         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 982         sgs[out_num++] = &hdr;
 983
 984         if (out)
 985                 sgs[out_num++] = out;
 986
 987         /* Add return status. */
 988         sg_init_one(&stat, &status, sizeof(status));
 989         sgs[out_num] = &stat;
 990
 991         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
 992         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
 993
 994         if (unlikely(!virtqueue_kick(vi->cvq)))
 995                 return status == VIRTIO_NET_OK;
 996
 997         /* Spin for a response, the kick causes an ioport write, trapping
 998          * into the hypervisor, so the request should be handled immediately.
 999          */
1000         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1001                !virtqueue_is_broken(vi->cvq))
1002                 cpu_relax();
1003
1004         return status == VIRTIO_NET_OK;
1005 }
1006
1007 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1008 {
1009         struct virtnet_info *vi = netdev_priv(dev);
1010         struct virtio_device *vdev = vi->vdev;
1011         int ret;
1012         struct sockaddr *addr = p;
1013         struct scatterlist sg;
1014
1015         ret = eth_prepare_mac_addr_change(dev, p);
1016         if (ret)
1017                 return ret;
1018
1019         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1020                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1021                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1022                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1023                         dev_warn(&vdev->dev,
1024                                  "Failed to set mac address by vq command.\n");
1025                         return -EINVAL;
1026                 }
1027         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1028                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1029                 unsigned int i;
1030
1031                 /* Naturally, this has an atomicity problem. */
1032                 for (i = 0; i < dev->addr_len; i++)
1033                         virtio_cwrite8(vdev,
1034                                        offsetof(struct virtio_net_config, mac) +
1035                                        i, addr->sa_data[i]);
1036         }
1037
1038         eth_commit_mac_addr_change(dev, p);
1039
1040         return 0;
1041 }
1042
1043 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1044                                                struct rtnl_link_stats64 *tot)
1045 {
1046         struct virtnet_info *vi = netdev_priv(dev);
1047         int cpu;
1048         unsigned int start;
1049
1050         for_each_possible_cpu(cpu) {
1051                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1052                 u64 tpackets, tbytes, rpackets, rbytes;
1053
1054                 do {
1055                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1056                         tpackets = stats->tx_packets;
1057                         tbytes   = stats->tx_bytes;
1058                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1059
1060                 do {
1061                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1062                         rpackets = stats->rx_packets;
1063                         rbytes   = stats->rx_bytes;
1064                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1065
1066                 tot->rx_packets += rpackets;
1067                 tot->tx_packets += tpackets;
1068                 tot->rx_bytes   += rbytes;
1069                 tot->tx_bytes   += tbytes;
1070         }
1071
1072         tot->tx_dropped = dev->stats.tx_dropped;
1073         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1074         tot->rx_dropped = dev->stats.rx_dropped;
1075         tot->rx_length_errors = dev->stats.rx_length_errors;
1076         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1077
1078         return tot;
1079 }
1080
1081 #ifdef CONFIG_NET_POLL_CONTROLLER
1082 static void virtnet_netpoll(struct net_device *dev)
1083 {
1084         struct virtnet_info *vi = netdev_priv(dev);
1085         int i;
1086
1087         for (i = 0; i < vi->curr_queue_pairs; i++)
1088                 napi_schedule(&vi->rq[i].napi);
1089 }
1090 #endif
1091
1092 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1093 {
1094         rtnl_lock();
1095         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1096                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1097                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1098         rtnl_unlock();
1099 }
1100
1101 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1102 {
1103         struct scatterlist sg;
1104         struct virtio_net_ctrl_mq s;
1105         struct net_device *dev = vi->dev;
1106
1107         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1108                 return 0;
1109
1110         s.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1111         sg_init_one(&sg, &s, sizeof(s));
1112
1113         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1114                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1115                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1116                          queue_pairs);
1117                 return -EINVAL;
1118         } else {
1119                 vi->curr_queue_pairs = queue_pairs;
1120                 /* virtnet_open() will refill when device is going to up. */
1121                 if (dev->flags & IFF_UP)
1122                         schedule_delayed_work(&vi->refill, 0);
1123         }
1124
1125         return 0;
1126 }
1127
1128 static int virtnet_close(struct net_device *dev)
1129 {
1130         struct virtnet_info *vi = netdev_priv(dev);
1131         int i;
1132
1133         /* Make sure refill_work doesn't re-enable napi! */
1134         cancel_delayed_work_sync(&vi->refill);
1135
1136         for (i = 0; i < vi->max_queue_pairs; i++)
1137                 napi_disable(&vi->rq[i].napi);
1138
1139         return 0;
1140 }
1141
1142 static void virtnet_set_rx_mode(struct net_device *dev)
1143 {
1144         struct virtnet_info *vi = netdev_priv(dev);
1145         struct scatterlist sg[2];
1146         u8 promisc, allmulti;
1147         struct virtio_net_ctrl_mac *mac_data;
1148         struct netdev_hw_addr *ha;
1149         int uc_count;
1150         int mc_count;
1151         void *buf;
1152         int i;
1153
1154         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1155         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1156                 return;
1157
1158         promisc = ((dev->flags & IFF_PROMISC) != 0);
1159         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1160
1161         sg_init_one(sg, &promisc, sizeof(promisc));
1162
1163         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1164                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1165                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1166                          promisc ? "en" : "dis");
1167
1168         sg_init_one(sg, &allmulti, sizeof(allmulti));
1169
1170         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1171                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1172                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1173                          allmulti ? "en" : "dis");
1174
1175         uc_count = netdev_uc_count(dev);
1176         mc_count = netdev_mc_count(dev);
1177         /* MAC filter - use one buffer for both lists */
1178         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1179                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1180         mac_data = buf;
1181         if (!buf)
1182                 return;
1183
1184         sg_init_table(sg, 2);
1185
1186         /* Store the unicast list and count in the front of the buffer */
1187         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1188         i = 0;
1189         netdev_for_each_uc_addr(ha, dev)
1190                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1191
1192         sg_set_buf(&sg[0], mac_data,
1193                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1194
1195         /* multicast list and count fill the end */
1196         mac_data = (void *)&mac_data->macs[uc_count][0];
1197
1198         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1199         i = 0;
1200         netdev_for_each_mc_addr(ha, dev)
1201                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1202
1203         sg_set_buf(&sg[1], mac_data,
1204                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1205
1206         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1207                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1208                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1209
1210         kfree(buf);
1211 }
1212
1213 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1214                                    __be16 proto, u16 vid)
1215 {
1216         struct virtnet_info *vi = netdev_priv(dev);
1217         struct scatterlist sg;
1218
1219         sg_init_one(&sg, &vid, sizeof(vid));
1220
1221         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1222                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1223                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1224         return 0;
1225 }
1226
1227 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1228                                     __be16 proto, u16 vid)
1229 {
1230         struct virtnet_info *vi = netdev_priv(dev);
1231         struct scatterlist sg;
1232
1233         sg_init_one(&sg, &vid, sizeof(vid));
1234
1235         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1236                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1237                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1238         return 0;
1239 }
1240
1241 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1242 {
1243         int i;
1244
1245         if (vi->affinity_hint_set) {
1246                 for (i = 0; i < vi->max_queue_pairs; i++) {
1247                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1248                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1249                 }
1250
1251                 vi->affinity_hint_set = false;
1252         }
1253 }
1254
1255 static void virtnet_set_affinity(struct virtnet_info *vi)
1256 {
1257         int i;
1258         int cpu;
1259
1260         /* In multiqueue mode, when the number of cpu is equal to the number of
1261          * queue pairs, we let the queue pairs to be private to one cpu by
1262          * setting the affinity hint to eliminate the contention.
1263          */
1264         if (vi->curr_queue_pairs == 1 ||
1265             vi->max_queue_pairs != num_online_cpus()) {
1266                 virtnet_clean_affinity(vi, -1);
1267                 return;
1268         }
1269
1270         i = 0;
1271         for_each_online_cpu(cpu) {
1272                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1273                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1274                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1275                 i++;
1276         }
1277
1278         vi->affinity_hint_set = true;
1279 }
1280
1281 static int virtnet_cpu_callback(struct notifier_block *nfb,
1282                                 unsigned long action, void *hcpu)
1283 {
1284         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1285
1286         switch(action & ~CPU_TASKS_FROZEN) {
1287         case CPU_ONLINE:
1288         case CPU_DOWN_FAILED:
1289         case CPU_DEAD:
1290                 virtnet_set_affinity(vi);
1291                 break;
1292         case CPU_DOWN_PREPARE:
1293                 virtnet_clean_affinity(vi, (long)hcpu);
1294                 break;
1295         default:
1296                 break;
1297         }
1298
1299         return NOTIFY_OK;
1300 }
1301
1302 static void virtnet_get_ringparam(struct net_device *dev,
1303                                 struct ethtool_ringparam *ring)
1304 {
1305         struct virtnet_info *vi = netdev_priv(dev);
1306
1307         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1308         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1309         ring->rx_pending = ring->rx_max_pending;
1310         ring->tx_pending = ring->tx_max_pending;
1311 }
1312
1313
1314 static void virtnet_get_drvinfo(struct net_device *dev,
1315                                 struct ethtool_drvinfo *info)
1316 {
1317         struct virtnet_info *vi = netdev_priv(dev);
1318         struct virtio_device *vdev = vi->vdev;
1319
1320         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1321         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1322         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1323
1324 }
1325
1326 /* TODO: Eliminate OOO packets during switching */
1327 static int virtnet_set_channels(struct net_device *dev,
1328                                 struct ethtool_channels *channels)
1329 {
1330         struct virtnet_info *vi = netdev_priv(dev);
1331         u16 queue_pairs = channels->combined_count;
1332         int err;
1333
1334         /* We don't support separate rx/tx channels.
1335          * We don't allow setting 'other' channels.
1336          */
1337         if (channels->rx_count || channels->tx_count || channels->other_count)
1338                 return -EINVAL;
1339
1340         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1341                 return -EINVAL;
1342
1343         get_online_cpus();
1344         err = virtnet_set_queues(vi, queue_pairs);
1345         if (!err) {
1346                 netif_set_real_num_tx_queues(dev, queue_pairs);
1347                 netif_set_real_num_rx_queues(dev, queue_pairs);
1348
1349                 virtnet_set_affinity(vi);
1350         }
1351         put_online_cpus();
1352
1353         return err;
1354 }
1355
1356 static void virtnet_get_channels(struct net_device *dev,
1357                                  struct ethtool_channels *channels)
1358 {
1359         struct virtnet_info *vi = netdev_priv(dev);
1360
1361         channels->combined_count = vi->curr_queue_pairs;
1362         channels->max_combined = vi->max_queue_pairs;
1363         channels->max_other = 0;
1364         channels->rx_count = 0;
1365         channels->tx_count = 0;
1366         channels->other_count = 0;
1367 }
1368
1369 static const struct ethtool_ops virtnet_ethtool_ops = {
1370         .get_drvinfo = virtnet_get_drvinfo,
1371         .get_link = ethtool_op_get_link,
1372         .get_ringparam = virtnet_get_ringparam,
1373         .set_channels = virtnet_set_channels,
1374         .get_channels = virtnet_get_channels,
1375         .get_ts_info = ethtool_op_get_ts_info,
1376 };
1377
1378 #define MIN_MTU 68
1379 #define MAX_MTU 65535
1380
1381 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1382 {
1383         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1384                 return -EINVAL;
1385         dev->mtu = new_mtu;
1386         return 0;
1387 }
1388
1389 static const struct net_device_ops virtnet_netdev = {
1390         .ndo_open            = virtnet_open,
1391         .ndo_stop            = virtnet_close,
1392         .ndo_start_xmit      = start_xmit,
1393         .ndo_validate_addr   = eth_validate_addr,
1394         .ndo_set_mac_address = virtnet_set_mac_address,
1395         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1396         .ndo_change_mtu      = virtnet_change_mtu,
1397         .ndo_get_stats64     = virtnet_stats,
1398         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1399         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1400 #ifdef CONFIG_NET_POLL_CONTROLLER
1401         .ndo_poll_controller = virtnet_netpoll,
1402 #endif
1403 #ifdef CONFIG_NET_RX_BUSY_POLL
1404         .ndo_busy_poll          = virtnet_busy_poll,
1405 #endif
1406 };
1407
1408 static void virtnet_config_changed_work(struct work_struct *work)
1409 {
1410         struct virtnet_info *vi =
1411                 container_of(work, struct virtnet_info, config_work);
1412         u16 v;
1413
1414         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1415                                  struct virtio_net_config, status, &v) < 0)
1416                 return;
1417
1418         if (v & VIRTIO_NET_S_ANNOUNCE) {
1419                 netdev_notify_peers(vi->dev);
1420                 virtnet_ack_link_announce(vi);
1421         }
1422
1423         /* Ignore unknown (future) status bits */
1424         v &= VIRTIO_NET_S_LINK_UP;
1425
1426         if (vi->status == v)
1427                 return;
1428
1429         vi->status = v;
1430
1431         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1432                 netif_carrier_on(vi->dev);
1433                 netif_tx_wake_all_queues(vi->dev);
1434         } else {
1435                 netif_carrier_off(vi->dev);
1436                 netif_tx_stop_all_queues(vi->dev);
1437         }
1438 }
1439
1440 static void virtnet_config_changed(struct virtio_device *vdev)
1441 {
1442         struct virtnet_info *vi = vdev->priv;
1443
1444         schedule_work(&vi->config_work);
1445 }
1446
1447 static void virtnet_free_queues(struct virtnet_info *vi)
1448 {
1449         int i;
1450
1451         for (i = 0; i < vi->max_queue_pairs; i++) {
1452                 napi_hash_del(&vi->rq[i].napi);
1453                 netif_napi_del(&vi->rq[i].napi);
1454         }
1455
1456         kfree(vi->rq);
1457         kfree(vi->sq);
1458 }
1459
1460 static void free_receive_bufs(struct virtnet_info *vi)
1461 {
1462         int i;
1463
1464         for (i = 0; i < vi->max_queue_pairs; i++) {
1465                 while (vi->rq[i].pages)
1466                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1467         }
1468 }
1469
1470 static void free_receive_page_frags(struct virtnet_info *vi)
1471 {
1472         int i;
1473         for (i = 0; i < vi->max_queue_pairs; i++)
1474                 if (vi->rq[i].alloc_frag.page)
1475                         put_page(vi->rq[i].alloc_frag.page);
1476 }
1477
1478 static void free_unused_bufs(struct virtnet_info *vi)
1479 {
1480         void *buf;
1481         int i;
1482
1483         for (i = 0; i < vi->max_queue_pairs; i++) {
1484                 struct virtqueue *vq = vi->sq[i].vq;
1485                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1486                         dev_kfree_skb(buf);
1487         }
1488
1489         for (i = 0; i < vi->max_queue_pairs; i++) {
1490                 struct virtqueue *vq = vi->rq[i].vq;
1491
1492                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1493                         if (vi->mergeable_rx_bufs) {
1494                                 unsigned long ctx = (unsigned long)buf;
1495                                 void *base = mergeable_ctx_to_buf_address(ctx);
1496                                 put_page(virt_to_head_page(base));
1497                         } else if (vi->big_packets) {
1498                                 give_pages(&vi->rq[i], buf);
1499                         } else {
1500                                 dev_kfree_skb(buf);
1501                         }
1502                 }
1503         }
1504 }
1505
1506 static void virtnet_del_vqs(struct virtnet_info *vi)
1507 {
1508         struct virtio_device *vdev = vi->vdev;
1509
1510         virtnet_clean_affinity(vi, -1);
1511
1512         vdev->config->del_vqs(vdev);
1513
1514         virtnet_free_queues(vi);
1515 }
1516
1517 static int virtnet_find_vqs(struct virtnet_info *vi)
1518 {
1519         vq_callback_t **callbacks;
1520         struct virtqueue **vqs;
1521         int ret = -ENOMEM;
1522         int i, total_vqs;
1523         const char **names;
1524
1525         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1526          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1527          * possible control vq.
1528          */
1529         total_vqs = vi->max_queue_pairs * 2 +
1530                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1531
1532         /* Allocate space for find_vqs parameters */
1533         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1534         if (!vqs)
1535                 goto err_vq;
1536         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1537         if (!callbacks)
1538                 goto err_callback;
1539         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1540         if (!names)
1541                 goto err_names;
1542
1543         /* Parameters for control virtqueue, if any */
1544         if (vi->has_cvq) {
1545                 callbacks[total_vqs - 1] = NULL;
1546                 names[total_vqs - 1] = "control";
1547         }
1548
1549         /* Allocate/initialize parameters for send/receive virtqueues */
1550         for (i = 0; i < vi->max_queue_pairs; i++) {
1551                 callbacks[rxq2vq(i)] = skb_recv_done;
1552                 callbacks[txq2vq(i)] = skb_xmit_done;
1553                 sprintf(vi->rq[i].name, "input.%d", i);
1554                 sprintf(vi->sq[i].name, "output.%d", i);
1555                 names[rxq2vq(i)] = vi->rq[i].name;
1556                 names[txq2vq(i)] = vi->sq[i].name;
1557         }
1558
1559         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1560                                          names);
1561         if (ret)
1562                 goto err_find;
1563
1564         if (vi->has_cvq) {
1565                 vi->cvq = vqs[total_vqs - 1];
1566                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1567                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1568         }
1569
1570         for (i = 0; i < vi->max_queue_pairs; i++) {
1571                 vi->rq[i].vq = vqs[rxq2vq(i)];
1572                 vi->sq[i].vq = vqs[txq2vq(i)];
1573         }
1574
1575         kfree(names);
1576         kfree(callbacks);
1577         kfree(vqs);
1578
1579         return 0;
1580
1581 err_find:
1582         kfree(names);
1583 err_names:
1584         kfree(callbacks);
1585 err_callback:
1586         kfree(vqs);
1587 err_vq:
1588         return ret;
1589 }
1590
1591 static int virtnet_alloc_queues(struct virtnet_info *vi)
1592 {
1593         int i;
1594
1595         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1596         if (!vi->sq)
1597                 goto err_sq;
1598         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1599         if (!vi->rq)
1600                 goto err_rq;
1601
1602         INIT_DELAYED_WORK(&vi->refill, refill_work);
1603         for (i = 0; i < vi->max_queue_pairs; i++) {
1604                 vi->rq[i].pages = NULL;
1605                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1606                                napi_weight);
1607                 napi_hash_add(&vi->rq[i].napi);
1608
1609                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1610                 ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
1611                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1612         }
1613
1614         return 0;
1615
1616 err_rq:
1617         kfree(vi->sq);
1618 err_sq:
1619         return -ENOMEM;
1620 }
1621
1622 static int init_vqs(struct virtnet_info *vi)
1623 {
1624         int ret;
1625
1626         /* Allocate send & receive queues */
1627         ret = virtnet_alloc_queues(vi);
1628         if (ret)
1629                 goto err;
1630
1631         ret = virtnet_find_vqs(vi);
1632         if (ret)
1633                 goto err_free;
1634
1635         get_online_cpus();
1636         virtnet_set_affinity(vi);
1637         put_online_cpus();
1638
1639         return 0;
1640
1641 err_free:
1642         virtnet_free_queues(vi);
1643 err:
1644         return ret;
1645 }
1646
1647 #ifdef CONFIG_SYSFS
1648 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1649                 struct rx_queue_attribute *attribute, char *buf)
1650 {
1651         struct virtnet_info *vi = netdev_priv(queue->dev);
1652         unsigned int queue_index = get_netdev_rx_queue_index(queue);
1653         struct ewma *avg;
1654
1655         BUG_ON(queue_index >= vi->max_queue_pairs);
1656         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1657         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1658 }
1659
1660 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1661         __ATTR_RO(mergeable_rx_buffer_size);
1662
1663 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1664         &mergeable_rx_buffer_size_attribute.attr,
1665         NULL
1666 };
1667
1668 static const struct attribute_group virtio_net_mrg_rx_group = {
1669         .name = "virtio_net",
1670         .attrs = virtio_net_mrg_rx_attrs
1671 };
1672 #endif
1673
1674 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1675                                     unsigned int fbit,
1676                                     const char *fname, const char *dname)
1677 {
1678         if (!virtio_has_feature(vdev, fbit))
1679                 return false;
1680
1681         dev_err(&vdev->dev, "device advertises feature %s but not %s",
1682                 fname, dname);
1683
1684         return true;
1685 }
1686
1687 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
1688         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1689
1690 static bool virtnet_validate_features(struct virtio_device *vdev)
1691 {
1692         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1693             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1694                              "VIRTIO_NET_F_CTRL_VQ") ||
1695              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1696                              "VIRTIO_NET_F_CTRL_VQ") ||
1697              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1698                              "VIRTIO_NET_F_CTRL_VQ") ||
1699              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1700              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1701                              "VIRTIO_NET_F_CTRL_VQ"))) {
1702                 return false;
1703         }
1704
1705         return true;
1706 }
1707
1708 static int virtnet_probe(struct virtio_device *vdev)
1709 {
1710         int i, err;
1711         struct net_device *dev;
1712         struct virtnet_info *vi;
1713         u16 max_queue_pairs;
1714
1715         if (!vdev->config->get) {
1716                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
1717                         __func__);
1718                 return -EINVAL;
1719         }
1720
1721         if (!virtnet_validate_features(vdev))
1722                 return -EINVAL;
1723
1724         /* Find if host supports multiqueue virtio_net device */
1725         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1726                                    struct virtio_net_config,
1727                                    max_virtqueue_pairs, &max_queue_pairs);
1728
1729         /* We need at least 2 queue's */
1730         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1731             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1732             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1733                 max_queue_pairs = 1;
1734
1735         /* Allocate ourselves a network device with room for our info */
1736         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1737         if (!dev)
1738                 return -ENOMEM;
1739
1740         /* Set up network device as normal. */
1741         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1742         dev->netdev_ops = &virtnet_netdev;
1743         dev->features = NETIF_F_HIGHDMA;
1744
1745         dev->ethtool_ops = &virtnet_ethtool_ops;
1746         SET_NETDEV_DEV(dev, &vdev->dev);
1747
1748         /* Do we support "hardware" checksums? */
1749         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1750                 /* This opens up the world of extra features. */
1751                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1752                 if (csum)
1753                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1754
1755                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1756                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1757                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1758                 }
1759                 /* Individual feature bits: what can host handle? */
1760                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1761                         dev->hw_features |= NETIF_F_TSO;
1762                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1763                         dev->hw_features |= NETIF_F_TSO6;
1764                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1765                         dev->hw_features |= NETIF_F_TSO_ECN;
1766                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1767                         dev->hw_features |= NETIF_F_UFO;
1768
1769                 dev->features |= NETIF_F_GSO_ROBUST;
1770
1771                 if (gso)
1772                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1773                 /* (!csum && gso) case will be fixed by register_netdev() */
1774         }
1775         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1776                 dev->features |= NETIF_F_RXCSUM;
1777
1778         dev->vlan_features = dev->features;
1779
1780         /* Configuration may specify what MAC to use.  Otherwise random. */
1781         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1782                 virtio_cread_bytes(vdev,
1783                                    offsetof(struct virtio_net_config, mac),
1784                                    dev->dev_addr, dev->addr_len);
1785         else
1786                 eth_hw_addr_random(dev);
1787
1788         /* Set up our device-specific information */
1789         vi = netdev_priv(dev);
1790         vi->dev = dev;
1791         vi->vdev = vdev;
1792         vdev->priv = vi;
1793         vi->stats = alloc_percpu(struct virtnet_stats);
1794         err = -ENOMEM;
1795         if (vi->stats == NULL)
1796                 goto free;
1797
1798         for_each_possible_cpu(i) {
1799                 struct virtnet_stats *virtnet_stats;
1800                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1801                 u64_stats_init(&virtnet_stats->tx_syncp);
1802                 u64_stats_init(&virtnet_stats->rx_syncp);
1803         }
1804
1805         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1806
1807         /* If we can receive ANY GSO packets, we must allocate large ones. */
1808         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1809             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1810             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1811             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1812                 vi->big_packets = true;
1813
1814         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1815                 vi->mergeable_rx_bufs = true;
1816
1817         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
1818             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1819                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1820         else
1821                 vi->hdr_len = sizeof(struct virtio_net_hdr);
1822
1823         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1824                 vi->any_header_sg = true;
1825
1826         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1827                 vi->has_cvq = true;
1828
1829         if (vi->any_header_sg)
1830                 dev->needed_headroom = vi->hdr_len;
1831
1832         /* Use single tx/rx queue pair as default */
1833         vi->curr_queue_pairs = 1;
1834         vi->max_queue_pairs = max_queue_pairs;
1835
1836         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1837         err = init_vqs(vi);
1838         if (err)
1839                 goto free_stats;
1840
1841 #ifdef CONFIG_SYSFS
1842         if (vi->mergeable_rx_bufs)
1843                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1844 #endif
1845         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1846         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1847
1848         err = register_netdev(dev);
1849         if (err) {
1850                 pr_debug("virtio_net: registering device failed\n");
1851                 goto free_vqs;
1852         }
1853
1854         virtio_device_ready(vdev);
1855
1856         /* Last of all, set up some receive buffers. */
1857         for (i = 0; i < vi->curr_queue_pairs; i++) {
1858                 try_fill_recv(vi, &vi->rq[i], GFP_KERNEL);
1859
1860                 /* If we didn't even get one input buffer, we're useless. */
1861                 if (vi->rq[i].vq->num_free ==
1862                     virtqueue_get_vring_size(vi->rq[i].vq)) {
1863                         free_unused_bufs(vi);
1864                         err = -ENOMEM;
1865                         goto free_recv_bufs;
1866                 }
1867         }
1868
1869         vi->nb.notifier_call = &virtnet_cpu_callback;
1870         err = register_hotcpu_notifier(&vi->nb);
1871         if (err) {
1872                 pr_debug("virtio_net: registering cpu notifier failed\n");
1873                 goto free_recv_bufs;
1874         }
1875
1876         /* Assume link up if device can't report link status,
1877            otherwise get link status from config. */
1878         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1879                 netif_carrier_off(dev);
1880                 schedule_work(&vi->config_work);
1881         } else {
1882                 vi->status = VIRTIO_NET_S_LINK_UP;
1883                 netif_carrier_on(dev);
1884         }
1885
1886         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1887                  dev->name, max_queue_pairs);
1888
1889         return 0;
1890
1891 free_recv_bufs:
1892         vi->vdev->config->reset(vdev);
1893
1894         free_receive_bufs(vi);
1895         unregister_netdev(dev);
1896 free_vqs:
1897         cancel_delayed_work_sync(&vi->refill);
1898         free_receive_page_frags(vi);
1899         virtnet_del_vqs(vi);
1900 free_stats:
1901         free_percpu(vi->stats);
1902 free:
1903         free_netdev(dev);
1904         return err;
1905 }
1906
1907 static void remove_vq_common(struct virtnet_info *vi)
1908 {
1909         vi->vdev->config->reset(vi->vdev);
1910
1911         /* Free unused buffers in both send and recv, if any. */
1912         free_unused_bufs(vi);
1913
1914         free_receive_bufs(vi);
1915
1916         free_receive_page_frags(vi);
1917
1918         virtnet_del_vqs(vi);
1919 }
1920
1921 static void virtnet_remove(struct virtio_device *vdev)
1922 {
1923         struct virtnet_info *vi = vdev->priv;
1924
1925         unregister_hotcpu_notifier(&vi->nb);
1926
1927         /* Make sure no work handler is accessing the device. */
1928         flush_work(&vi->config_work);
1929
1930         unregister_netdev(vi->dev);
1931
1932         remove_vq_common(vi);
1933
1934         free_percpu(vi->stats);
1935         free_netdev(vi->dev);
1936 }
1937
1938 #ifdef CONFIG_PM_SLEEP
1939 static int virtnet_freeze(struct virtio_device *vdev)
1940 {
1941         struct virtnet_info *vi = vdev->priv;
1942         int i;
1943
1944         unregister_hotcpu_notifier(&vi->nb);
1945
1946         /* Make sure no work handler is accessing the device */
1947         flush_work(&vi->config_work);
1948
1949         netif_device_detach(vi->dev);
1950         cancel_delayed_work_sync(&vi->refill);
1951
1952         if (netif_running(vi->dev)) {
1953                 for (i = 0; i < vi->max_queue_pairs; i++)
1954                         napi_disable(&vi->rq[i].napi);
1955         }
1956
1957         remove_vq_common(vi);
1958
1959         return 0;
1960 }
1961
1962 static int virtnet_restore(struct virtio_device *vdev)
1963 {
1964         struct virtnet_info *vi = vdev->priv;
1965         int err, i;
1966
1967         err = init_vqs(vi);
1968         if (err)
1969                 return err;
1970
1971         virtio_device_ready(vdev);
1972
1973         if (netif_running(vi->dev)) {
1974                 for (i = 0; i < vi->curr_queue_pairs; i++)
1975                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1976                                 schedule_delayed_work(&vi->refill, 0);
1977
1978                 for (i = 0; i < vi->max_queue_pairs; i++)
1979                         virtnet_napi_enable(&vi->rq[i]);
1980         }
1981
1982         netif_device_attach(vi->dev);
1983
1984         rtnl_lock();
1985         virtnet_set_queues(vi, vi->curr_queue_pairs);
1986         rtnl_unlock();
1987
1988         err = register_hotcpu_notifier(&vi->nb);
1989         if (err)
1990                 return err;
1991
1992         return 0;
1993 }
1994 #endif
1995
1996 static struct virtio_device_id id_table[] = {
1997         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1998         { 0 },
1999 };
2000
2001 static unsigned int features[] = {
2002         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
2003         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
2004         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
2005         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
2006         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
2007         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
2008         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
2009         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
2010         VIRTIO_NET_F_CTRL_MAC_ADDR,
2011         VIRTIO_F_ANY_LAYOUT,
2012 };
2013
2014 static struct virtio_driver virtio_net_driver = {
2015         .feature_table = features,
2016         .feature_table_size = ARRAY_SIZE(features),
2017         .driver.name =  KBUILD_MODNAME,
2018         .driver.owner = THIS_MODULE,
2019         .id_table =     id_table,
2020         .probe =        virtnet_probe,
2021         .remove =       virtnet_remove,
2022         .config_changed = virtnet_config_changed,
2023 #ifdef CONFIG_PM_SLEEP
2024         .freeze =       virtnet_freeze,
2025         .restore =      virtnet_restore,
2026 #endif
2027 };
2028
2029 module_virtio_driver(virtio_net_driver);
2030
2031 MODULE_DEVICE_TABLE(virtio, id_table);
2032 MODULE_DESCRIPTION("Virtio network driver");
2033 MODULE_LICENSE("GPL");