kernel/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 /* RX packet size EWMA. The average packet size is used to determine the packet
  44  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  45  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  46  * term, transient changes in packet size.
  47  */
  48 DECLARE_EWMA(pkt_len, 1, 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.$index */
  74         char name[40];
  75 };
  76
  77 /* Internal representation of a receive virtqueue */
  78 struct receive_queue {
  79         /* Virtqueue associated with this receive_queue */
  80         struct virtqueue *vq;
  81
  82         struct napi_struct napi;
  83
  84         /* Chain pages by the private ptr. */
  85         struct page *pages;
  86
  87         /* Average packet length for mergeable receive buffers. */
  88         struct ewma_pkt_len mrg_avg_pkt_len;
  89
  90         /* Page frag for packet buffer allocation. */
  91         struct page_frag alloc_frag;
  92
  93         /* RX: fragments + linear part + virtio header */
  94         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  95
  96         /* Name of this receive queue: input.$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         /* Control VQ buffers: protected by the rtnl lock */
 145         struct virtio_net_ctrl_hdr ctrl_hdr;
 146         virtio_net_ctrl_ack ctrl_status;
 147         u8 ctrl_promisc;
 148         u8 ctrl_allmulti;
 149 };
 150
 151 struct padded_vnet_hdr {
 152         struct virtio_net_hdr_mrg_rxbuf hdr;
 153         /*
 154          * hdr is in a separate sg buffer, and data sg buffer shares same page
 155          * with this header sg. This padding makes next sg 16 byte aligned
 156          * after the header.
 157          */
 158         char padding[4];
 159 };
 160
 161 /* Converting between virtqueue no. and kernel tx/rx queue no.
 162  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 163  */
 164 static int vq2txq(struct virtqueue *vq)
 165 {
 166         return (vq->index - 1) / 2;
 167 }
 168
 169 static int txq2vq(int txq)
 170 {
 171         return txq * 2 + 1;
 172 }
 173
 174 static int vq2rxq(struct virtqueue *vq)
 175 {
 176         return vq->index / 2;
 177 }
 178
 179 static int rxq2vq(int rxq)
 180 {
 181         return rxq * 2;
 182 }
 183
 184 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 185 {
 186         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 187 }
 188
 189 /*
 190  * private is used to chain pages for big packets, put the whole
 191  * most recent used list in the beginning for reuse
 192  */
 193 static void give_pages(struct receive_queue *rq, struct page *page)
 194 {
 195         struct page *end;
 196
 197         /* Find end of list, sew whole thing into vi->rq.pages. */
 198         for (end = page; end->private; end = (struct page *)end->private);
 199         end->private = (unsigned long)rq->pages;
 200         rq->pages = page;
 201 }
 202
 203 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 204 {
 205         struct page *p = rq->pages;
 206
 207         if (p) {
 208                 rq->pages = (struct page *)p->private;
 209                 /* clear private here, it is used to chain pages */
 210                 p->private = 0;
 211         } else
 212                 p = alloc_page(gfp_mask);
 213         return p;
 214 }
 215
 216 static void skb_xmit_done(struct virtqueue *vq)
 217 {
 218         struct virtnet_info *vi = vq->vdev->priv;
 219
 220         /* Suppress further interrupts. */
 221         virtqueue_disable_cb(vq);
 222
 223         /* We were probably waiting for more output buffers. */
 224         netif_wake_subqueue(vi->dev, vq2txq(vq));
 225 }
 226
 227 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
 228 {
 229         unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
 230         return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
 231 }
 232
 233 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
 234 {
 235         return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
 236
 237 }
 238
 239 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
 240 {
 241         unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
 242         return (unsigned long)buf | (size - 1);
 243 }
 244
 245 /* Called from bottom half context */
 246 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 247                                    struct receive_queue *rq,
 248                                    struct page *page, unsigned int offset,
 249                                    unsigned int len, unsigned int truesize)
 250 {
 251         struct sk_buff *skb;
 252         struct virtio_net_hdr_mrg_rxbuf *hdr;
 253         unsigned int copy, hdr_len, hdr_padded_len;
 254         char *p;
 255
 256         p = page_address(page) + offset;
 257
 258         /* copy small packet so we can reuse these pages for small data */
 259         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 260         if (unlikely(!skb))
 261                 return NULL;
 262
 263         hdr = skb_vnet_hdr(skb);
 264
 265         hdr_len = vi->hdr_len;
 266         if (vi->mergeable_rx_bufs)
 267                 hdr_padded_len = sizeof *hdr;
 268         else
 269                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 270
 271         memcpy(hdr, p, hdr_len);
 272
 273         len -= hdr_len;
 274         offset += hdr_padded_len;
 275         p += hdr_padded_len;
 276
 277         copy = len;
 278         if (copy > skb_tailroom(skb))
 279                 copy = skb_tailroom(skb);
 280         memcpy(skb_put(skb, copy), p, copy);
 281
 282         len -= copy;
 283         offset += copy;
 284
 285         if (vi->mergeable_rx_bufs) {
 286                 if (len)
 287                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 288                 else
 289                         put_page(page);
 290                 return skb;
 291         }
 292
 293         /*
 294          * Verify that we can indeed put this data into a skb.
 295          * This is here to handle cases when the device erroneously
 296          * tries to receive more than is possible. This is usually
 297          * the case of a broken device.
 298          */
 299         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 300                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 301                 dev_kfree_skb(skb);
 302                 return NULL;
 303         }
 304         BUG_ON(offset >= PAGE_SIZE);
 305         while (len) {
 306                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 307                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 308                                 frag_size, truesize);
 309                 len -= frag_size;
 310                 page = (struct page *)page->private;
 311                 offset = 0;
 312         }
 313
 314         if (page)
 315                 give_pages(rq, page);
 316
 317         return skb;
 318 }
 319
 320 static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
 321 {
 322         struct sk_buff * skb = buf;
 323
 324         len -= vi->hdr_len;
 325         skb_trim(skb, len);
 326
 327         return skb;
 328 }
 329
 330 static struct sk_buff *receive_big(struct net_device *dev,
 331                                    struct virtnet_info *vi,
 332                                    struct receive_queue *rq,
 333                                    void *buf,
 334                                    unsigned int len)
 335 {
 336         struct page *page = buf;
 337         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 338
 339         if (unlikely(!skb))
 340                 goto err;
 341
 342         return skb;
 343
 344 err:
 345         dev->stats.rx_dropped++;
 346         give_pages(rq, page);
 347         return NULL;
 348 }
 349
 350 static struct sk_buff *receive_mergeable(struct net_device *dev,
 351                                          struct virtnet_info *vi,
 352                                          struct receive_queue *rq,
 353                                          unsigned long ctx,
 354                                          unsigned int len)
 355 {
 356         void *buf = mergeable_ctx_to_buf_address(ctx);
 357         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 358         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 359         struct page *page = virt_to_head_page(buf);
 360         int offset = buf - page_address(page);
 361         unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 362
 363         struct sk_buff *head_skb = page_to_skb(vi, rq, page, offset, len,
 364                                                truesize);
 365         struct sk_buff *curr_skb = head_skb;
 366
 367         if (unlikely(!curr_skb))
 368                 goto err_skb;
 369         while (--num_buf) {
 370                 int num_skb_frags;
 371
 372                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 373                 if (unlikely(!ctx)) {
 374                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 375                                  dev->name, num_buf,
 376                                  virtio16_to_cpu(vi->vdev,
 377                                                  hdr->num_buffers));
 378                         dev->stats.rx_length_errors++;
 379                         goto err_buf;
 380                 }
 381
 382                 buf = mergeable_ctx_to_buf_address(ctx);
 383                 page = virt_to_head_page(buf);
 384
 385                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 386                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 387                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 388
 389                         if (unlikely(!nskb))
 390                                 goto err_skb;
 391                         if (curr_skb == head_skb)
 392                                 skb_shinfo(curr_skb)->frag_list = nskb;
 393                         else
 394                                 curr_skb->next = nskb;
 395                         curr_skb = nskb;
 396                         head_skb->truesize += nskb->truesize;
 397                         num_skb_frags = 0;
 398                 }
 399                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 400                 if (curr_skb != head_skb) {
 401                         head_skb->data_len += len;
 402                         head_skb->len += len;
 403                         head_skb->truesize += truesize;
 404                 }
 405                 offset = buf - page_address(page);
 406                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 407                         put_page(page);
 408                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 409                                              len, truesize);
 410                 } else {
 411                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 412                                         offset, len, truesize);
 413                 }
 414         }
 415
 416         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 417         return head_skb;
 418
 419 err_skb:
 420         put_page(page);
 421         while (--num_buf) {
 422                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 423                 if (unlikely(!ctx)) {
 424                         pr_debug("%s: rx error: %d buffers missing\n",
 425                                  dev->name, num_buf);
 426                         dev->stats.rx_length_errors++;
 427                         break;
 428                 }
 429                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 430                 put_page(page);
 431         }
 432 err_buf:
 433         dev->stats.rx_dropped++;
 434         dev_kfree_skb(head_skb);
 435         return NULL;
 436 }
 437
 438 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 439                         void *buf, unsigned int len)
 440 {
 441         struct net_device *dev = vi->dev;
 442         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 443         struct sk_buff *skb;
 444         struct virtio_net_hdr_mrg_rxbuf *hdr;
 445
 446         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 447                 pr_debug("%s: short packet %i\n", dev->name, len);
 448                 dev->stats.rx_length_errors++;
 449                 if (vi->mergeable_rx_bufs) {
 450                         unsigned long ctx = (unsigned long)buf;
 451                         void *base = mergeable_ctx_to_buf_address(ctx);
 452                         put_page(virt_to_head_page(base));
 453                 } else if (vi->big_packets) {
 454                         give_pages(rq, buf);
 455                 } else {
 456                         dev_kfree_skb(buf);
 457                 }
 458                 return;
 459         }
 460
 461         if (vi->mergeable_rx_bufs)
 462                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
 463         else if (vi->big_packets)
 464                 skb = receive_big(dev, vi, rq, buf, len);
 465         else
 466                 skb = receive_small(vi, buf, len);
 467
 468         if (unlikely(!skb))
 469                 return;
 470
 471         hdr = skb_vnet_hdr(skb);
 472
 473         u64_stats_update_begin(&stats->rx_syncp);
 474         stats->rx_bytes += skb->len;
 475         stats->rx_packets++;
 476         u64_stats_update_end(&stats->rx_syncp);
 477
 478         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 479                 pr_debug("Needs csum!\n");
 480                 if (!skb_partial_csum_set(skb,
 481                           virtio16_to_cpu(vi->vdev, hdr->hdr.csum_start),
 482                           virtio16_to_cpu(vi->vdev, hdr->hdr.csum_offset)))
 483                         goto frame_err;
 484         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 485                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 486         }
 487
 488         skb->protocol = eth_type_trans(skb, dev);
 489         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 490                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 491
 492         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 493                 pr_debug("GSO!\n");
 494                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 495                 case VIRTIO_NET_HDR_GSO_TCPV4:
 496                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 497                         break;
 498                 case VIRTIO_NET_HDR_GSO_UDP:
 499                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 500                         break;
 501                 case VIRTIO_NET_HDR_GSO_TCPV6:
 502                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 503                         break;
 504                 default:
 505                         net_warn_ratelimited("%s: bad gso type %u.\n",
 506                                              dev->name, hdr->hdr.gso_type);
 507                         goto frame_err;
 508                 }
 509
 510                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 511                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 512
 513                 skb_shinfo(skb)->gso_size = virtio16_to_cpu(vi->vdev,
 514                                                             hdr->hdr.gso_size);
 515                 if (skb_shinfo(skb)->gso_size == 0) {
 516                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 517                         goto frame_err;
 518                 }
 519
 520                 /* Header must be checked, and gso_segs computed. */
 521                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 522                 skb_shinfo(skb)->gso_segs = 0;
 523         }
 524
 525         skb_mark_napi_id(skb, &rq->napi);
 526
 527         napi_gro_receive(&rq->napi, skb);
 528         return;
 529
 530 frame_err:
 531         dev->stats.rx_frame_errors++;
 532         dev_kfree_skb(skb);
 533 }
 534
 535 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 536                              gfp_t gfp)
 537 {
 538         struct sk_buff *skb;
 539         struct virtio_net_hdr_mrg_rxbuf *hdr;
 540         int err;
 541
 542         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 543         if (unlikely(!skb))
 544                 return -ENOMEM;
 545
 546         skb_put(skb, GOOD_PACKET_LEN);
 547
 548         hdr = skb_vnet_hdr(skb);
 549         sg_init_table(rq->sg, 2);
 550         sg_set_buf(rq->sg, hdr, vi->hdr_len);
 551         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 552
 553         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 554         if (err < 0)
 555                 dev_kfree_skb(skb);
 556
 557         return err;
 558 }
 559
 560 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 561                            gfp_t gfp)
 562 {
 563         struct page *first, *list = NULL;
 564         char *p;
 565         int i, err, offset;
 566
 567         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 568
 569         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 570         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 571                 first = get_a_page(rq, gfp);
 572                 if (!first) {
 573                         if (list)
 574                                 give_pages(rq, list);
 575                         return -ENOMEM;
 576                 }
 577                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 578
 579                 /* chain new page in list head to match sg */
 580                 first->private = (unsigned long)list;
 581                 list = first;
 582         }
 583
 584         first = get_a_page(rq, gfp);
 585         if (!first) {
 586                 give_pages(rq, list);
 587                 return -ENOMEM;
 588         }
 589         p = page_address(first);
 590
 591         /* rq->sg[0], rq->sg[1] share the same page */
 592         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 593         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 594
 595         /* rq->sg[1] for data packet, from offset */
 596         offset = sizeof(struct padded_vnet_hdr);
 597         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 598
 599         /* chain first in list head */
 600         first->private = (unsigned long)list;
 601         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 602                                   first, gfp);
 603         if (err < 0)
 604                 give_pages(rq, first);
 605
 606         return err;
 607 }
 608
 609 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
 610 {
 611         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 612         unsigned int len;
 613
 614         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
 615                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 616         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 617 }
 618
 619 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 620 {
 621         struct page_frag *alloc_frag = &rq->alloc_frag;
 622         char *buf;
 623         unsigned long ctx;
 624         int err;
 625         unsigned int len, hole;
 626
 627         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 628         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 629                 return -ENOMEM;
 630
 631         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 632         ctx = mergeable_buf_to_ctx(buf, len);
 633         get_page(alloc_frag->page);
 634         alloc_frag->offset += len;
 635         hole = alloc_frag->size - alloc_frag->offset;
 636         if (hole < len) {
 637                 /* To avoid internal fragmentation, if there is very likely not
 638                  * enough space for another buffer, add the remaining space to
 639                  * the current buffer. This extra space is not included in
 640                  * the truesize stored in ctx.
 641                  */
 642                 len += hole;
 643                 alloc_frag->offset += hole;
 644         }
 645
 646         sg_init_one(rq->sg, buf, len);
 647         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 648         if (err < 0)
 649                 put_page(virt_to_head_page(buf));
 650
 651         return err;
 652 }
 653
 654 /*
 655  * Returns false if we couldn't fill entirely (OOM).
 656  *
 657  * Normally run in the receive path, but can also be run from ndo_open
 658  * before we're receiving packets, or from refill_work which is
 659  * careful to disable receiving (using napi_disable).
 660  */
 661 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 662                           gfp_t gfp)
 663 {
 664         int err;
 665         bool oom;
 666
 667         gfp |= __GFP_COLD;
 668         do {
 669                 if (vi->mergeable_rx_bufs)
 670                         err = add_recvbuf_mergeable(rq, gfp);
 671                 else if (vi->big_packets)
 672                         err = add_recvbuf_big(vi, rq, gfp);
 673                 else
 674                         err = add_recvbuf_small(vi, rq, gfp);
 675
 676                 oom = err == -ENOMEM;
 677                 if (err)
 678                         break;
 679         } while (rq->vq->num_free);
 680         virtqueue_kick(rq->vq);
 681         return !oom;
 682 }
 683
 684 static void skb_recv_done(struct virtqueue *rvq)
 685 {
 686         struct virtnet_info *vi = rvq->vdev->priv;
 687         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 688
 689         /* Schedule NAPI, Suppress further interrupts if successful. */
 690         if (napi_schedule_prep(&rq->napi)) {
 691                 virtqueue_disable_cb(rvq);
 692                 __napi_schedule(&rq->napi);
 693         }
 694 }
 695
 696 static void virtnet_napi_enable(struct receive_queue *rq)
 697 {
 698         napi_enable(&rq->napi);
 699
 700         /* If all buffers were filled by other side before we napi_enabled, we
 701          * won't get another interrupt, so process any outstanding packets
 702          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 703          * We synchronize against interrupts via NAPI_STATE_SCHED */
 704         if (napi_schedule_prep(&rq->napi)) {
 705                 virtqueue_disable_cb(rq->vq);
 706                 local_bh_disable();
 707                 __napi_schedule(&rq->napi);
 708                 local_bh_enable();
 709         }
 710 }
 711
 712 static void refill_work(struct work_struct *work)
 713 {
 714         struct virtnet_info *vi =
 715                 container_of(work, struct virtnet_info, refill.work);
 716         bool still_empty;
 717         int i;
 718
 719         for (i = 0; i < vi->curr_queue_pairs; i++) {
 720                 struct receive_queue *rq = &vi->rq[i];
 721
 722                 napi_disable(&rq->napi);
 723                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
 724                 virtnet_napi_enable(rq);
 725
 726                 /* In theory, this can happen: if we don't get any buffers in
 727                  * we will *never* try to fill again.
 728                  */
 729                 if (still_empty)
 730                         schedule_delayed_work(&vi->refill, HZ/2);
 731         }
 732 }
 733
 734 static int virtnet_receive(struct receive_queue *rq, int budget)
 735 {
 736         struct virtnet_info *vi = rq->vq->vdev->priv;
 737         unsigned int len, received = 0;
 738         void *buf;
 739
 740         while (received < budget &&
 741                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 742                 receive_buf(vi, rq, buf, len);
 743                 received++;
 744         }
 745
 746         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 747                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 748                         schedule_delayed_work(&vi->refill, 0);
 749         }
 750
 751         return received;
 752 }
 753
 754 static int virtnet_poll(struct napi_struct *napi, int budget)
 755 {
 756         struct receive_queue *rq =
 757                 container_of(napi, struct receive_queue, napi);
 758         unsigned int r, received;
 759
 760         received = virtnet_receive(rq, budget);
 761
 762         /* Out of packets? */
 763         if (received < budget) {
 764                 r = virtqueue_enable_cb_prepare(rq->vq);
 765                 napi_complete_done(napi, received);
 766                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 767                     napi_schedule_prep(napi)) {
 768                         virtqueue_disable_cb(rq->vq);
 769                         __napi_schedule(napi);
 770                 }
 771         }
 772
 773         return received;
 774 }
 775
 776 #ifdef CONFIG_NET_RX_BUSY_POLL
 777 /* must be called with local_bh_disable()d */
 778 static int virtnet_busy_poll(struct napi_struct *napi)
 779 {
 780         struct receive_queue *rq =
 781                 container_of(napi, struct receive_queue, napi);
 782         struct virtnet_info *vi = rq->vq->vdev->priv;
 783         int r, received = 0, budget = 4;
 784
 785         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
 786                 return LL_FLUSH_FAILED;
 787
 788         if (!napi_schedule_prep(napi))
 789                 return LL_FLUSH_BUSY;
 790
 791         virtqueue_disable_cb(rq->vq);
 792
 793 again:
 794         received += virtnet_receive(rq, budget);
 795
 796         r = virtqueue_enable_cb_prepare(rq->vq);
 797         clear_bit(NAPI_STATE_SCHED, &napi->state);
 798         if (unlikely(virtqueue_poll(rq->vq, r)) &&
 799             napi_schedule_prep(napi)) {
 800                 virtqueue_disable_cb(rq->vq);
 801                 if (received < budget) {
 802                         budget -= received;
 803                         goto again;
 804                 } else {
 805                         __napi_schedule(napi);
 806                 }
 807         }
 808
 809         return received;
 810 }
 811 #endif  /* CONFIG_NET_RX_BUSY_POLL */
 812
 813 static int virtnet_open(struct net_device *dev)
 814 {
 815         struct virtnet_info *vi = netdev_priv(dev);
 816         int i;
 817
 818         for (i = 0; i < vi->max_queue_pairs; i++) {
 819                 if (i < vi->curr_queue_pairs)
 820                         /* Make sure we have some buffers: if oom use wq. */
 821                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
 822                                 schedule_delayed_work(&vi->refill, 0);
 823                 virtnet_napi_enable(&vi->rq[i]);
 824         }
 825
 826         return 0;
 827 }
 828
 829 static void free_old_xmit_skbs(struct send_queue *sq)
 830 {
 831         struct sk_buff *skb;
 832         unsigned int len;
 833         struct virtnet_info *vi = sq->vq->vdev->priv;
 834         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 835
 836         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 837                 pr_debug("Sent skb %p\n", skb);
 838
 839                 u64_stats_update_begin(&stats->tx_syncp);
 840                 stats->tx_bytes += skb->len;
 841                 stats->tx_packets++;
 842                 u64_stats_update_end(&stats->tx_syncp);
 843
 844                 dev_kfree_skb_any(skb);
 845         }
 846 }
 847
 848 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 849 {
 850         struct virtio_net_hdr_mrg_rxbuf *hdr;
 851         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 852         struct virtnet_info *vi = sq->vq->vdev->priv;
 853         unsigned num_sg;
 854         unsigned hdr_len = vi->hdr_len;
 855         bool can_push;
 856
 857         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 858
 859         can_push = vi->any_header_sg &&
 860                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 861                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 862         /* Even if we can, don't push here yet as this would skew
 863          * csum_start offset below. */
 864         if (can_push)
 865                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
 866         else
 867                 hdr = skb_vnet_hdr(skb);
 868
 869         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 870                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 871                 hdr->hdr.csum_start = cpu_to_virtio16(vi->vdev,
 872                                                 skb_checksum_start_offset(skb));
 873                 hdr->hdr.csum_offset = cpu_to_virtio16(vi->vdev,
 874                                                          skb->csum_offset);
 875         } else {
 876                 hdr->hdr.flags = 0;
 877                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 878         }
 879
 880         if (skb_is_gso(skb)) {
 881                 hdr->hdr.hdr_len = cpu_to_virtio16(vi->vdev, skb_headlen(skb));
 882                 hdr->hdr.gso_size = cpu_to_virtio16(vi->vdev,
 883                                                     skb_shinfo(skb)->gso_size);
 884                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 885                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 886                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 887                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 888                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 889                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 890                 else
 891                         BUG();
 892                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 893                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 894         } else {
 895                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 896                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 897         }
 898
 899         if (vi->mergeable_rx_bufs)
 900                 hdr->num_buffers = 0;
 901
 902         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
 903         if (can_push) {
 904                 __skb_push(skb, hdr_len);
 905                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 906                 /* Pull header back to avoid skew in tx bytes calculations. */
 907                 __skb_pull(skb, hdr_len);
 908         } else {
 909                 sg_set_buf(sq->sg, hdr, hdr_len);
 910                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 911         }
 912         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 913 }
 914
 915 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 916 {
 917         struct virtnet_info *vi = netdev_priv(dev);
 918         int qnum = skb_get_queue_mapping(skb);
 919         struct send_queue *sq = &vi->sq[qnum];
 920         int err;
 921         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
 922         bool kick = !skb->xmit_more;
 923
 924         /* Free up any pending old buffers before queueing new ones. */
 925         free_old_xmit_skbs(sq);
 926
 927         /* timestamp packet in software */
 928         skb_tx_timestamp(skb);
 929
 930         /* Try to transmit */
 931         err = xmit_skb(sq, skb);
 932
 933         /* This should not happen! */
 934         if (unlikely(err)) {
 935                 dev->stats.tx_fifo_errors++;
 936                 if (net_ratelimit())
 937                         dev_warn(&dev->dev,
 938                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 939                 dev->stats.tx_dropped++;
 940                 dev_kfree_skb_any(skb);
 941                 return NETDEV_TX_OK;
 942         }
 943
 944         /* Don't wait up for transmitted skbs to be freed. */
 945         skb_orphan(skb);
 946         nf_reset(skb);
 947
 948         /* If running out of space, stop queue to avoid getting packets that we
 949          * are then unable to transmit.
 950          * An alternative would be to force queuing layer to requeue the skb by
 951          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
 952          * returned in a normal path of operation: it means that driver is not
 953          * maintaining the TX queue stop/start state properly, and causes
 954          * the stack to do a non-trivial amount of useless work.
 955          * Since most packets only take 1 or 2 ring slots, stopping the queue
 956          * early means 16 slots are typically wasted.
 957          */
 958         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 959                 netif_stop_subqueue(dev, qnum);
 960                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 961                         /* More just got used, free them then recheck. */
 962                         free_old_xmit_skbs(sq);
 963                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 964                                 netif_start_subqueue(dev, qnum);
 965                                 virtqueue_disable_cb(sq->vq);
 966                         }
 967                 }
 968         }
 969
 970         if (kick || netif_xmit_stopped(txq))
 971                 virtqueue_kick(sq->vq);
 972
 973         return NETDEV_TX_OK;
 974 }
 975
 976 /*
 977  * Send command via the control virtqueue and check status.  Commands
 978  * supported by the hypervisor, as indicated by feature bits, should
 979  * never fail unless improperly formatted.
 980  */
 981 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 982                                  struct scatterlist *out)
 983 {
 984         struct scatterlist *sgs[4], hdr, stat;
 985         unsigned out_num = 0, tmp;
 986
 987         /* Caller should know better */
 988         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 989
 990         vi->ctrl_status = ~0;
 991         vi->ctrl_hdr.class = class;
 992         vi->ctrl_hdr.cmd = cmd;
 993         /* Add header */
 994         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
 995         sgs[out_num++] = &hdr;
 996
 997         if (out)
 998                 sgs[out_num++] = out;
 999
1000         /* Add return status. */
1001         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1002         sgs[out_num] = &stat;
1003
1004         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1005         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1006
1007         if (unlikely(!virtqueue_kick(vi->cvq)))
1008                 return vi->ctrl_status == VIRTIO_NET_OK;
1009
1010         /* Spin for a response, the kick causes an ioport write, trapping
1011          * into the hypervisor, so the request should be handled immediately.
1012          */
1013         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1014                !virtqueue_is_broken(vi->cvq))
1015                 cpu_relax();
1016
1017         return vi->ctrl_status == VIRTIO_NET_OK;
1018 }
1019
1020 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1021 {
1022         struct virtnet_info *vi = netdev_priv(dev);
1023         struct virtio_device *vdev = vi->vdev;
1024         int ret;
1025         struct sockaddr *addr = p;
1026         struct scatterlist sg;
1027
1028         ret = eth_prepare_mac_addr_change(dev, p);
1029         if (ret)
1030                 return ret;
1031
1032         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1033                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1034                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1035                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1036                         dev_warn(&vdev->dev,
1037                                  "Failed to set mac address by vq command.\n");
1038                         return -EINVAL;
1039                 }
1040         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1041                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1042                 unsigned int i;
1043
1044                 /* Naturally, this has an atomicity problem. */
1045                 for (i = 0; i < dev->addr_len; i++)
1046                         virtio_cwrite8(vdev,
1047                                        offsetof(struct virtio_net_config, mac) +
1048                                        i, addr->sa_data[i]);
1049         }
1050
1051         eth_commit_mac_addr_change(dev, p);
1052
1053         return 0;
1054 }
1055
1056 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1057                                                struct rtnl_link_stats64 *tot)
1058 {
1059         struct virtnet_info *vi = netdev_priv(dev);
1060         int cpu;
1061         unsigned int start;
1062
1063         for_each_possible_cpu(cpu) {
1064                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1065                 u64 tpackets, tbytes, rpackets, rbytes;
1066
1067                 do {
1068                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1069                         tpackets = stats->tx_packets;
1070                         tbytes   = stats->tx_bytes;
1071                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1072
1073                 do {
1074                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1075                         rpackets = stats->rx_packets;
1076                         rbytes   = stats->rx_bytes;
1077                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1078
1079                 tot->rx_packets += rpackets;
1080                 tot->tx_packets += tpackets;
1081                 tot->rx_bytes   += rbytes;
1082                 tot->tx_bytes   += tbytes;
1083         }
1084
1085         tot->tx_dropped = dev->stats.tx_dropped;
1086         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1087         tot->rx_dropped = dev->stats.rx_dropped;
1088         tot->rx_length_errors = dev->stats.rx_length_errors;
1089         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1090
1091         return tot;
1092 }
1093
1094 #ifdef CONFIG_NET_POLL_CONTROLLER
1095 static void virtnet_netpoll(struct net_device *dev)
1096 {
1097         struct virtnet_info *vi = netdev_priv(dev);
1098         int i;
1099
1100         for (i = 0; i < vi->curr_queue_pairs; i++)
1101                 napi_schedule(&vi->rq[i].napi);
1102 }
1103 #endif
1104
1105 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1106 {
1107         rtnl_lock();
1108         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1109                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1110                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1111         rtnl_unlock();
1112 }
1113
1114 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1115 {
1116         struct scatterlist sg;
1117         struct virtio_net_ctrl_mq s;
1118         struct net_device *dev = vi->dev;
1119
1120         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1121                 return 0;
1122
1123         s.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1124         sg_init_one(&sg, &s, sizeof(s));
1125
1126         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1127                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1128                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1129                          queue_pairs);
1130                 return -EINVAL;
1131         } else {
1132                 vi->curr_queue_pairs = queue_pairs;
1133                 /* virtnet_open() will refill when device is going to up. */
1134                 if (dev->flags & IFF_UP)
1135                         schedule_delayed_work(&vi->refill, 0);
1136         }
1137
1138         return 0;
1139 }
1140
1141 static int virtnet_close(struct net_device *dev)
1142 {
1143         struct virtnet_info *vi = netdev_priv(dev);
1144         int i;
1145
1146         /* Make sure refill_work doesn't re-enable napi! */
1147         cancel_delayed_work_sync(&vi->refill);
1148
1149         for (i = 0; i < vi->max_queue_pairs; i++)
1150                 napi_disable(&vi->rq[i].napi);
1151
1152         return 0;
1153 }
1154
1155 static void virtnet_set_rx_mode(struct net_device *dev)
1156 {
1157         struct virtnet_info *vi = netdev_priv(dev);
1158         struct scatterlist sg[2];
1159         struct virtio_net_ctrl_mac *mac_data;
1160         struct netdev_hw_addr *ha;
1161         int uc_count;
1162         int mc_count;
1163         void *buf;
1164         int i;
1165
1166         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1167         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1168                 return;
1169
1170         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1171         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1172
1173         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1174
1175         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1176                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1177                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1178                          vi->ctrl_promisc ? "en" : "dis");
1179
1180         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1181
1182         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1183                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1184                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1185                          vi->ctrl_allmulti ? "en" : "dis");
1186
1187         uc_count = netdev_uc_count(dev);
1188         mc_count = netdev_mc_count(dev);
1189         /* MAC filter - use one buffer for both lists */
1190         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1191                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1192         mac_data = buf;
1193         if (!buf)
1194                 return;
1195
1196         sg_init_table(sg, 2);
1197
1198         /* Store the unicast list and count in the front of the buffer */
1199         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1200         i = 0;
1201         netdev_for_each_uc_addr(ha, dev)
1202                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1203
1204         sg_set_buf(&sg[0], mac_data,
1205                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1206
1207         /* multicast list and count fill the end */
1208         mac_data = (void *)&mac_data->macs[uc_count][0];
1209
1210         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1211         i = 0;
1212         netdev_for_each_mc_addr(ha, dev)
1213                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1214
1215         sg_set_buf(&sg[1], mac_data,
1216                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1217
1218         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1219                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1220                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1221
1222         kfree(buf);
1223 }
1224
1225 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1226                                    __be16 proto, u16 vid)
1227 {
1228         struct virtnet_info *vi = netdev_priv(dev);
1229         struct scatterlist sg;
1230
1231         sg_init_one(&sg, &vid, sizeof(vid));
1232
1233         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1234                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1235                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1236         return 0;
1237 }
1238
1239 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1240                                     __be16 proto, u16 vid)
1241 {
1242         struct virtnet_info *vi = netdev_priv(dev);
1243         struct scatterlist sg;
1244
1245         sg_init_one(&sg, &vid, sizeof(vid));
1246
1247         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1248                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1249                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1250         return 0;
1251 }
1252
1253 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1254 {
1255         int i;
1256
1257         if (vi->affinity_hint_set) {
1258                 for (i = 0; i < vi->max_queue_pairs; i++) {
1259                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1260                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1261                 }
1262
1263                 vi->affinity_hint_set = false;
1264         }
1265 }
1266
1267 static void virtnet_set_affinity(struct virtnet_info *vi)
1268 {
1269         int i;
1270         int cpu;
1271
1272         /* In multiqueue mode, when the number of cpu is equal to the number of
1273          * queue pairs, we let the queue pairs to be private to one cpu by
1274          * setting the affinity hint to eliminate the contention.
1275          */
1276         if (vi->curr_queue_pairs == 1 ||
1277             vi->max_queue_pairs != num_online_cpus()) {
1278                 virtnet_clean_affinity(vi, -1);
1279                 return;
1280         }
1281
1282         i = 0;
1283         for_each_online_cpu(cpu) {
1284                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1285                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1286                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1287                 i++;
1288         }
1289
1290         vi->affinity_hint_set = true;
1291 }
1292
1293 static int virtnet_cpu_callback(struct notifier_block *nfb,
1294                                 unsigned long action, void *hcpu)
1295 {
1296         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1297
1298         switch(action & ~CPU_TASKS_FROZEN) {
1299         case CPU_ONLINE:
1300         case CPU_DOWN_FAILED:
1301         case CPU_DEAD:
1302                 virtnet_set_affinity(vi);
1303                 break;
1304         case CPU_DOWN_PREPARE:
1305                 virtnet_clean_affinity(vi, (long)hcpu);
1306                 break;
1307         default:
1308                 break;
1309         }
1310
1311         return NOTIFY_OK;
1312 }
1313
1314 static void virtnet_get_ringparam(struct net_device *dev,
1315                                 struct ethtool_ringparam *ring)
1316 {
1317         struct virtnet_info *vi = netdev_priv(dev);
1318
1319         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1320         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1321         ring->rx_pending = ring->rx_max_pending;
1322         ring->tx_pending = ring->tx_max_pending;
1323 }
1324
1325
1326 static void virtnet_get_drvinfo(struct net_device *dev,
1327                                 struct ethtool_drvinfo *info)
1328 {
1329         struct virtnet_info *vi = netdev_priv(dev);
1330         struct virtio_device *vdev = vi->vdev;
1331
1332         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1333         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1334         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1335
1336 }
1337
1338 /* TODO: Eliminate OOO packets during switching */
1339 static int virtnet_set_channels(struct net_device *dev,
1340                                 struct ethtool_channels *channels)
1341 {
1342         struct virtnet_info *vi = netdev_priv(dev);
1343         u16 queue_pairs = channels->combined_count;
1344         int err;
1345
1346         /* We don't support separate rx/tx channels.
1347          * We don't allow setting 'other' channels.
1348          */
1349         if (channels->rx_count || channels->tx_count || channels->other_count)
1350                 return -EINVAL;
1351
1352         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1353                 return -EINVAL;
1354
1355         get_online_cpus();
1356         err = virtnet_set_queues(vi, queue_pairs);
1357         if (!err) {
1358                 netif_set_real_num_tx_queues(dev, queue_pairs);
1359                 netif_set_real_num_rx_queues(dev, queue_pairs);
1360
1361                 virtnet_set_affinity(vi);
1362         }
1363         put_online_cpus();
1364
1365         return err;
1366 }
1367
1368 static void virtnet_get_channels(struct net_device *dev,
1369                                  struct ethtool_channels *channels)
1370 {
1371         struct virtnet_info *vi = netdev_priv(dev);
1372
1373         channels->combined_count = vi->curr_queue_pairs;
1374         channels->max_combined = vi->max_queue_pairs;
1375         channels->max_other = 0;
1376         channels->rx_count = 0;
1377         channels->tx_count = 0;
1378         channels->other_count = 0;
1379 }
1380
1381 static const struct ethtool_ops virtnet_ethtool_ops = {
1382         .get_drvinfo = virtnet_get_drvinfo,
1383         .get_link = ethtool_op_get_link,
1384         .get_ringparam = virtnet_get_ringparam,
1385         .set_channels = virtnet_set_channels,
1386         .get_channels = virtnet_get_channels,
1387         .get_ts_info = ethtool_op_get_ts_info,
1388 };
1389
1390 #define MIN_MTU 68
1391 #define MAX_MTU 65535
1392
1393 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1394 {
1395         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1396                 return -EINVAL;
1397         dev->mtu = new_mtu;
1398         return 0;
1399 }
1400
1401 static const struct net_device_ops virtnet_netdev = {
1402         .ndo_open            = virtnet_open,
1403         .ndo_stop            = virtnet_close,
1404         .ndo_start_xmit      = start_xmit,
1405         .ndo_validate_addr   = eth_validate_addr,
1406         .ndo_set_mac_address = virtnet_set_mac_address,
1407         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1408         .ndo_change_mtu      = virtnet_change_mtu,
1409         .ndo_get_stats64     = virtnet_stats,
1410         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1411         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1412 #ifdef CONFIG_NET_POLL_CONTROLLER
1413         .ndo_poll_controller = virtnet_netpoll,
1414 #endif
1415 #ifdef CONFIG_NET_RX_BUSY_POLL
1416         .ndo_busy_poll          = virtnet_busy_poll,
1417 #endif
1418 };
1419
1420 static void virtnet_config_changed_work(struct work_struct *work)
1421 {
1422         struct virtnet_info *vi =
1423                 container_of(work, struct virtnet_info, config_work);
1424         u16 v;
1425
1426         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1427                                  struct virtio_net_config, status, &v) < 0)
1428                 return;
1429
1430         if (v & VIRTIO_NET_S_ANNOUNCE) {
1431                 netdev_notify_peers(vi->dev);
1432                 virtnet_ack_link_announce(vi);
1433         }
1434
1435         /* Ignore unknown (future) status bits */
1436         v &= VIRTIO_NET_S_LINK_UP;
1437
1438         if (vi->status == v)
1439                 return;
1440
1441         vi->status = v;
1442
1443         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1444                 netif_carrier_on(vi->dev);
1445                 netif_tx_wake_all_queues(vi->dev);
1446         } else {
1447                 netif_carrier_off(vi->dev);
1448                 netif_tx_stop_all_queues(vi->dev);
1449         }
1450 }
1451
1452 static void virtnet_config_changed(struct virtio_device *vdev)
1453 {
1454         struct virtnet_info *vi = vdev->priv;
1455
1456         schedule_work(&vi->config_work);
1457 }
1458
1459 static void virtnet_free_queues(struct virtnet_info *vi)
1460 {
1461         int i;
1462
1463         for (i = 0; i < vi->max_queue_pairs; i++) {
1464                 napi_hash_del(&vi->rq[i].napi);
1465                 netif_napi_del(&vi->rq[i].napi);
1466         }
1467
1468         kfree(vi->rq);
1469         kfree(vi->sq);
1470 }
1471
1472 static void free_receive_bufs(struct virtnet_info *vi)
1473 {
1474         int i;
1475
1476         for (i = 0; i < vi->max_queue_pairs; i++) {
1477                 while (vi->rq[i].pages)
1478                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1479         }
1480 }
1481
1482 static void free_receive_page_frags(struct virtnet_info *vi)
1483 {
1484         int i;
1485         for (i = 0; i < vi->max_queue_pairs; i++)
1486                 if (vi->rq[i].alloc_frag.page)
1487                         put_page(vi->rq[i].alloc_frag.page);
1488 }
1489
1490 static void free_unused_bufs(struct virtnet_info *vi)
1491 {
1492         void *buf;
1493         int i;
1494
1495         for (i = 0; i < vi->max_queue_pairs; i++) {
1496                 struct virtqueue *vq = vi->sq[i].vq;
1497                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1498                         dev_kfree_skb(buf);
1499         }
1500
1501         for (i = 0; i < vi->max_queue_pairs; i++) {
1502                 struct virtqueue *vq = vi->rq[i].vq;
1503
1504                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1505                         if (vi->mergeable_rx_bufs) {
1506                                 unsigned long ctx = (unsigned long)buf;
1507                                 void *base = mergeable_ctx_to_buf_address(ctx);
1508                                 put_page(virt_to_head_page(base));
1509                         } else if (vi->big_packets) {
1510                                 give_pages(&vi->rq[i], buf);
1511                         } else {
1512                                 dev_kfree_skb(buf);
1513                         }
1514                 }
1515         }
1516 }
1517
1518 static void virtnet_del_vqs(struct virtnet_info *vi)
1519 {
1520         struct virtio_device *vdev = vi->vdev;
1521
1522         virtnet_clean_affinity(vi, -1);
1523
1524         vdev->config->del_vqs(vdev);
1525
1526         virtnet_free_queues(vi);
1527 }
1528
1529 static int virtnet_find_vqs(struct virtnet_info *vi)
1530 {
1531         vq_callback_t **callbacks;
1532         struct virtqueue **vqs;
1533         int ret = -ENOMEM;
1534         int i, total_vqs;
1535         const char **names;
1536
1537         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1538          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1539          * possible control vq.
1540          */
1541         total_vqs = vi->max_queue_pairs * 2 +
1542                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1543
1544         /* Allocate space for find_vqs parameters */
1545         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1546         if (!vqs)
1547                 goto err_vq;
1548         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1549         if (!callbacks)
1550                 goto err_callback;
1551         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1552         if (!names)
1553                 goto err_names;
1554
1555         /* Parameters for control virtqueue, if any */
1556         if (vi->has_cvq) {
1557                 callbacks[total_vqs - 1] = NULL;
1558                 names[total_vqs - 1] = "control";
1559         }
1560
1561         /* Allocate/initialize parameters for send/receive virtqueues */
1562         for (i = 0; i < vi->max_queue_pairs; i++) {
1563                 callbacks[rxq2vq(i)] = skb_recv_done;
1564                 callbacks[txq2vq(i)] = skb_xmit_done;
1565                 sprintf(vi->rq[i].name, "input.%d", i);
1566                 sprintf(vi->sq[i].name, "output.%d", i);
1567                 names[rxq2vq(i)] = vi->rq[i].name;
1568                 names[txq2vq(i)] = vi->sq[i].name;
1569         }
1570
1571         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1572                                          names);
1573         if (ret)
1574                 goto err_find;
1575
1576         if (vi->has_cvq) {
1577                 vi->cvq = vqs[total_vqs - 1];
1578                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1579                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1580         }
1581
1582         for (i = 0; i < vi->max_queue_pairs; i++) {
1583                 vi->rq[i].vq = vqs[rxq2vq(i)];
1584                 vi->sq[i].vq = vqs[txq2vq(i)];
1585         }
1586
1587         kfree(names);
1588         kfree(callbacks);
1589         kfree(vqs);
1590
1591         return 0;
1592
1593 err_find:
1594         kfree(names);
1595 err_names:
1596         kfree(callbacks);
1597 err_callback:
1598         kfree(vqs);
1599 err_vq:
1600         return ret;
1601 }
1602
1603 static int virtnet_alloc_queues(struct virtnet_info *vi)
1604 {
1605         int i;
1606
1607         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1608         if (!vi->sq)
1609                 goto err_sq;
1610         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1611         if (!vi->rq)
1612                 goto err_rq;
1613
1614         INIT_DELAYED_WORK(&vi->refill, refill_work);
1615         for (i = 0; i < vi->max_queue_pairs; i++) {
1616                 vi->rq[i].pages = NULL;
1617                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1618                                napi_weight);
1619                 napi_hash_add(&vi->rq[i].napi);
1620
1621                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1622                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
1623                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1624         }
1625
1626         return 0;
1627
1628 err_rq:
1629         kfree(vi->sq);
1630 err_sq:
1631         return -ENOMEM;
1632 }
1633
1634 static int init_vqs(struct virtnet_info *vi)
1635 {
1636         int ret;
1637
1638         /* Allocate send & receive queues */
1639         ret = virtnet_alloc_queues(vi);
1640         if (ret)
1641                 goto err;
1642
1643         ret = virtnet_find_vqs(vi);
1644         if (ret)
1645                 goto err_free;
1646
1647         get_online_cpus();
1648         virtnet_set_affinity(vi);
1649         put_online_cpus();
1650
1651         return 0;
1652
1653 err_free:
1654         virtnet_free_queues(vi);
1655 err:
1656         return ret;
1657 }
1658
1659 #ifdef CONFIG_SYSFS
1660 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1661                 struct rx_queue_attribute *attribute, char *buf)
1662 {
1663         struct virtnet_info *vi = netdev_priv(queue->dev);
1664         unsigned int queue_index = get_netdev_rx_queue_index(queue);
1665         struct ewma_pkt_len *avg;
1666
1667         BUG_ON(queue_index >= vi->max_queue_pairs);
1668         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1669         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1670 }
1671
1672 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1673         __ATTR_RO(mergeable_rx_buffer_size);
1674
1675 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1676         &mergeable_rx_buffer_size_attribute.attr,
1677         NULL
1678 };
1679
1680 static const struct attribute_group virtio_net_mrg_rx_group = {
1681         .name = "virtio_net",
1682         .attrs = virtio_net_mrg_rx_attrs
1683 };
1684 #endif
1685
1686 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1687                                     unsigned int fbit,
1688                                     const char *fname, const char *dname)
1689 {
1690         if (!virtio_has_feature(vdev, fbit))
1691                 return false;
1692
1693         dev_err(&vdev->dev, "device advertises feature %s but not %s",
1694                 fname, dname);
1695
1696         return true;
1697 }
1698
1699 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
1700         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1701
1702 static bool virtnet_validate_features(struct virtio_device *vdev)
1703 {
1704         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1705             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1706                              "VIRTIO_NET_F_CTRL_VQ") ||
1707              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1708                              "VIRTIO_NET_F_CTRL_VQ") ||
1709              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1710                              "VIRTIO_NET_F_CTRL_VQ") ||
1711              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1712              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1713                              "VIRTIO_NET_F_CTRL_VQ"))) {
1714                 return false;
1715         }
1716
1717         return true;
1718 }
1719
1720 static int virtnet_probe(struct virtio_device *vdev)
1721 {
1722         int i, err;
1723         struct net_device *dev;
1724         struct virtnet_info *vi;
1725         u16 max_queue_pairs;
1726
1727         if (!vdev->config->get) {
1728                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
1729                         __func__);
1730                 return -EINVAL;
1731         }
1732
1733         if (!virtnet_validate_features(vdev))
1734                 return -EINVAL;
1735
1736         /* Find if host supports multiqueue virtio_net device */
1737         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1738                                    struct virtio_net_config,
1739                                    max_virtqueue_pairs, &max_queue_pairs);
1740
1741         /* We need at least 2 queue's */
1742         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1743             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1744             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1745                 max_queue_pairs = 1;
1746
1747         /* Allocate ourselves a network device with room for our info */
1748         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1749         if (!dev)
1750                 return -ENOMEM;
1751
1752         /* Set up network device as normal. */
1753         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1754         dev->netdev_ops = &virtnet_netdev;
1755         dev->features = NETIF_F_HIGHDMA;
1756
1757         dev->ethtool_ops = &virtnet_ethtool_ops;
1758         SET_NETDEV_DEV(dev, &vdev->dev);
1759
1760         /* Do we support "hardware" checksums? */
1761         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1762                 /* This opens up the world of extra features. */
1763                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1764                 if (csum)
1765                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1766
1767                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1768                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1769                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1770                 }
1771                 /* Individual feature bits: what can host handle? */
1772                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1773                         dev->hw_features |= NETIF_F_TSO;
1774                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1775                         dev->hw_features |= NETIF_F_TSO6;
1776                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1777                         dev->hw_features |= NETIF_F_TSO_ECN;
1778                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1779                         dev->hw_features |= NETIF_F_UFO;
1780
1781                 dev->features |= NETIF_F_GSO_ROBUST;
1782
1783                 if (gso)
1784                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1785                 /* (!csum && gso) case will be fixed by register_netdev() */
1786         }
1787         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1788                 dev->features |= NETIF_F_RXCSUM;
1789
1790         dev->vlan_features = dev->features;
1791
1792         /* Configuration may specify what MAC to use.  Otherwise random. */
1793         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1794                 virtio_cread_bytes(vdev,
1795                                    offsetof(struct virtio_net_config, mac),
1796                                    dev->dev_addr, dev->addr_len);
1797         else
1798                 eth_hw_addr_random(dev);
1799
1800         /* Set up our device-specific information */
1801         vi = netdev_priv(dev);
1802         vi->dev = dev;
1803         vi->vdev = vdev;
1804         vdev->priv = vi;
1805         vi->stats = alloc_percpu(struct virtnet_stats);
1806         err = -ENOMEM;
1807         if (vi->stats == NULL)
1808                 goto free;
1809
1810         for_each_possible_cpu(i) {
1811                 struct virtnet_stats *virtnet_stats;
1812                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1813                 u64_stats_init(&virtnet_stats->tx_syncp);
1814                 u64_stats_init(&virtnet_stats->rx_syncp);
1815         }
1816
1817         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1818
1819         /* If we can receive ANY GSO packets, we must allocate large ones. */
1820         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1821             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1822             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1823             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1824                 vi->big_packets = true;
1825
1826         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1827                 vi->mergeable_rx_bufs = true;
1828
1829         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
1830             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1831                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1832         else
1833                 vi->hdr_len = sizeof(struct virtio_net_hdr);
1834
1835         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
1836             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1837                 vi->any_header_sg = true;
1838
1839         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1840                 vi->has_cvq = true;
1841
1842         if (vi->any_header_sg)
1843                 dev->needed_headroom = vi->hdr_len;
1844
1845         /* Use single tx/rx queue pair as default */
1846         vi->curr_queue_pairs = 1;
1847         vi->max_queue_pairs = max_queue_pairs;
1848
1849         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1850         err = init_vqs(vi);
1851         if (err)
1852                 goto free_stats;
1853
1854 #ifdef CONFIG_SYSFS
1855         if (vi->mergeable_rx_bufs)
1856                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1857 #endif
1858         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1859         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1860
1861         err = register_netdev(dev);
1862         if (err) {
1863                 pr_debug("virtio_net: registering device failed\n");
1864                 goto free_vqs;
1865         }
1866
1867         virtio_device_ready(vdev);
1868
1869         /* Last of all, set up some receive buffers. */
1870         for (i = 0; i < vi->curr_queue_pairs; i++) {
1871                 try_fill_recv(vi, &vi->rq[i], GFP_KERNEL);
1872
1873                 /* If we didn't even get one input buffer, we're useless. */
1874                 if (vi->rq[i].vq->num_free ==
1875                     virtqueue_get_vring_size(vi->rq[i].vq)) {
1876                         free_unused_bufs(vi);
1877                         err = -ENOMEM;
1878                         goto free_recv_bufs;
1879                 }
1880         }
1881
1882         vi->nb.notifier_call = &virtnet_cpu_callback;
1883         err = register_hotcpu_notifier(&vi->nb);
1884         if (err) {
1885                 pr_debug("virtio_net: registering cpu notifier failed\n");
1886                 goto free_recv_bufs;
1887         }
1888
1889         /* Assume link up if device can't report link status,
1890            otherwise get link status from config. */
1891         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1892                 netif_carrier_off(dev);
1893                 schedule_work(&vi->config_work);
1894         } else {
1895                 vi->status = VIRTIO_NET_S_LINK_UP;
1896                 netif_carrier_on(dev);
1897         }
1898
1899         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1900                  dev->name, max_queue_pairs);
1901
1902         return 0;
1903
1904 free_recv_bufs:
1905         vi->vdev->config->reset(vdev);
1906
1907         free_receive_bufs(vi);
1908         unregister_netdev(dev);
1909 free_vqs:
1910         cancel_delayed_work_sync(&vi->refill);
1911         free_receive_page_frags(vi);
1912         virtnet_del_vqs(vi);
1913 free_stats:
1914         free_percpu(vi->stats);
1915 free:
1916         free_netdev(dev);
1917         return err;
1918 }
1919
1920 static void remove_vq_common(struct virtnet_info *vi)
1921 {
1922         vi->vdev->config->reset(vi->vdev);
1923
1924         /* Free unused buffers in both send and recv, if any. */
1925         free_unused_bufs(vi);
1926
1927         free_receive_bufs(vi);
1928
1929         free_receive_page_frags(vi);
1930
1931         virtnet_del_vqs(vi);
1932 }
1933
1934 static void virtnet_remove(struct virtio_device *vdev)
1935 {
1936         struct virtnet_info *vi = vdev->priv;
1937
1938         unregister_hotcpu_notifier(&vi->nb);
1939
1940         /* Make sure no work handler is accessing the device. */
1941         flush_work(&vi->config_work);
1942
1943         unregister_netdev(vi->dev);
1944
1945         remove_vq_common(vi);
1946
1947         free_percpu(vi->stats);
1948         free_netdev(vi->dev);
1949 }
1950
1951 #ifdef CONFIG_PM_SLEEP
1952 static int virtnet_freeze(struct virtio_device *vdev)
1953 {
1954         struct virtnet_info *vi = vdev->priv;
1955         int i;
1956
1957         unregister_hotcpu_notifier(&vi->nb);
1958
1959         /* Make sure no work handler is accessing the device */
1960         flush_work(&vi->config_work);
1961
1962         netif_device_detach(vi->dev);
1963         cancel_delayed_work_sync(&vi->refill);
1964
1965         if (netif_running(vi->dev)) {
1966                 for (i = 0; i < vi->max_queue_pairs; i++)
1967                         napi_disable(&vi->rq[i].napi);
1968         }
1969
1970         remove_vq_common(vi);
1971
1972         return 0;
1973 }
1974
1975 static int virtnet_restore(struct virtio_device *vdev)
1976 {
1977         struct virtnet_info *vi = vdev->priv;
1978         int err, i;
1979
1980         err = init_vqs(vi);
1981         if (err)
1982                 return err;
1983
1984         virtio_device_ready(vdev);
1985
1986         if (netif_running(vi->dev)) {
1987                 for (i = 0; i < vi->curr_queue_pairs; i++)
1988                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1989                                 schedule_delayed_work(&vi->refill, 0);
1990
1991                 for (i = 0; i < vi->max_queue_pairs; i++)
1992                         virtnet_napi_enable(&vi->rq[i]);
1993         }
1994
1995         netif_device_attach(vi->dev);
1996
1997         rtnl_lock();
1998         virtnet_set_queues(vi, vi->curr_queue_pairs);
1999         rtnl_unlock();
2000
2001         err = register_hotcpu_notifier(&vi->nb);
2002         if (err)
2003                 return err;
2004
2005         return 0;
2006 }
2007 #endif
2008
2009 static struct virtio_device_id id_table[] = {
2010         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2011         { 0 },
2012 };
2013
2014 static unsigned int features[] = {
2015         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
2016         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
2017         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
2018         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
2019         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
2020         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
2021         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
2022         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
2023         VIRTIO_NET_F_CTRL_MAC_ADDR,
2024         VIRTIO_F_ANY_LAYOUT,
2025 };
2026
2027 static struct virtio_driver virtio_net_driver = {
2028         .feature_table = features,
2029         .feature_table_size = ARRAY_SIZE(features),
2030         .driver.name =  KBUILD_MODNAME,
2031         .driver.owner = THIS_MODULE,
2032         .id_table =     id_table,
2033         .probe =        virtnet_probe,
2034         .remove =       virtnet_remove,
2035         .config_changed = virtnet_config_changed,
2036 #ifdef CONFIG_PM_SLEEP
2037         .freeze =       virtnet_freeze,
2038         .restore =      virtnet_restore,
2039 #endif
2040 };
2041
2042 module_virtio_driver(virtio_net_driver);
2043
2044 MODULE_DEVICE_TABLE(virtio, id_table);
2045 MODULE_DESCRIPTION("Virtio network driver");
2046 MODULE_LICENSE("GPL");