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 /* 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;
 753
 754         received = virtnet_receive(rq, budget);
 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         /* If running out of space, stop queue to avoid getting packets that we
 943          * are then unable to transmit.
 944          * An alternative would be to force queuing layer to requeue the skb by
 945          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
 946          * returned in a normal path of operation: it means that driver is not
 947          * maintaining the TX queue stop/start state properly, and causes
 948          * the stack to do a non-trivial amount of useless work.
 949          * Since most packets only take 1 or 2 ring slots, stopping the queue
 950          * early means 16 slots are typically wasted.
 951          */
 952         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 953                 netif_stop_subqueue(dev, qnum);
 954                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 955                         /* More just got used, free them then recheck. */
 956                         free_old_xmit_skbs(sq);
 957                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 958                                 netif_start_subqueue(dev, qnum);
 959                                 virtqueue_disable_cb(sq->vq);
 960                         }
 961                 }
 962         }
 963
 964         if (kick || netif_xmit_stopped(txq))
 965                 virtqueue_kick(sq->vq);
 966
 967         return NETDEV_TX_OK;
 968 }
 969
 970 /*
 971  * Send command via the control virtqueue and check status.  Commands
 972  * supported by the hypervisor, as indicated by feature bits, should
 973  * never fail unless improperly formatted.
 974  */
 975 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 976                                  struct scatterlist *out)
 977 {
 978         struct scatterlist *sgs[4], hdr, stat;
 979         struct virtio_net_ctrl_hdr ctrl;
 980         virtio_net_ctrl_ack status = ~0;
 981         unsigned out_num = 0, tmp;
 982
 983         /* Caller should know better */
 984         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 985
 986         ctrl.class = class;
 987         ctrl.cmd = cmd;
 988         /* Add header */
 989         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 990         sgs[out_num++] = &hdr;
 991
 992         if (out)
 993                 sgs[out_num++] = out;
 994
 995         /* Add return status. */
 996         sg_init_one(&stat, &status, sizeof(status));
 997         sgs[out_num] = &stat;
 998
 999         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1000         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1001
1002         if (unlikely(!virtqueue_kick(vi->cvq)))
1003                 return status == VIRTIO_NET_OK;
1004
1005         /* Spin for a response, the kick causes an ioport write, trapping
1006          * into the hypervisor, so the request should be handled immediately.
1007          */
1008         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1009                !virtqueue_is_broken(vi->cvq))
1010                 cpu_relax();
1011
1012         return status == VIRTIO_NET_OK;
1013 }
1014
1015 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1016 {
1017         struct virtnet_info *vi = netdev_priv(dev);
1018         struct virtio_device *vdev = vi->vdev;
1019         int ret;
1020         struct sockaddr *addr = p;
1021         struct scatterlist sg;
1022
1023         ret = eth_prepare_mac_addr_change(dev, p);
1024         if (ret)
1025                 return ret;
1026
1027         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1028                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1029                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1030                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1031                         dev_warn(&vdev->dev,
1032                                  "Failed to set mac address by vq command.\n");
1033                         return -EINVAL;
1034                 }
1035         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1036                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1037                 unsigned int i;
1038
1039                 /* Naturally, this has an atomicity problem. */
1040                 for (i = 0; i < dev->addr_len; i++)
1041                         virtio_cwrite8(vdev,
1042                                        offsetof(struct virtio_net_config, mac) +
1043                                        i, addr->sa_data[i]);
1044         }
1045
1046         eth_commit_mac_addr_change(dev, p);
1047
1048         return 0;
1049 }
1050
1051 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1052                                                struct rtnl_link_stats64 *tot)
1053 {
1054         struct virtnet_info *vi = netdev_priv(dev);
1055         int cpu;
1056         unsigned int start;
1057
1058         for_each_possible_cpu(cpu) {
1059                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1060                 u64 tpackets, tbytes, rpackets, rbytes;
1061
1062                 do {
1063                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1064                         tpackets = stats->tx_packets;
1065                         tbytes   = stats->tx_bytes;
1066                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1067
1068                 do {
1069                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1070                         rpackets = stats->rx_packets;
1071                         rbytes   = stats->rx_bytes;
1072                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1073
1074                 tot->rx_packets += rpackets;
1075                 tot->tx_packets += tpackets;
1076                 tot->rx_bytes   += rbytes;
1077                 tot->tx_bytes   += tbytes;
1078         }
1079
1080         tot->tx_dropped = dev->stats.tx_dropped;
1081         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1082         tot->rx_dropped = dev->stats.rx_dropped;
1083         tot->rx_length_errors = dev->stats.rx_length_errors;
1084         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1085
1086         return tot;
1087 }
1088
1089 #ifdef CONFIG_NET_POLL_CONTROLLER
1090 static void virtnet_netpoll(struct net_device *dev)
1091 {
1092         struct virtnet_info *vi = netdev_priv(dev);
1093         int i;
1094
1095         for (i = 0; i < vi->curr_queue_pairs; i++)
1096                 napi_schedule(&vi->rq[i].napi);
1097 }
1098 #endif
1099
1100 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1101 {
1102         rtnl_lock();
1103         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1104                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1105                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1106         rtnl_unlock();
1107 }
1108
1109 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1110 {
1111         struct scatterlist sg;
1112         struct virtio_net_ctrl_mq s;
1113         struct net_device *dev = vi->dev;
1114
1115         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1116                 return 0;
1117
1118         s.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1119         sg_init_one(&sg, &s, sizeof(s));
1120
1121         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1122                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1123                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1124                          queue_pairs);
1125                 return -EINVAL;
1126         } else {
1127                 vi->curr_queue_pairs = queue_pairs;
1128                 /* virtnet_open() will refill when device is going to up. */
1129                 if (dev->flags & IFF_UP)
1130                         schedule_delayed_work(&vi->refill, 0);
1131         }
1132
1133         return 0;
1134 }
1135
1136 static int virtnet_close(struct net_device *dev)
1137 {
1138         struct virtnet_info *vi = netdev_priv(dev);
1139         int i;
1140
1141         /* Make sure refill_work doesn't re-enable napi! */
1142         cancel_delayed_work_sync(&vi->refill);
1143
1144         for (i = 0; i < vi->max_queue_pairs; i++)
1145                 napi_disable(&vi->rq[i].napi);
1146
1147         return 0;
1148 }
1149
1150 static void virtnet_set_rx_mode(struct net_device *dev)
1151 {
1152         struct virtnet_info *vi = netdev_priv(dev);
1153         struct scatterlist sg[2];
1154         u8 promisc, allmulti;
1155         struct virtio_net_ctrl_mac *mac_data;
1156         struct netdev_hw_addr *ha;
1157         int uc_count;
1158         int mc_count;
1159         void *buf;
1160         int i;
1161
1162         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1163         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1164                 return;
1165
1166         promisc = ((dev->flags & IFF_PROMISC) != 0);
1167         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1168
1169         sg_init_one(sg, &promisc, sizeof(promisc));
1170
1171         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1172                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1173                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1174                          promisc ? "en" : "dis");
1175
1176         sg_init_one(sg, &allmulti, sizeof(allmulti));
1177
1178         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1179                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1180                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1181                          allmulti ? "en" : "dis");
1182
1183         uc_count = netdev_uc_count(dev);
1184         mc_count = netdev_mc_count(dev);
1185         /* MAC filter - use one buffer for both lists */
1186         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1187                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1188         mac_data = buf;
1189         if (!buf)
1190                 return;
1191
1192         sg_init_table(sg, 2);
1193
1194         /* Store the unicast list and count in the front of the buffer */
1195         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1196         i = 0;
1197         netdev_for_each_uc_addr(ha, dev)
1198                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1199
1200         sg_set_buf(&sg[0], mac_data,
1201                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1202
1203         /* multicast list and count fill the end */
1204         mac_data = (void *)&mac_data->macs[uc_count][0];
1205
1206         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1207         i = 0;
1208         netdev_for_each_mc_addr(ha, dev)
1209                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1210
1211         sg_set_buf(&sg[1], mac_data,
1212                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1213
1214         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1215                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1216                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1217
1218         kfree(buf);
1219 }
1220
1221 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1222                                    __be16 proto, u16 vid)
1223 {
1224         struct virtnet_info *vi = netdev_priv(dev);
1225         struct scatterlist sg;
1226
1227         sg_init_one(&sg, &vid, sizeof(vid));
1228
1229         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1230                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1231                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1232         return 0;
1233 }
1234
1235 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1236                                     __be16 proto, u16 vid)
1237 {
1238         struct virtnet_info *vi = netdev_priv(dev);
1239         struct scatterlist sg;
1240
1241         sg_init_one(&sg, &vid, sizeof(vid));
1242
1243         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1244                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1245                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1246         return 0;
1247 }
1248
1249 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1250 {
1251         int i;
1252
1253         if (vi->affinity_hint_set) {
1254                 for (i = 0; i < vi->max_queue_pairs; i++) {
1255                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1256                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1257                 }
1258
1259                 vi->affinity_hint_set = false;
1260         }
1261 }
1262
1263 static void virtnet_set_affinity(struct virtnet_info *vi)
1264 {
1265         int i;
1266         int cpu;
1267
1268         /* In multiqueue mode, when the number of cpu is equal to the number of
1269          * queue pairs, we let the queue pairs to be private to one cpu by
1270          * setting the affinity hint to eliminate the contention.
1271          */
1272         if (vi->curr_queue_pairs == 1 ||
1273             vi->max_queue_pairs != num_online_cpus()) {
1274                 virtnet_clean_affinity(vi, -1);
1275                 return;
1276         }
1277
1278         i = 0;
1279         for_each_online_cpu(cpu) {
1280                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1281                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1282                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1283                 i++;
1284         }
1285
1286         vi->affinity_hint_set = true;
1287 }
1288
1289 static int virtnet_cpu_callback(struct notifier_block *nfb,
1290                                 unsigned long action, void *hcpu)
1291 {
1292         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1293
1294         switch(action & ~CPU_TASKS_FROZEN) {
1295         case CPU_ONLINE:
1296         case CPU_DOWN_FAILED:
1297         case CPU_DEAD:
1298                 virtnet_set_affinity(vi);
1299                 break;
1300         case CPU_DOWN_PREPARE:
1301                 virtnet_clean_affinity(vi, (long)hcpu);
1302                 break;
1303         default:
1304                 break;
1305         }
1306
1307         return NOTIFY_OK;
1308 }
1309
1310 static void virtnet_get_ringparam(struct net_device *dev,
1311                                 struct ethtool_ringparam *ring)
1312 {
1313         struct virtnet_info *vi = netdev_priv(dev);
1314
1315         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1316         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1317         ring->rx_pending = ring->rx_max_pending;
1318         ring->tx_pending = ring->tx_max_pending;
1319 }
1320
1321
1322 static void virtnet_get_drvinfo(struct net_device *dev,
1323                                 struct ethtool_drvinfo *info)
1324 {
1325         struct virtnet_info *vi = netdev_priv(dev);
1326         struct virtio_device *vdev = vi->vdev;
1327
1328         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1329         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1330         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1331
1332 }
1333
1334 /* TODO: Eliminate OOO packets during switching */
1335 static int virtnet_set_channels(struct net_device *dev,
1336                                 struct ethtool_channels *channels)
1337 {
1338         struct virtnet_info *vi = netdev_priv(dev);
1339         u16 queue_pairs = channels->combined_count;
1340         int err;
1341
1342         /* We don't support separate rx/tx channels.
1343          * We don't allow setting 'other' channels.
1344          */
1345         if (channels->rx_count || channels->tx_count || channels->other_count)
1346                 return -EINVAL;
1347
1348         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1349                 return -EINVAL;
1350
1351         get_online_cpus();
1352         err = virtnet_set_queues(vi, queue_pairs);
1353         if (!err) {
1354                 netif_set_real_num_tx_queues(dev, queue_pairs);
1355                 netif_set_real_num_rx_queues(dev, queue_pairs);
1356
1357                 virtnet_set_affinity(vi);
1358         }
1359         put_online_cpus();
1360
1361         return err;
1362 }
1363
1364 static void virtnet_get_channels(struct net_device *dev,
1365                                  struct ethtool_channels *channels)
1366 {
1367         struct virtnet_info *vi = netdev_priv(dev);
1368
1369         channels->combined_count = vi->curr_queue_pairs;
1370         channels->max_combined = vi->max_queue_pairs;
1371         channels->max_other = 0;
1372         channels->rx_count = 0;
1373         channels->tx_count = 0;
1374         channels->other_count = 0;
1375 }
1376
1377 static const struct ethtool_ops virtnet_ethtool_ops = {
1378         .get_drvinfo = virtnet_get_drvinfo,
1379         .get_link = ethtool_op_get_link,
1380         .get_ringparam = virtnet_get_ringparam,
1381         .set_channels = virtnet_set_channels,
1382         .get_channels = virtnet_get_channels,
1383         .get_ts_info = ethtool_op_get_ts_info,
1384 };
1385
1386 #define MIN_MTU 68
1387 #define MAX_MTU 65535
1388
1389 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1390 {
1391         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1392                 return -EINVAL;
1393         dev->mtu = new_mtu;
1394         return 0;
1395 }
1396
1397 static const struct net_device_ops virtnet_netdev = {
1398         .ndo_open            = virtnet_open,
1399         .ndo_stop            = virtnet_close,
1400         .ndo_start_xmit      = start_xmit,
1401         .ndo_validate_addr   = eth_validate_addr,
1402         .ndo_set_mac_address = virtnet_set_mac_address,
1403         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1404         .ndo_change_mtu      = virtnet_change_mtu,
1405         .ndo_get_stats64     = virtnet_stats,
1406         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1407         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1408 #ifdef CONFIG_NET_POLL_CONTROLLER
1409         .ndo_poll_controller = virtnet_netpoll,
1410 #endif
1411 #ifdef CONFIG_NET_RX_BUSY_POLL
1412         .ndo_busy_poll          = virtnet_busy_poll,
1413 #endif
1414 };
1415
1416 static void virtnet_config_changed_work(struct work_struct *work)
1417 {
1418         struct virtnet_info *vi =
1419                 container_of(work, struct virtnet_info, config_work);
1420         u16 v;
1421
1422         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1423                                  struct virtio_net_config, status, &v) < 0)
1424                 return;
1425
1426         if (v & VIRTIO_NET_S_ANNOUNCE) {
1427                 netdev_notify_peers(vi->dev);
1428                 virtnet_ack_link_announce(vi);
1429         }
1430
1431         /* Ignore unknown (future) status bits */
1432         v &= VIRTIO_NET_S_LINK_UP;
1433
1434         if (vi->status == v)
1435                 return;
1436
1437         vi->status = v;
1438
1439         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1440                 netif_carrier_on(vi->dev);
1441                 netif_tx_wake_all_queues(vi->dev);
1442         } else {
1443                 netif_carrier_off(vi->dev);
1444                 netif_tx_stop_all_queues(vi->dev);
1445         }
1446 }
1447
1448 static void virtnet_config_changed(struct virtio_device *vdev)
1449 {
1450         struct virtnet_info *vi = vdev->priv;
1451
1452         schedule_work(&vi->config_work);
1453 }
1454
1455 static void virtnet_free_queues(struct virtnet_info *vi)
1456 {
1457         int i;
1458
1459         for (i = 0; i < vi->max_queue_pairs; i++) {
1460                 napi_hash_del(&vi->rq[i].napi);
1461                 netif_napi_del(&vi->rq[i].napi);
1462         }
1463
1464         kfree(vi->rq);
1465         kfree(vi->sq);
1466 }
1467
1468 static void free_receive_bufs(struct virtnet_info *vi)
1469 {
1470         int i;
1471
1472         for (i = 0; i < vi->max_queue_pairs; i++) {
1473                 while (vi->rq[i].pages)
1474                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1475         }
1476 }
1477
1478 static void free_receive_page_frags(struct virtnet_info *vi)
1479 {
1480         int i;
1481         for (i = 0; i < vi->max_queue_pairs; i++)
1482                 if (vi->rq[i].alloc_frag.page)
1483                         put_page(vi->rq[i].alloc_frag.page);
1484 }
1485
1486 static void free_unused_bufs(struct virtnet_info *vi)
1487 {
1488         void *buf;
1489         int i;
1490
1491         for (i = 0; i < vi->max_queue_pairs; i++) {
1492                 struct virtqueue *vq = vi->sq[i].vq;
1493                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1494                         dev_kfree_skb(buf);
1495         }
1496
1497         for (i = 0; i < vi->max_queue_pairs; i++) {
1498                 struct virtqueue *vq = vi->rq[i].vq;
1499
1500                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1501                         if (vi->mergeable_rx_bufs) {
1502                                 unsigned long ctx = (unsigned long)buf;
1503                                 void *base = mergeable_ctx_to_buf_address(ctx);
1504                                 put_page(virt_to_head_page(base));
1505                         } else if (vi->big_packets) {
1506                                 give_pages(&vi->rq[i], buf);
1507                         } else {
1508                                 dev_kfree_skb(buf);
1509                         }
1510                 }
1511         }
1512 }
1513
1514 static void virtnet_del_vqs(struct virtnet_info *vi)
1515 {
1516         struct virtio_device *vdev = vi->vdev;
1517
1518         virtnet_clean_affinity(vi, -1);
1519
1520         vdev->config->del_vqs(vdev);
1521
1522         virtnet_free_queues(vi);
1523 }
1524
1525 static int virtnet_find_vqs(struct virtnet_info *vi)
1526 {
1527         vq_callback_t **callbacks;
1528         struct virtqueue **vqs;
1529         int ret = -ENOMEM;
1530         int i, total_vqs;
1531         const char **names;
1532
1533         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1534          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1535          * possible control vq.
1536          */
1537         total_vqs = vi->max_queue_pairs * 2 +
1538                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1539
1540         /* Allocate space for find_vqs parameters */
1541         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1542         if (!vqs)
1543                 goto err_vq;
1544         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1545         if (!callbacks)
1546                 goto err_callback;
1547         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1548         if (!names)
1549                 goto err_names;
1550
1551         /* Parameters for control virtqueue, if any */
1552         if (vi->has_cvq) {
1553                 callbacks[total_vqs - 1] = NULL;
1554                 names[total_vqs - 1] = "control";
1555         }
1556
1557         /* Allocate/initialize parameters for send/receive virtqueues */
1558         for (i = 0; i < vi->max_queue_pairs; i++) {
1559                 callbacks[rxq2vq(i)] = skb_recv_done;
1560                 callbacks[txq2vq(i)] = skb_xmit_done;
1561                 sprintf(vi->rq[i].name, "input.%d", i);
1562                 sprintf(vi->sq[i].name, "output.%d", i);
1563                 names[rxq2vq(i)] = vi->rq[i].name;
1564                 names[txq2vq(i)] = vi->sq[i].name;
1565         }
1566
1567         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1568                                          names);
1569         if (ret)
1570                 goto err_find;
1571
1572         if (vi->has_cvq) {
1573                 vi->cvq = vqs[total_vqs - 1];
1574                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1575                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1576         }
1577
1578         for (i = 0; i < vi->max_queue_pairs; i++) {
1579                 vi->rq[i].vq = vqs[rxq2vq(i)];
1580                 vi->sq[i].vq = vqs[txq2vq(i)];
1581         }
1582
1583         kfree(names);
1584         kfree(callbacks);
1585         kfree(vqs);
1586
1587         return 0;
1588
1589 err_find:
1590         kfree(names);
1591 err_names:
1592         kfree(callbacks);
1593 err_callback:
1594         kfree(vqs);
1595 err_vq:
1596         return ret;
1597 }
1598
1599 static int virtnet_alloc_queues(struct virtnet_info *vi)
1600 {
1601         int i;
1602
1603         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1604         if (!vi->sq)
1605                 goto err_sq;
1606         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1607         if (!vi->rq)
1608                 goto err_rq;
1609
1610         INIT_DELAYED_WORK(&vi->refill, refill_work);
1611         for (i = 0; i < vi->max_queue_pairs; i++) {
1612                 vi->rq[i].pages = NULL;
1613                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1614                                napi_weight);
1615                 napi_hash_add(&vi->rq[i].napi);
1616
1617                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1618                 ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
1619                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1620         }
1621
1622         return 0;
1623
1624 err_rq:
1625         kfree(vi->sq);
1626 err_sq:
1627         return -ENOMEM;
1628 }
1629
1630 static int init_vqs(struct virtnet_info *vi)
1631 {
1632         int ret;
1633
1634         /* Allocate send & receive queues */
1635         ret = virtnet_alloc_queues(vi);
1636         if (ret)
1637                 goto err;
1638
1639         ret = virtnet_find_vqs(vi);
1640         if (ret)
1641                 goto err_free;
1642
1643         get_online_cpus();
1644         virtnet_set_affinity(vi);
1645         put_online_cpus();
1646
1647         return 0;
1648
1649 err_free:
1650         virtnet_free_queues(vi);
1651 err:
1652         return ret;
1653 }
1654
1655 #ifdef CONFIG_SYSFS
1656 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1657                 struct rx_queue_attribute *attribute, char *buf)
1658 {
1659         struct virtnet_info *vi = netdev_priv(queue->dev);
1660         unsigned int queue_index = get_netdev_rx_queue_index(queue);
1661         struct ewma *avg;
1662
1663         BUG_ON(queue_index >= vi->max_queue_pairs);
1664         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1665         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1666 }
1667
1668 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1669         __ATTR_RO(mergeable_rx_buffer_size);
1670
1671 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1672         &mergeable_rx_buffer_size_attribute.attr,
1673         NULL
1674 };
1675
1676 static const struct attribute_group virtio_net_mrg_rx_group = {
1677         .name = "virtio_net",
1678         .attrs = virtio_net_mrg_rx_attrs
1679 };
1680 #endif
1681
1682 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1683                                     unsigned int fbit,
1684                                     const char *fname, const char *dname)
1685 {
1686         if (!virtio_has_feature(vdev, fbit))
1687                 return false;
1688
1689         dev_err(&vdev->dev, "device advertises feature %s but not %s",
1690                 fname, dname);
1691
1692         return true;
1693 }
1694
1695 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
1696         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1697
1698 static bool virtnet_validate_features(struct virtio_device *vdev)
1699 {
1700         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1701             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1702                              "VIRTIO_NET_F_CTRL_VQ") ||
1703              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1704                              "VIRTIO_NET_F_CTRL_VQ") ||
1705              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1706                              "VIRTIO_NET_F_CTRL_VQ") ||
1707              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1708              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1709                              "VIRTIO_NET_F_CTRL_VQ"))) {
1710                 return false;
1711         }
1712
1713         return true;
1714 }
1715
1716 static int virtnet_probe(struct virtio_device *vdev)
1717 {
1718         int i, err;
1719         struct net_device *dev;
1720         struct virtnet_info *vi;
1721         u16 max_queue_pairs;
1722
1723         if (!vdev->config->get) {
1724                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
1725                         __func__);
1726                 return -EINVAL;
1727         }
1728
1729         if (!virtnet_validate_features(vdev))
1730                 return -EINVAL;
1731
1732         /* Find if host supports multiqueue virtio_net device */
1733         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1734                                    struct virtio_net_config,
1735                                    max_virtqueue_pairs, &max_queue_pairs);
1736
1737         /* We need at least 2 queue's */
1738         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1739             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1740             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1741                 max_queue_pairs = 1;
1742
1743         /* Allocate ourselves a network device with room for our info */
1744         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1745         if (!dev)
1746                 return -ENOMEM;
1747
1748         /* Set up network device as normal. */
1749         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1750         dev->netdev_ops = &virtnet_netdev;
1751         dev->features = NETIF_F_HIGHDMA;
1752
1753         dev->ethtool_ops = &virtnet_ethtool_ops;
1754         SET_NETDEV_DEV(dev, &vdev->dev);
1755
1756         /* Do we support "hardware" checksums? */
1757         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1758                 /* This opens up the world of extra features. */
1759                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1760                 if (csum)
1761                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1762
1763                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1764                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1765                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1766                 }
1767                 /* Individual feature bits: what can host handle? */
1768                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1769                         dev->hw_features |= NETIF_F_TSO;
1770                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1771                         dev->hw_features |= NETIF_F_TSO6;
1772                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1773                         dev->hw_features |= NETIF_F_TSO_ECN;
1774                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1775                         dev->hw_features |= NETIF_F_UFO;
1776
1777                 dev->features |= NETIF_F_GSO_ROBUST;
1778
1779                 if (gso)
1780                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1781                 /* (!csum && gso) case will be fixed by register_netdev() */
1782         }
1783         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1784                 dev->features |= NETIF_F_RXCSUM;
1785
1786         dev->vlan_features = dev->features;
1787
1788         /* Configuration may specify what MAC to use.  Otherwise random. */
1789         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1790                 virtio_cread_bytes(vdev,
1791                                    offsetof(struct virtio_net_config, mac),
1792                                    dev->dev_addr, dev->addr_len);
1793         else
1794                 eth_hw_addr_random(dev);
1795
1796         /* Set up our device-specific information */
1797         vi = netdev_priv(dev);
1798         vi->dev = dev;
1799         vi->vdev = vdev;
1800         vdev->priv = vi;
1801         vi->stats = alloc_percpu(struct virtnet_stats);
1802         err = -ENOMEM;
1803         if (vi->stats == NULL)
1804                 goto free;
1805
1806         for_each_possible_cpu(i) {
1807                 struct virtnet_stats *virtnet_stats;
1808                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1809                 u64_stats_init(&virtnet_stats->tx_syncp);
1810                 u64_stats_init(&virtnet_stats->rx_syncp);
1811         }
1812
1813         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1814
1815         /* If we can receive ANY GSO packets, we must allocate large ones. */
1816         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1817             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1818             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1819             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1820                 vi->big_packets = true;
1821
1822         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1823                 vi->mergeable_rx_bufs = true;
1824
1825         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
1826             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1827                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1828         else
1829                 vi->hdr_len = sizeof(struct virtio_net_hdr);
1830
1831         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
1832             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1833                 vi->any_header_sg = true;
1834
1835         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1836                 vi->has_cvq = true;
1837
1838         if (vi->any_header_sg)
1839                 dev->needed_headroom = vi->hdr_len;
1840
1841         /* Use single tx/rx queue pair as default */
1842         vi->curr_queue_pairs = 1;
1843         vi->max_queue_pairs = max_queue_pairs;
1844
1845         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1846         err = init_vqs(vi);
1847         if (err)
1848                 goto free_stats;
1849
1850 #ifdef CONFIG_SYSFS
1851         if (vi->mergeable_rx_bufs)
1852                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1853 #endif
1854         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1855         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1856
1857         err = register_netdev(dev);
1858         if (err) {
1859                 pr_debug("virtio_net: registering device failed\n");
1860                 goto free_vqs;
1861         }
1862
1863         virtio_device_ready(vdev);
1864
1865         /* Last of all, set up some receive buffers. */
1866         for (i = 0; i < vi->curr_queue_pairs; i++) {
1867                 try_fill_recv(vi, &vi->rq[i], GFP_KERNEL);
1868
1869                 /* If we didn't even get one input buffer, we're useless. */
1870                 if (vi->rq[i].vq->num_free ==
1871                     virtqueue_get_vring_size(vi->rq[i].vq)) {
1872                         free_unused_bufs(vi);
1873                         err = -ENOMEM;
1874                         goto free_recv_bufs;
1875                 }
1876         }
1877
1878         vi->nb.notifier_call = &virtnet_cpu_callback;
1879         err = register_hotcpu_notifier(&vi->nb);
1880         if (err) {
1881                 pr_debug("virtio_net: registering cpu notifier failed\n");
1882                 goto free_recv_bufs;
1883         }
1884
1885         /* Assume link up if device can't report link status,
1886            otherwise get link status from config. */
1887         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1888                 netif_carrier_off(dev);
1889                 schedule_work(&vi->config_work);
1890         } else {
1891                 vi->status = VIRTIO_NET_S_LINK_UP;
1892                 netif_carrier_on(dev);
1893         }
1894
1895         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1896                  dev->name, max_queue_pairs);
1897
1898         return 0;
1899
1900 free_recv_bufs:
1901         vi->vdev->config->reset(vdev);
1902
1903         free_receive_bufs(vi);
1904         unregister_netdev(dev);
1905 free_vqs:
1906         cancel_delayed_work_sync(&vi->refill);
1907         free_receive_page_frags(vi);
1908         virtnet_del_vqs(vi);
1909 free_stats:
1910         free_percpu(vi->stats);
1911 free:
1912         free_netdev(dev);
1913         return err;
1914 }
1915
1916 static void remove_vq_common(struct virtnet_info *vi)
1917 {
1918         vi->vdev->config->reset(vi->vdev);
1919
1920         /* Free unused buffers in both send and recv, if any. */
1921         free_unused_bufs(vi);
1922
1923         free_receive_bufs(vi);
1924
1925         free_receive_page_frags(vi);
1926
1927         virtnet_del_vqs(vi);
1928 }
1929
1930 static void virtnet_remove(struct virtio_device *vdev)
1931 {
1932         struct virtnet_info *vi = vdev->priv;
1933
1934         unregister_hotcpu_notifier(&vi->nb);
1935
1936         /* Make sure no work handler is accessing the device. */
1937         flush_work(&vi->config_work);
1938
1939         unregister_netdev(vi->dev);
1940
1941         remove_vq_common(vi);
1942
1943         free_percpu(vi->stats);
1944         free_netdev(vi->dev);
1945 }
1946
1947 #ifdef CONFIG_PM_SLEEP
1948 static int virtnet_freeze(struct virtio_device *vdev)
1949 {
1950         struct virtnet_info *vi = vdev->priv;
1951         int i;
1952
1953         unregister_hotcpu_notifier(&vi->nb);
1954
1955         /* Make sure no work handler is accessing the device */
1956         flush_work(&vi->config_work);
1957
1958         netif_device_detach(vi->dev);
1959         cancel_delayed_work_sync(&vi->refill);
1960
1961         if (netif_running(vi->dev)) {
1962                 for (i = 0; i < vi->max_queue_pairs; i++)
1963                         napi_disable(&vi->rq[i].napi);
1964         }
1965
1966         remove_vq_common(vi);
1967
1968         return 0;
1969 }
1970
1971 static int virtnet_restore(struct virtio_device *vdev)
1972 {
1973         struct virtnet_info *vi = vdev->priv;
1974         int err, i;
1975
1976         err = init_vqs(vi);
1977         if (err)
1978                 return err;
1979
1980         virtio_device_ready(vdev);
1981
1982         if (netif_running(vi->dev)) {
1983                 for (i = 0; i < vi->curr_queue_pairs; i++)
1984                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1985                                 schedule_delayed_work(&vi->refill, 0);
1986
1987                 for (i = 0; i < vi->max_queue_pairs; i++)
1988                         virtnet_napi_enable(&vi->rq[i]);
1989         }
1990
1991         netif_device_attach(vi->dev);
1992
1993         rtnl_lock();
1994         virtnet_set_queues(vi, vi->curr_queue_pairs);
1995         rtnl_unlock();
1996
1997         err = register_hotcpu_notifier(&vi->nb);
1998         if (err)
1999                 return err;
2000
2001         return 0;
2002 }
2003 #endif
2004
2005 static struct virtio_device_id id_table[] = {
2006         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2007         { 0 },
2008 };
2009
2010 static unsigned int features[] = {
2011         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
2012         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
2013         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
2014         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
2015         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
2016         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
2017         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
2018         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
2019         VIRTIO_NET_F_CTRL_MAC_ADDR,
2020         VIRTIO_F_ANY_LAYOUT,
2021 };
2022
2023 static struct virtio_driver virtio_net_driver = {
2024         .feature_table = features,
2025         .feature_table_size = ARRAY_SIZE(features),
2026         .driver.name =  KBUILD_MODNAME,
2027         .driver.owner = THIS_MODULE,
2028         .id_table =     id_table,
2029         .probe =        virtnet_probe,
2030         .remove =       virtnet_remove,
2031         .config_changed = virtnet_config_changed,
2032 #ifdef CONFIG_PM_SLEEP
2033         .freeze =       virtnet_freeze,
2034         .restore =      virtnet_restore,
2035 #endif
2036 };
2037
2038 module_virtio_driver(virtio_net_driver);
2039
2040 MODULE_DEVICE_TABLE(virtio, id_table);
2041 MODULE_DESCRIPTION("Virtio network driver");
2042 MODULE_LICENSE("GPL");