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         /* Ethtool settings */
 151         u8 duplex;
 152         u32 speed;
 153 };
 154
 155 struct padded_vnet_hdr {
 156         struct virtio_net_hdr_mrg_rxbuf hdr;
 157         /*
 158          * hdr is in a separate sg buffer, and data sg buffer shares same page
 159          * with this header sg. This padding makes next sg 16 byte aligned
 160          * after the header.
 161          */
 162         char padding[4];
 163 };
 164
 165 /* Converting between virtqueue no. and kernel tx/rx queue no.
 166  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 167  */
 168 static int vq2txq(struct virtqueue *vq)
 169 {
 170         return (vq->index - 1) / 2;
 171 }
 172
 173 static int txq2vq(int txq)
 174 {
 175         return txq * 2 + 1;
 176 }
 177
 178 static int vq2rxq(struct virtqueue *vq)
 179 {
 180         return vq->index / 2;
 181 }
 182
 183 static int rxq2vq(int rxq)
 184 {
 185         return rxq * 2;
 186 }
 187
 188 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 189 {
 190         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 191 }
 192
 193 /*
 194  * private is used to chain pages for big packets, put the whole
 195  * most recent used list in the beginning for reuse
 196  */
 197 static void give_pages(struct receive_queue *rq, struct page *page)
 198 {
 199         struct page *end;
 200
 201         /* Find end of list, sew whole thing into vi->rq.pages. */
 202         for (end = page; end->private; end = (struct page *)end->private);
 203         end->private = (unsigned long)rq->pages;
 204         rq->pages = page;
 205 }
 206
 207 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 208 {
 209         struct page *p = rq->pages;
 210
 211         if (p) {
 212                 rq->pages = (struct page *)p->private;
 213                 /* clear private here, it is used to chain pages */
 214                 p->private = 0;
 215         } else
 216                 p = alloc_page(gfp_mask);
 217         return p;
 218 }
 219
 220 static void skb_xmit_done(struct virtqueue *vq)
 221 {
 222         struct virtnet_info *vi = vq->vdev->priv;
 223
 224         /* Suppress further interrupts. */
 225         virtqueue_disable_cb(vq);
 226
 227         /* We were probably waiting for more output buffers. */
 228         netif_wake_subqueue(vi->dev, vq2txq(vq));
 229 }
 230
 231 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
 232 {
 233         unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
 234         return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
 235 }
 236
 237 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
 238 {
 239         return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
 240
 241 }
 242
 243 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
 244 {
 245         unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
 246         return (unsigned long)buf | (size - 1);
 247 }
 248
 249 /* Called from bottom half context */
 250 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 251                                    struct receive_queue *rq,
 252                                    struct page *page, unsigned int offset,
 253                                    unsigned int len, unsigned int truesize)
 254 {
 255         struct sk_buff *skb;
 256         struct virtio_net_hdr_mrg_rxbuf *hdr;
 257         unsigned int copy, hdr_len, hdr_padded_len;
 258         char *p;
 259
 260         p = page_address(page) + offset;
 261
 262         /* copy small packet so we can reuse these pages for small data */
 263         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 264         if (unlikely(!skb))
 265                 return NULL;
 266
 267         hdr = skb_vnet_hdr(skb);
 268
 269         hdr_len = vi->hdr_len;
 270         if (vi->mergeable_rx_bufs)
 271                 hdr_padded_len = sizeof *hdr;
 272         else
 273                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 274
 275         memcpy(hdr, p, hdr_len);
 276
 277         len -= hdr_len;
 278         offset += hdr_padded_len;
 279         p += hdr_padded_len;
 280
 281         copy = len;
 282         if (copy > skb_tailroom(skb))
 283                 copy = skb_tailroom(skb);
 284         memcpy(skb_put(skb, copy), p, copy);
 285
 286         len -= copy;
 287         offset += copy;
 288
 289         if (vi->mergeable_rx_bufs) {
 290                 if (len)
 291                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 292                 else
 293                         put_page(page);
 294                 return skb;
 295         }
 296
 297         /*
 298          * Verify that we can indeed put this data into a skb.
 299          * This is here to handle cases when the device erroneously
 300          * tries to receive more than is possible. This is usually
 301          * the case of a broken device.
 302          */
 303         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 304                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 305                 dev_kfree_skb(skb);
 306                 return NULL;
 307         }
 308         BUG_ON(offset >= PAGE_SIZE);
 309         while (len) {
 310                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 311                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 312                                 frag_size, truesize);
 313                 len -= frag_size;
 314                 page = (struct page *)page->private;
 315                 offset = 0;
 316         }
 317
 318         if (page)
 319                 give_pages(rq, page);
 320
 321         return skb;
 322 }
 323
 324 static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
 325 {
 326         struct sk_buff * skb = buf;
 327
 328         len -= vi->hdr_len;
 329         skb_trim(skb, len);
 330
 331         return skb;
 332 }
 333
 334 static struct sk_buff *receive_big(struct net_device *dev,
 335                                    struct virtnet_info *vi,
 336                                    struct receive_queue *rq,
 337                                    void *buf,
 338                                    unsigned int len)
 339 {
 340         struct page *page = buf;
 341         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 342
 343         if (unlikely(!skb))
 344                 goto err;
 345
 346         return skb;
 347
 348 err:
 349         dev->stats.rx_dropped++;
 350         give_pages(rq, page);
 351         return NULL;
 352 }
 353
 354 static struct sk_buff *receive_mergeable(struct net_device *dev,
 355                                          struct virtnet_info *vi,
 356                                          struct receive_queue *rq,
 357                                          unsigned long ctx,
 358                                          unsigned int len)
 359 {
 360         void *buf = mergeable_ctx_to_buf_address(ctx);
 361         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 362         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 363         struct page *page = virt_to_head_page(buf);
 364         int offset = buf - page_address(page);
 365         unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 366
 367         struct sk_buff *head_skb = page_to_skb(vi, rq, page, offset, len,
 368                                                truesize);
 369         struct sk_buff *curr_skb = head_skb;
 370
 371         if (unlikely(!curr_skb))
 372                 goto err_skb;
 373         while (--num_buf) {
 374                 int num_skb_frags;
 375
 376                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 377                 if (unlikely(!ctx)) {
 378                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 379                                  dev->name, num_buf,
 380                                  virtio16_to_cpu(vi->vdev,
 381                                                  hdr->num_buffers));
 382                         dev->stats.rx_length_errors++;
 383                         goto err_buf;
 384                 }
 385
 386                 buf = mergeable_ctx_to_buf_address(ctx);
 387                 page = virt_to_head_page(buf);
 388
 389                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 390                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 391                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 392
 393                         if (unlikely(!nskb))
 394                                 goto err_skb;
 395                         if (curr_skb == head_skb)
 396                                 skb_shinfo(curr_skb)->frag_list = nskb;
 397                         else
 398                                 curr_skb->next = nskb;
 399                         curr_skb = nskb;
 400                         head_skb->truesize += nskb->truesize;
 401                         num_skb_frags = 0;
 402                 }
 403                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 404                 if (curr_skb != head_skb) {
 405                         head_skb->data_len += len;
 406                         head_skb->len += len;
 407                         head_skb->truesize += truesize;
 408                 }
 409                 offset = buf - page_address(page);
 410                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 411                         put_page(page);
 412                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 413                                              len, truesize);
 414                 } else {
 415                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 416                                         offset, len, truesize);
 417                 }
 418         }
 419
 420         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 421         return head_skb;
 422
 423 err_skb:
 424         put_page(page);
 425         while (--num_buf) {
 426                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 427                 if (unlikely(!ctx)) {
 428                         pr_debug("%s: rx error: %d buffers missing\n",
 429                                  dev->name, num_buf);
 430                         dev->stats.rx_length_errors++;
 431                         break;
 432                 }
 433                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 434                 put_page(page);
 435         }
 436 err_buf:
 437         dev->stats.rx_dropped++;
 438         dev_kfree_skb(head_skb);
 439         return NULL;
 440 }
 441
 442 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 443                         void *buf, unsigned int len)
 444 {
 445         struct net_device *dev = vi->dev;
 446         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 447         struct sk_buff *skb;
 448         struct virtio_net_hdr_mrg_rxbuf *hdr;
 449
 450         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 451                 pr_debug("%s: short packet %i\n", dev->name, len);
 452                 dev->stats.rx_length_errors++;
 453                 if (vi->mergeable_rx_bufs) {
 454                         unsigned long ctx = (unsigned long)buf;
 455                         void *base = mergeable_ctx_to_buf_address(ctx);
 456                         put_page(virt_to_head_page(base));
 457                 } else if (vi->big_packets) {
 458                         give_pages(rq, buf);
 459                 } else {
 460                         dev_kfree_skb(buf);
 461                 }
 462                 return;
 463         }
 464
 465         if (vi->mergeable_rx_bufs)
 466                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
 467         else if (vi->big_packets)
 468                 skb = receive_big(dev, vi, rq, buf, len);
 469         else
 470                 skb = receive_small(vi, buf, len);
 471
 472         if (unlikely(!skb))
 473                 return;
 474
 475         hdr = skb_vnet_hdr(skb);
 476
 477         u64_stats_update_begin(&stats->rx_syncp);
 478         stats->rx_bytes += skb->len;
 479         stats->rx_packets++;
 480         u64_stats_update_end(&stats->rx_syncp);
 481
 482         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 483                 pr_debug("Needs csum!\n");
 484                 if (!skb_partial_csum_set(skb,
 485                           virtio16_to_cpu(vi->vdev, hdr->hdr.csum_start),
 486                           virtio16_to_cpu(vi->vdev, hdr->hdr.csum_offset)))
 487                         goto frame_err;
 488         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 489                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 490         }
 491
 492         skb->protocol = eth_type_trans(skb, dev);
 493         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 494                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 495
 496         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 497                 pr_debug("GSO!\n");
 498                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 499                 case VIRTIO_NET_HDR_GSO_TCPV4:
 500                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 501                         break;
 502                 case VIRTIO_NET_HDR_GSO_UDP:
 503                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 504                         break;
 505                 case VIRTIO_NET_HDR_GSO_TCPV6:
 506                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 507                         break;
 508                 default:
 509                         net_warn_ratelimited("%s: bad gso type %u.\n",
 510                                              dev->name, hdr->hdr.gso_type);
 511                         goto frame_err;
 512                 }
 513
 514                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 515                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 516
 517                 skb_shinfo(skb)->gso_size = virtio16_to_cpu(vi->vdev,
 518                                                             hdr->hdr.gso_size);
 519                 if (skb_shinfo(skb)->gso_size == 0) {
 520                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 521                         goto frame_err;
 522                 }
 523
 524                 /* Header must be checked, and gso_segs computed. */
 525                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 526                 skb_shinfo(skb)->gso_segs = 0;
 527         }
 528
 529         napi_gro_receive(&rq->napi, skb);
 530         return;
 531
 532 frame_err:
 533         dev->stats.rx_frame_errors++;
 534         dev_kfree_skb(skb);
 535 }
 536
 537 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 538                              gfp_t gfp)
 539 {
 540         struct sk_buff *skb;
 541         struct virtio_net_hdr_mrg_rxbuf *hdr;
 542         int err;
 543
 544         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 545         if (unlikely(!skb))
 546                 return -ENOMEM;
 547
 548         skb_put(skb, GOOD_PACKET_LEN);
 549
 550         hdr = skb_vnet_hdr(skb);
 551         sg_init_table(rq->sg, 2);
 552         sg_set_buf(rq->sg, hdr, vi->hdr_len);
 553         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 554
 555         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 556         if (err < 0)
 557                 dev_kfree_skb(skb);
 558
 559         return err;
 560 }
 561
 562 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 563                            gfp_t gfp)
 564 {
 565         struct page *first, *list = NULL;
 566         char *p;
 567         int i, err, offset;
 568
 569         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 570
 571         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 572         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 573                 first = get_a_page(rq, gfp);
 574                 if (!first) {
 575                         if (list)
 576                                 give_pages(rq, list);
 577                         return -ENOMEM;
 578                 }
 579                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 580
 581                 /* chain new page in list head to match sg */
 582                 first->private = (unsigned long)list;
 583                 list = first;
 584         }
 585
 586         first = get_a_page(rq, gfp);
 587         if (!first) {
 588                 give_pages(rq, list);
 589                 return -ENOMEM;
 590         }
 591         p = page_address(first);
 592
 593         /* rq->sg[0], rq->sg[1] share the same page */
 594         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 595         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 596
 597         /* rq->sg[1] for data packet, from offset */
 598         offset = sizeof(struct padded_vnet_hdr);
 599         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 600
 601         /* chain first in list head */
 602         first->private = (unsigned long)list;
 603         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 604                                   first, gfp);
 605         if (err < 0)
 606                 give_pages(rq, first);
 607
 608         return err;
 609 }
 610
 611 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
 612 {
 613         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 614         unsigned int len;
 615
 616         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
 617                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 618         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 619 }
 620
 621 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 622 {
 623         struct page_frag *alloc_frag = &rq->alloc_frag;
 624         char *buf;
 625         unsigned long ctx;
 626         int err;
 627         unsigned int len, hole;
 628
 629         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 630         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 631                 return -ENOMEM;
 632
 633         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 634         ctx = mergeable_buf_to_ctx(buf, len);
 635         get_page(alloc_frag->page);
 636         alloc_frag->offset += len;
 637         hole = alloc_frag->size - alloc_frag->offset;
 638         if (hole < len) {
 639                 /* To avoid internal fragmentation, if there is very likely not
 640                  * enough space for another buffer, add the remaining space to
 641                  * the current buffer. This extra space is not included in
 642                  * the truesize stored in ctx.
 643                  */
 644                 len += hole;
 645                 alloc_frag->offset += hole;
 646         }
 647
 648         sg_init_one(rq->sg, buf, len);
 649         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 650         if (err < 0)
 651                 put_page(virt_to_head_page(buf));
 652
 653         return err;
 654 }
 655
 656 /*
 657  * Returns false if we couldn't fill entirely (OOM).
 658  *
 659  * Normally run in the receive path, but can also be run from ndo_open
 660  * before we're receiving packets, or from refill_work which is
 661  * careful to disable receiving (using napi_disable).
 662  */
 663 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 664                           gfp_t gfp)
 665 {
 666         int err;
 667         bool oom;
 668
 669         gfp |= __GFP_COLD;
 670         do {
 671                 if (vi->mergeable_rx_bufs)
 672                         err = add_recvbuf_mergeable(rq, gfp);
 673                 else if (vi->big_packets)
 674                         err = add_recvbuf_big(vi, rq, gfp);
 675                 else
 676                         err = add_recvbuf_small(vi, rq, gfp);
 677
 678                 oom = err == -ENOMEM;
 679                 if (err)
 680                         break;
 681         } while (rq->vq->num_free);
 682         virtqueue_kick(rq->vq);
 683         return !oom;
 684 }
 685
 686 static void skb_recv_done(struct virtqueue *rvq)
 687 {
 688         struct virtnet_info *vi = rvq->vdev->priv;
 689         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 690
 691         /* Schedule NAPI, Suppress further interrupts if successful. */
 692         if (napi_schedule_prep(&rq->napi)) {
 693                 virtqueue_disable_cb(rvq);
 694                 __napi_schedule(&rq->napi);
 695         }
 696 }
 697
 698 static void virtnet_napi_enable(struct receive_queue *rq)
 699 {
 700         napi_enable(&rq->napi);
 701
 702         /* If all buffers were filled by other side before we napi_enabled, we
 703          * won't get another interrupt, so process any outstanding packets
 704          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 705          * We synchronize against interrupts via NAPI_STATE_SCHED */
 706         if (napi_schedule_prep(&rq->napi)) {
 707                 virtqueue_disable_cb(rq->vq);
 708                 local_bh_disable();
 709                 __napi_schedule(&rq->napi);
 710                 local_bh_enable();
 711         }
 712 }
 713
 714 static void refill_work(struct work_struct *work)
 715 {
 716         struct virtnet_info *vi =
 717                 container_of(work, struct virtnet_info, refill.work);
 718         bool still_empty;
 719         int i;
 720
 721         for (i = 0; i < vi->curr_queue_pairs; i++) {
 722                 struct receive_queue *rq = &vi->rq[i];
 723
 724                 napi_disable(&rq->napi);
 725                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
 726                 virtnet_napi_enable(rq);
 727
 728                 /* In theory, this can happen: if we don't get any buffers in
 729                  * we will *never* try to fill again.
 730                  */
 731                 if (still_empty)
 732                         schedule_delayed_work(&vi->refill, HZ/2);
 733         }
 734 }
 735
 736 static int virtnet_receive(struct receive_queue *rq, int budget)
 737 {
 738         struct virtnet_info *vi = rq->vq->vdev->priv;
 739         unsigned int len, received = 0;
 740         void *buf;
 741
 742         while (received < budget &&
 743                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 744                 receive_buf(vi, rq, buf, len);
 745                 received++;
 746         }
 747
 748         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 749                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 750                         schedule_delayed_work(&vi->refill, 0);
 751         }
 752
 753         return received;
 754 }
 755
 756 static int virtnet_poll(struct napi_struct *napi, int budget)
 757 {
 758         struct receive_queue *rq =
 759                 container_of(napi, struct receive_queue, napi);
 760         unsigned int r, received;
 761
 762         received = virtnet_receive(rq, budget);
 763
 764         /* Out of packets? */
 765         if (received < budget) {
 766                 r = virtqueue_enable_cb_prepare(rq->vq);
 767                 napi_complete_done(napi, received);
 768                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 769                     napi_schedule_prep(napi)) {
 770                         virtqueue_disable_cb(rq->vq);
 771                         __napi_schedule(napi);
 772                 }
 773         }
 774
 775         return received;
 776 }
 777
 778 #ifdef CONFIG_NET_RX_BUSY_POLL
 779 /* must be called with local_bh_disable()d */
 780 static int virtnet_busy_poll(struct napi_struct *napi)
 781 {
 782         struct receive_queue *rq =
 783                 container_of(napi, struct receive_queue, napi);
 784         struct virtnet_info *vi = rq->vq->vdev->priv;
 785         int r, received = 0, budget = 4;
 786
 787         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
 788                 return LL_FLUSH_FAILED;
 789
 790         if (!napi_schedule_prep(napi))
 791                 return LL_FLUSH_BUSY;
 792
 793         virtqueue_disable_cb(rq->vq);
 794
 795 again:
 796         received += virtnet_receive(rq, budget);
 797
 798         r = virtqueue_enable_cb_prepare(rq->vq);
 799         clear_bit(NAPI_STATE_SCHED, &napi->state);
 800         if (unlikely(virtqueue_poll(rq->vq, r)) &&
 801             napi_schedule_prep(napi)) {
 802                 virtqueue_disable_cb(rq->vq);
 803                 if (received < budget) {
 804                         budget -= received;
 805                         goto again;
 806                 } else {
 807                         __napi_schedule(napi);
 808                 }
 809         }
 810
 811         return received;
 812 }
 813 #endif  /* CONFIG_NET_RX_BUSY_POLL */
 814
 815 static int virtnet_open(struct net_device *dev)
 816 {
 817         struct virtnet_info *vi = netdev_priv(dev);
 818         int i;
 819
 820         for (i = 0; i < vi->max_queue_pairs; i++) {
 821                 if (i < vi->curr_queue_pairs)
 822                         /* Make sure we have some buffers: if oom use wq. */
 823                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
 824                                 schedule_delayed_work(&vi->refill, 0);
 825                 virtnet_napi_enable(&vi->rq[i]);
 826         }
 827
 828         return 0;
 829 }
 830
 831 static void free_old_xmit_skbs(struct send_queue *sq)
 832 {
 833         struct sk_buff *skb;
 834         unsigned int len;
 835         struct virtnet_info *vi = sq->vq->vdev->priv;
 836         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 837
 838         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 839                 pr_debug("Sent skb %p\n", skb);
 840
 841                 u64_stats_update_begin(&stats->tx_syncp);
 842                 stats->tx_bytes += skb->len;
 843                 stats->tx_packets++;
 844                 u64_stats_update_end(&stats->tx_syncp);
 845
 846                 dev_kfree_skb_any(skb);
 847         }
 848 }
 849
 850 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 851 {
 852         struct virtio_net_hdr_mrg_rxbuf *hdr;
 853         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 854         struct virtnet_info *vi = sq->vq->vdev->priv;
 855         unsigned num_sg;
 856         unsigned hdr_len = vi->hdr_len;
 857         bool can_push;
 858
 859         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 860
 861         can_push = vi->any_header_sg &&
 862                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 863                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 864         /* Even if we can, don't push here yet as this would skew
 865          * csum_start offset below. */
 866         if (can_push)
 867                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
 868         else
 869                 hdr = skb_vnet_hdr(skb);
 870
 871         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 872                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 873                 hdr->hdr.csum_start = cpu_to_virtio16(vi->vdev,
 874                                                 skb_checksum_start_offset(skb));
 875                 hdr->hdr.csum_offset = cpu_to_virtio16(vi->vdev,
 876                                                          skb->csum_offset);
 877         } else {
 878                 hdr->hdr.flags = 0;
 879                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 880         }
 881
 882         if (skb_is_gso(skb)) {
 883                 hdr->hdr.hdr_len = cpu_to_virtio16(vi->vdev, skb_headlen(skb));
 884                 hdr->hdr.gso_size = cpu_to_virtio16(vi->vdev,
 885                                                     skb_shinfo(skb)->gso_size);
 886                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 887                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 888                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 889                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 890                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 891                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 892                 else
 893                         BUG();
 894                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 895                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 896         } else {
 897                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 898                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 899         }
 900
 901         if (vi->mergeable_rx_bufs)
 902                 hdr->num_buffers = 0;
 903
 904         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
 905         if (can_push) {
 906                 __skb_push(skb, hdr_len);
 907                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 908                 /* Pull header back to avoid skew in tx bytes calculations. */
 909                 __skb_pull(skb, hdr_len);
 910         } else {
 911                 sg_set_buf(sq->sg, hdr, hdr_len);
 912                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 913         }
 914         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 915 }
 916
 917 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 918 {
 919         struct virtnet_info *vi = netdev_priv(dev);
 920         int qnum = skb_get_queue_mapping(skb);
 921         struct send_queue *sq = &vi->sq[qnum];
 922         int err;
 923         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
 924         bool kick = !skb->xmit_more;
 925
 926         /* Free up any pending old buffers before queueing new ones. */
 927         free_old_xmit_skbs(sq);
 928
 929         /* timestamp packet in software */
 930         skb_tx_timestamp(skb);
 931
 932         /* Try to transmit */
 933         err = xmit_skb(sq, skb);
 934
 935         /* This should not happen! */
 936         if (unlikely(err)) {
 937                 dev->stats.tx_fifo_errors++;
 938                 if (net_ratelimit())
 939                         dev_warn(&dev->dev,
 940                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 941                 dev->stats.tx_dropped++;
 942                 dev_kfree_skb_any(skb);
 943                 return NETDEV_TX_OK;
 944         }
 945
 946         /* Don't wait up for transmitted skbs to be freed. */
 947         skb_orphan(skb);
 948         nf_reset(skb);
 949
 950         /* If running out of space, stop queue to avoid getting packets that we
 951          * are then unable to transmit.
 952          * An alternative would be to force queuing layer to requeue the skb by
 953          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
 954          * returned in a normal path of operation: it means that driver is not
 955          * maintaining the TX queue stop/start state properly, and causes
 956          * the stack to do a non-trivial amount of useless work.
 957          * Since most packets only take 1 or 2 ring slots, stopping the queue
 958          * early means 16 slots are typically wasted.
 959          */
 960         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 961                 netif_stop_subqueue(dev, qnum);
 962                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 963                         /* More just got used, free them then recheck. */
 964                         free_old_xmit_skbs(sq);
 965                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 966                                 netif_start_subqueue(dev, qnum);
 967                                 virtqueue_disable_cb(sq->vq);
 968                         }
 969                 }
 970         }
 971
 972         if (kick || netif_xmit_stopped(txq))
 973                 virtqueue_kick(sq->vq);
 974
 975         return NETDEV_TX_OK;
 976 }
 977
 978 /*
 979  * Send command via the control virtqueue and check status.  Commands
 980  * supported by the hypervisor, as indicated by feature bits, should
 981  * never fail unless improperly formatted.
 982  */
 983 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 984                                  struct scatterlist *out)
 985 {
 986         struct scatterlist *sgs[4], hdr, stat;
 987         unsigned out_num = 0, tmp;
 988
 989         /* Caller should know better */
 990         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 991
 992         vi->ctrl_status = ~0;
 993         vi->ctrl_hdr.class = class;
 994         vi->ctrl_hdr.cmd = cmd;
 995         /* Add header */
 996         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
 997         sgs[out_num++] = &hdr;
 998
 999         if (out)
1000                 sgs[out_num++] = out;
1001
1002         /* Add return status. */
1003         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1004         sgs[out_num] = &stat;
1005
1006         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1007         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1008
1009         if (unlikely(!virtqueue_kick(vi->cvq)))
1010                 return vi->ctrl_status == VIRTIO_NET_OK;
1011
1012         /* Spin for a response, the kick causes an ioport write, trapping
1013          * into the hypervisor, so the request should be handled immediately.
1014          */
1015         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1016                !virtqueue_is_broken(vi->cvq))
1017                 cpu_relax();
1018
1019         return vi->ctrl_status == VIRTIO_NET_OK;
1020 }
1021
1022 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1023 {
1024         struct virtnet_info *vi = netdev_priv(dev);
1025         struct virtio_device *vdev = vi->vdev;
1026         int ret;
1027         struct sockaddr *addr = p;
1028         struct scatterlist sg;
1029
1030         ret = eth_prepare_mac_addr_change(dev, p);
1031         if (ret)
1032                 return ret;
1033
1034         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1035                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1036                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1037                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1038                         dev_warn(&vdev->dev,
1039                                  "Failed to set mac address by vq command.\n");
1040                         return -EINVAL;
1041                 }
1042         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1043                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1044                 unsigned int i;
1045
1046                 /* Naturally, this has an atomicity problem. */
1047                 for (i = 0; i < dev->addr_len; i++)
1048                         virtio_cwrite8(vdev,
1049                                        offsetof(struct virtio_net_config, mac) +
1050                                        i, addr->sa_data[i]);
1051         }
1052
1053         eth_commit_mac_addr_change(dev, p);
1054
1055         return 0;
1056 }
1057
1058 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1059                                                struct rtnl_link_stats64 *tot)
1060 {
1061         struct virtnet_info *vi = netdev_priv(dev);
1062         int cpu;
1063         unsigned int start;
1064
1065         for_each_possible_cpu(cpu) {
1066                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1067                 u64 tpackets, tbytes, rpackets, rbytes;
1068
1069                 do {
1070                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1071                         tpackets = stats->tx_packets;
1072                         tbytes   = stats->tx_bytes;
1073                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1074
1075                 do {
1076                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1077                         rpackets = stats->rx_packets;
1078                         rbytes   = stats->rx_bytes;
1079                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1080
1081                 tot->rx_packets += rpackets;
1082                 tot->tx_packets += tpackets;
1083                 tot->rx_bytes   += rbytes;
1084                 tot->tx_bytes   += tbytes;
1085         }
1086
1087         tot->tx_dropped = dev->stats.tx_dropped;
1088         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1089         tot->rx_dropped = dev->stats.rx_dropped;
1090         tot->rx_length_errors = dev->stats.rx_length_errors;
1091         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1092
1093         return tot;
1094 }
1095
1096 #ifdef CONFIG_NET_POLL_CONTROLLER
1097 static void virtnet_netpoll(struct net_device *dev)
1098 {
1099         struct virtnet_info *vi = netdev_priv(dev);
1100         int i;
1101
1102         for (i = 0; i < vi->curr_queue_pairs; i++)
1103                 napi_schedule(&vi->rq[i].napi);
1104 }
1105 #endif
1106
1107 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1108 {
1109         rtnl_lock();
1110         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1111                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1112                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1113         rtnl_unlock();
1114 }
1115
1116 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1117 {
1118         struct scatterlist sg;
1119         struct virtio_net_ctrl_mq s;
1120         struct net_device *dev = vi->dev;
1121
1122         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1123                 return 0;
1124
1125         s.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1126         sg_init_one(&sg, &s, sizeof(s));
1127
1128         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1129                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1130                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1131                          queue_pairs);
1132                 return -EINVAL;
1133         } else {
1134                 vi->curr_queue_pairs = queue_pairs;
1135                 /* virtnet_open() will refill when device is going to up. */
1136                 if (dev->flags & IFF_UP)
1137                         schedule_delayed_work(&vi->refill, 0);
1138         }
1139
1140         return 0;
1141 }
1142
1143 static int virtnet_close(struct net_device *dev)
1144 {
1145         struct virtnet_info *vi = netdev_priv(dev);
1146         int i;
1147
1148         /* Make sure refill_work doesn't re-enable napi! */
1149         cancel_delayed_work_sync(&vi->refill);
1150
1151         for (i = 0; i < vi->max_queue_pairs; i++)
1152                 napi_disable(&vi->rq[i].napi);
1153
1154         return 0;
1155 }
1156
1157 static void virtnet_set_rx_mode(struct net_device *dev)
1158 {
1159         struct virtnet_info *vi = netdev_priv(dev);
1160         struct scatterlist sg[2];
1161         struct virtio_net_ctrl_mac *mac_data;
1162         struct netdev_hw_addr *ha;
1163         int uc_count;
1164         int mc_count;
1165         void *buf;
1166         int i;
1167
1168         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1169         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1170                 return;
1171
1172         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1173         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1174
1175         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1176
1177         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1178                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1179                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1180                          vi->ctrl_promisc ? "en" : "dis");
1181
1182         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1183
1184         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1185                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1186                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1187                          vi->ctrl_allmulti ? "en" : "dis");
1188
1189         uc_count = netdev_uc_count(dev);
1190         mc_count = netdev_mc_count(dev);
1191         /* MAC filter - use one buffer for both lists */
1192         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1193                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1194         mac_data = buf;
1195         if (!buf)
1196                 return;
1197
1198         sg_init_table(sg, 2);
1199
1200         /* Store the unicast list and count in the front of the buffer */
1201         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1202         i = 0;
1203         netdev_for_each_uc_addr(ha, dev)
1204                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1205
1206         sg_set_buf(&sg[0], mac_data,
1207                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1208
1209         /* multicast list and count fill the end */
1210         mac_data = (void *)&mac_data->macs[uc_count][0];
1211
1212         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1213         i = 0;
1214         netdev_for_each_mc_addr(ha, dev)
1215                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1216
1217         sg_set_buf(&sg[1], mac_data,
1218                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1219
1220         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1221                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1222                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1223
1224         kfree(buf);
1225 }
1226
1227 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1228                                    __be16 proto, u16 vid)
1229 {
1230         struct virtnet_info *vi = netdev_priv(dev);
1231         struct scatterlist sg;
1232
1233         sg_init_one(&sg, &vid, sizeof(vid));
1234
1235         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1236                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1237                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1238         return 0;
1239 }
1240
1241 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1242                                     __be16 proto, u16 vid)
1243 {
1244         struct virtnet_info *vi = netdev_priv(dev);
1245         struct scatterlist sg;
1246
1247         sg_init_one(&sg, &vid, sizeof(vid));
1248
1249         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1250                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1251                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1252         return 0;
1253 }
1254
1255 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1256 {
1257         int i;
1258
1259         if (vi->affinity_hint_set) {
1260                 for (i = 0; i < vi->max_queue_pairs; i++) {
1261                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1262                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1263                 }
1264
1265                 vi->affinity_hint_set = false;
1266         }
1267 }
1268
1269 static void virtnet_set_affinity(struct virtnet_info *vi)
1270 {
1271         int i;
1272         int cpu;
1273
1274         /* In multiqueue mode, when the number of cpu is equal to the number of
1275          * queue pairs, we let the queue pairs to be private to one cpu by
1276          * setting the affinity hint to eliminate the contention.
1277          */
1278         if (vi->curr_queue_pairs == 1 ||
1279             vi->max_queue_pairs != num_online_cpus()) {
1280                 virtnet_clean_affinity(vi, -1);
1281                 return;
1282         }
1283
1284         i = 0;
1285         for_each_online_cpu(cpu) {
1286                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1287                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1288                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1289                 i++;
1290         }
1291
1292         vi->affinity_hint_set = true;
1293 }
1294
1295 static int virtnet_cpu_callback(struct notifier_block *nfb,
1296                                 unsigned long action, void *hcpu)
1297 {
1298         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1299
1300         switch(action & ~CPU_TASKS_FROZEN) {
1301         case CPU_ONLINE:
1302         case CPU_DOWN_FAILED:
1303         case CPU_DEAD:
1304                 virtnet_set_affinity(vi);
1305                 break;
1306         case CPU_DOWN_PREPARE:
1307                 virtnet_clean_affinity(vi, (long)hcpu);
1308                 break;
1309         default:
1310                 break;
1311         }
1312
1313         return NOTIFY_OK;
1314 }
1315
1316 static void virtnet_get_ringparam(struct net_device *dev,
1317                                 struct ethtool_ringparam *ring)
1318 {
1319         struct virtnet_info *vi = netdev_priv(dev);
1320
1321         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1322         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1323         ring->rx_pending = ring->rx_max_pending;
1324         ring->tx_pending = ring->tx_max_pending;
1325 }
1326
1327
1328 static void virtnet_get_drvinfo(struct net_device *dev,
1329                                 struct ethtool_drvinfo *info)
1330 {
1331         struct virtnet_info *vi = netdev_priv(dev);
1332         struct virtio_device *vdev = vi->vdev;
1333
1334         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1335         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1336         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1337
1338 }
1339
1340 /* TODO: Eliminate OOO packets during switching */
1341 static int virtnet_set_channels(struct net_device *dev,
1342                                 struct ethtool_channels *channels)
1343 {
1344         struct virtnet_info *vi = netdev_priv(dev);
1345         u16 queue_pairs = channels->combined_count;
1346         int err;
1347
1348         /* We don't support separate rx/tx channels.
1349          * We don't allow setting 'other' channels.
1350          */
1351         if (channels->rx_count || channels->tx_count || channels->other_count)
1352                 return -EINVAL;
1353
1354         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1355                 return -EINVAL;
1356
1357         get_online_cpus();
1358         err = virtnet_set_queues(vi, queue_pairs);
1359         if (!err) {
1360                 netif_set_real_num_tx_queues(dev, queue_pairs);
1361                 netif_set_real_num_rx_queues(dev, queue_pairs);
1362
1363                 virtnet_set_affinity(vi);
1364         }
1365         put_online_cpus();
1366
1367         return err;
1368 }
1369
1370 static void virtnet_get_channels(struct net_device *dev,
1371                                  struct ethtool_channels *channels)
1372 {
1373         struct virtnet_info *vi = netdev_priv(dev);
1374
1375         channels->combined_count = vi->curr_queue_pairs;
1376         channels->max_combined = vi->max_queue_pairs;
1377         channels->max_other = 0;
1378         channels->rx_count = 0;
1379         channels->tx_count = 0;
1380         channels->other_count = 0;
1381 }
1382
1383 /* Check if the user is trying to change anything besides speed/duplex */
1384 static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1385 {
1386         struct ethtool_cmd diff1 = *cmd;
1387         struct ethtool_cmd diff2 = {};
1388
1389         /* advertising and cmd are usually set, ignore port because we set it */
1390         ethtool_cmd_speed_set(&diff1, 0);
1391         diff1.advertising = 0;
1392         diff1.duplex = 0;
1393         diff1.port = 0;
1394         diff1.cmd = 0;
1395
1396         return !memcmp(&diff1, &diff2, sizeof(diff1));
1397 }
1398
1399 static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1400 {
1401         struct virtnet_info *vi = netdev_priv(dev);
1402         u32 speed;
1403
1404         speed = ethtool_cmd_speed(cmd);
1405         /* don't allow custom speed and duplex */
1406         if (!ethtool_validate_speed(speed) ||
1407             !ethtool_validate_duplex(cmd->duplex) ||
1408             !virtnet_validate_ethtool_cmd(cmd))
1409                 return -EINVAL;
1410         vi->speed = speed;
1411         vi->duplex = cmd->duplex;
1412
1413         return 0;
1414 }
1415
1416 static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1417 {
1418         struct virtnet_info *vi = netdev_priv(dev);
1419
1420         ethtool_cmd_speed_set(cmd, vi->speed);
1421         cmd->duplex = vi->duplex;
1422         cmd->port = PORT_OTHER;
1423
1424         return 0;
1425 }
1426
1427 static void virtnet_init_settings(struct net_device *dev)
1428 {
1429         struct virtnet_info *vi = netdev_priv(dev);
1430
1431         vi->speed = SPEED_UNKNOWN;
1432         vi->duplex = DUPLEX_UNKNOWN;
1433 }
1434
1435 static const struct ethtool_ops virtnet_ethtool_ops = {
1436         .get_drvinfo = virtnet_get_drvinfo,
1437         .get_link = ethtool_op_get_link,
1438         .get_ringparam = virtnet_get_ringparam,
1439         .set_channels = virtnet_set_channels,
1440         .get_channels = virtnet_get_channels,
1441         .get_ts_info = ethtool_op_get_ts_info,
1442         .get_settings = virtnet_get_settings,
1443         .set_settings = virtnet_set_settings,
1444 };
1445
1446 #define MIN_MTU 68
1447 #define MAX_MTU 65535
1448
1449 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1450 {
1451         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1452                 return -EINVAL;
1453         dev->mtu = new_mtu;
1454         return 0;
1455 }
1456
1457 static const struct net_device_ops virtnet_netdev = {
1458         .ndo_open            = virtnet_open,
1459         .ndo_stop            = virtnet_close,
1460         .ndo_start_xmit      = start_xmit,
1461         .ndo_validate_addr   = eth_validate_addr,
1462         .ndo_set_mac_address = virtnet_set_mac_address,
1463         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1464         .ndo_change_mtu      = virtnet_change_mtu,
1465         .ndo_get_stats64     = virtnet_stats,
1466         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1467         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1468 #ifdef CONFIG_NET_POLL_CONTROLLER
1469         .ndo_poll_controller = virtnet_netpoll,
1470 #endif
1471 #ifdef CONFIG_NET_RX_BUSY_POLL
1472         .ndo_busy_poll          = virtnet_busy_poll,
1473 #endif
1474 };
1475
1476 static void virtnet_config_changed_work(struct work_struct *work)
1477 {
1478         struct virtnet_info *vi =
1479                 container_of(work, struct virtnet_info, config_work);
1480         u16 v;
1481
1482         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1483                                  struct virtio_net_config, status, &v) < 0)
1484                 return;
1485
1486         if (v & VIRTIO_NET_S_ANNOUNCE) {
1487                 netdev_notify_peers(vi->dev);
1488                 virtnet_ack_link_announce(vi);
1489         }
1490
1491         /* Ignore unknown (future) status bits */
1492         v &= VIRTIO_NET_S_LINK_UP;
1493
1494         if (vi->status == v)
1495                 return;
1496
1497         vi->status = v;
1498
1499         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1500                 netif_carrier_on(vi->dev);
1501                 netif_tx_wake_all_queues(vi->dev);
1502         } else {
1503                 netif_carrier_off(vi->dev);
1504                 netif_tx_stop_all_queues(vi->dev);
1505         }
1506 }
1507
1508 static void virtnet_config_changed(struct virtio_device *vdev)
1509 {
1510         struct virtnet_info *vi = vdev->priv;
1511
1512         schedule_work(&vi->config_work);
1513 }
1514
1515 static void virtnet_free_queues(struct virtnet_info *vi)
1516 {
1517         int i;
1518
1519         for (i = 0; i < vi->max_queue_pairs; i++) {
1520                 napi_hash_del(&vi->rq[i].napi);
1521                 netif_napi_del(&vi->rq[i].napi);
1522         }
1523
1524         kfree(vi->rq);
1525         kfree(vi->sq);
1526 }
1527
1528 static void free_receive_bufs(struct virtnet_info *vi)
1529 {
1530         int i;
1531
1532         for (i = 0; i < vi->max_queue_pairs; i++) {
1533                 while (vi->rq[i].pages)
1534                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1535         }
1536 }
1537
1538 static void free_receive_page_frags(struct virtnet_info *vi)
1539 {
1540         int i;
1541         for (i = 0; i < vi->max_queue_pairs; i++)
1542                 if (vi->rq[i].alloc_frag.page)
1543                         put_page(vi->rq[i].alloc_frag.page);
1544 }
1545
1546 static void free_unused_bufs(struct virtnet_info *vi)
1547 {
1548         void *buf;
1549         int i;
1550
1551         for (i = 0; i < vi->max_queue_pairs; i++) {
1552                 struct virtqueue *vq = vi->sq[i].vq;
1553                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1554                         dev_kfree_skb(buf);
1555         }
1556
1557         for (i = 0; i < vi->max_queue_pairs; i++) {
1558                 struct virtqueue *vq = vi->rq[i].vq;
1559
1560                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1561                         if (vi->mergeable_rx_bufs) {
1562                                 unsigned long ctx = (unsigned long)buf;
1563                                 void *base = mergeable_ctx_to_buf_address(ctx);
1564                                 put_page(virt_to_head_page(base));
1565                         } else if (vi->big_packets) {
1566                                 give_pages(&vi->rq[i], buf);
1567                         } else {
1568                                 dev_kfree_skb(buf);
1569                         }
1570                 }
1571         }
1572 }
1573
1574 static void virtnet_del_vqs(struct virtnet_info *vi)
1575 {
1576         struct virtio_device *vdev = vi->vdev;
1577
1578         virtnet_clean_affinity(vi, -1);
1579
1580         vdev->config->del_vqs(vdev);
1581
1582         virtnet_free_queues(vi);
1583 }
1584
1585 static int virtnet_find_vqs(struct virtnet_info *vi)
1586 {
1587         vq_callback_t **callbacks;
1588         struct virtqueue **vqs;
1589         int ret = -ENOMEM;
1590         int i, total_vqs;
1591         const char **names;
1592
1593         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1594          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1595          * possible control vq.
1596          */
1597         total_vqs = vi->max_queue_pairs * 2 +
1598                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1599
1600         /* Allocate space for find_vqs parameters */
1601         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1602         if (!vqs)
1603                 goto err_vq;
1604         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1605         if (!callbacks)
1606                 goto err_callback;
1607         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1608         if (!names)
1609                 goto err_names;
1610
1611         /* Parameters for control virtqueue, if any */
1612         if (vi->has_cvq) {
1613                 callbacks[total_vqs - 1] = NULL;
1614                 names[total_vqs - 1] = "control";
1615         }
1616
1617         /* Allocate/initialize parameters for send/receive virtqueues */
1618         for (i = 0; i < vi->max_queue_pairs; i++) {
1619                 callbacks[rxq2vq(i)] = skb_recv_done;
1620                 callbacks[txq2vq(i)] = skb_xmit_done;
1621                 sprintf(vi->rq[i].name, "input.%d", i);
1622                 sprintf(vi->sq[i].name, "output.%d", i);
1623                 names[rxq2vq(i)] = vi->rq[i].name;
1624                 names[txq2vq(i)] = vi->sq[i].name;
1625         }
1626
1627         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1628                                          names);
1629         if (ret)
1630                 goto err_find;
1631
1632         if (vi->has_cvq) {
1633                 vi->cvq = vqs[total_vqs - 1];
1634                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1635                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1636         }
1637
1638         for (i = 0; i < vi->max_queue_pairs; i++) {
1639                 vi->rq[i].vq = vqs[rxq2vq(i)];
1640                 vi->sq[i].vq = vqs[txq2vq(i)];
1641         }
1642
1643         kfree(names);
1644         kfree(callbacks);
1645         kfree(vqs);
1646
1647         return 0;
1648
1649 err_find:
1650         kfree(names);
1651 err_names:
1652         kfree(callbacks);
1653 err_callback:
1654         kfree(vqs);
1655 err_vq:
1656         return ret;
1657 }
1658
1659 static int virtnet_alloc_queues(struct virtnet_info *vi)
1660 {
1661         int i;
1662
1663         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1664         if (!vi->sq)
1665                 goto err_sq;
1666         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1667         if (!vi->rq)
1668                 goto err_rq;
1669
1670         INIT_DELAYED_WORK(&vi->refill, refill_work);
1671         for (i = 0; i < vi->max_queue_pairs; i++) {
1672                 vi->rq[i].pages = NULL;
1673                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1674                                napi_weight);
1675
1676                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1677                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
1678                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1679         }
1680
1681         return 0;
1682
1683 err_rq:
1684         kfree(vi->sq);
1685 err_sq:
1686         return -ENOMEM;
1687 }
1688
1689 static int init_vqs(struct virtnet_info *vi)
1690 {
1691         int ret;
1692
1693         /* Allocate send & receive queues */
1694         ret = virtnet_alloc_queues(vi);
1695         if (ret)
1696                 goto err;
1697
1698         ret = virtnet_find_vqs(vi);
1699         if (ret)
1700                 goto err_free;
1701
1702         get_online_cpus();
1703         virtnet_set_affinity(vi);
1704         put_online_cpus();
1705
1706         return 0;
1707
1708 err_free:
1709         virtnet_free_queues(vi);
1710 err:
1711         return ret;
1712 }
1713
1714 #ifdef CONFIG_SYSFS
1715 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1716                 struct rx_queue_attribute *attribute, char *buf)
1717 {
1718         struct virtnet_info *vi = netdev_priv(queue->dev);
1719         unsigned int queue_index = get_netdev_rx_queue_index(queue);
1720         struct ewma_pkt_len *avg;
1721
1722         BUG_ON(queue_index >= vi->max_queue_pairs);
1723         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1724         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1725 }
1726
1727 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1728         __ATTR_RO(mergeable_rx_buffer_size);
1729
1730 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1731         &mergeable_rx_buffer_size_attribute.attr,
1732         NULL
1733 };
1734
1735 static const struct attribute_group virtio_net_mrg_rx_group = {
1736         .name = "virtio_net",
1737         .attrs = virtio_net_mrg_rx_attrs
1738 };
1739 #endif
1740
1741 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1742                                     unsigned int fbit,
1743                                     const char *fname, const char *dname)
1744 {
1745         if (!virtio_has_feature(vdev, fbit))
1746                 return false;
1747
1748         dev_err(&vdev->dev, "device advertises feature %s but not %s",
1749                 fname, dname);
1750
1751         return true;
1752 }
1753
1754 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
1755         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1756
1757 static bool virtnet_validate_features(struct virtio_device *vdev)
1758 {
1759         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1760             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1761                              "VIRTIO_NET_F_CTRL_VQ") ||
1762              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1763                              "VIRTIO_NET_F_CTRL_VQ") ||
1764              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1765                              "VIRTIO_NET_F_CTRL_VQ") ||
1766              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1767              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1768                              "VIRTIO_NET_F_CTRL_VQ"))) {
1769                 return false;
1770         }
1771
1772         return true;
1773 }
1774
1775 static int virtnet_probe(struct virtio_device *vdev)
1776 {
1777         int i, err;
1778         struct net_device *dev;
1779         struct virtnet_info *vi;
1780         u16 max_queue_pairs;
1781
1782         if (!vdev->config->get) {
1783                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
1784                         __func__);
1785                 return -EINVAL;
1786         }
1787
1788         if (!virtnet_validate_features(vdev))
1789                 return -EINVAL;
1790
1791         /* Find if host supports multiqueue virtio_net device */
1792         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1793                                    struct virtio_net_config,
1794                                    max_virtqueue_pairs, &max_queue_pairs);
1795
1796         /* We need at least 2 queue's */
1797         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1798             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1799             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1800                 max_queue_pairs = 1;
1801
1802         /* Allocate ourselves a network device with room for our info */
1803         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1804         if (!dev)
1805                 return -ENOMEM;
1806
1807         /* Set up network device as normal. */
1808         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1809         dev->netdev_ops = &virtnet_netdev;
1810         dev->features = NETIF_F_HIGHDMA;
1811
1812         dev->ethtool_ops = &virtnet_ethtool_ops;
1813         SET_NETDEV_DEV(dev, &vdev->dev);
1814
1815         /* Do we support "hardware" checksums? */
1816         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1817                 /* This opens up the world of extra features. */
1818                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1819                 if (csum)
1820                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1821
1822                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1823                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1824                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1825                 }
1826                 /* Individual feature bits: what can host handle? */
1827                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1828                         dev->hw_features |= NETIF_F_TSO;
1829                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1830                         dev->hw_features |= NETIF_F_TSO6;
1831                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1832                         dev->hw_features |= NETIF_F_TSO_ECN;
1833                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1834                         dev->hw_features |= NETIF_F_UFO;
1835
1836                 dev->features |= NETIF_F_GSO_ROBUST;
1837
1838                 if (gso)
1839                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1840                 /* (!csum && gso) case will be fixed by register_netdev() */
1841         }
1842         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1843                 dev->features |= NETIF_F_RXCSUM;
1844
1845         dev->vlan_features = dev->features;
1846
1847         /* Configuration may specify what MAC to use.  Otherwise random. */
1848         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1849                 virtio_cread_bytes(vdev,
1850                                    offsetof(struct virtio_net_config, mac),
1851                                    dev->dev_addr, dev->addr_len);
1852         else
1853                 eth_hw_addr_random(dev);
1854
1855         /* Set up our device-specific information */
1856         vi = netdev_priv(dev);
1857         vi->dev = dev;
1858         vi->vdev = vdev;
1859         vdev->priv = vi;
1860         vi->stats = alloc_percpu(struct virtnet_stats);
1861         err = -ENOMEM;
1862         if (vi->stats == NULL)
1863                 goto free;
1864
1865         for_each_possible_cpu(i) {
1866                 struct virtnet_stats *virtnet_stats;
1867                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1868                 u64_stats_init(&virtnet_stats->tx_syncp);
1869                 u64_stats_init(&virtnet_stats->rx_syncp);
1870         }
1871
1872         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1873
1874         /* If we can receive ANY GSO packets, we must allocate large ones. */
1875         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1876             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1877             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1878             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1879                 vi->big_packets = true;
1880
1881         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1882                 vi->mergeable_rx_bufs = true;
1883
1884         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
1885             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1886                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1887         else
1888                 vi->hdr_len = sizeof(struct virtio_net_hdr);
1889
1890         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
1891             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1892                 vi->any_header_sg = true;
1893
1894         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1895                 vi->has_cvq = true;
1896
1897         if (vi->any_header_sg)
1898                 dev->needed_headroom = vi->hdr_len;
1899
1900         /* Use single tx/rx queue pair as default */
1901         vi->curr_queue_pairs = 1;
1902         vi->max_queue_pairs = max_queue_pairs;
1903
1904         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1905         err = init_vqs(vi);
1906         if (err)
1907                 goto free_stats;
1908
1909 #ifdef CONFIG_SYSFS
1910         if (vi->mergeable_rx_bufs)
1911                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1912 #endif
1913         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1914         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1915
1916         virtnet_init_settings(dev);
1917
1918         err = register_netdev(dev);
1919         if (err) {
1920                 pr_debug("virtio_net: registering device failed\n");
1921                 goto free_vqs;
1922         }
1923
1924         virtio_device_ready(vdev);
1925
1926         /* Last of all, set up some receive buffers. */
1927         for (i = 0; i < vi->curr_queue_pairs; i++) {
1928                 try_fill_recv(vi, &vi->rq[i], GFP_KERNEL);
1929
1930                 /* If we didn't even get one input buffer, we're useless. */
1931                 if (vi->rq[i].vq->num_free ==
1932                     virtqueue_get_vring_size(vi->rq[i].vq)) {
1933                         free_unused_bufs(vi);
1934                         err = -ENOMEM;
1935                         goto free_recv_bufs;
1936                 }
1937         }
1938
1939         vi->nb.notifier_call = &virtnet_cpu_callback;
1940         err = register_hotcpu_notifier(&vi->nb);
1941         if (err) {
1942                 pr_debug("virtio_net: registering cpu notifier failed\n");
1943                 goto free_recv_bufs;
1944         }
1945
1946         /* Assume link up if device can't report link status,
1947            otherwise get link status from config. */
1948         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1949                 netif_carrier_off(dev);
1950                 schedule_work(&vi->config_work);
1951         } else {
1952                 vi->status = VIRTIO_NET_S_LINK_UP;
1953                 netif_carrier_on(dev);
1954         }
1955
1956         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1957                  dev->name, max_queue_pairs);
1958
1959         return 0;
1960
1961 free_recv_bufs:
1962         vi->vdev->config->reset(vdev);
1963
1964         free_receive_bufs(vi);
1965         unregister_netdev(dev);
1966 free_vqs:
1967         cancel_delayed_work_sync(&vi->refill);
1968         free_receive_page_frags(vi);
1969         virtnet_del_vqs(vi);
1970 free_stats:
1971         free_percpu(vi->stats);
1972 free:
1973         free_netdev(dev);
1974         return err;
1975 }
1976
1977 static void remove_vq_common(struct virtnet_info *vi)
1978 {
1979         vi->vdev->config->reset(vi->vdev);
1980
1981         /* Free unused buffers in both send and recv, if any. */
1982         free_unused_bufs(vi);
1983
1984         free_receive_bufs(vi);
1985
1986         free_receive_page_frags(vi);
1987
1988         virtnet_del_vqs(vi);
1989 }
1990
1991 static void virtnet_remove(struct virtio_device *vdev)
1992 {
1993         struct virtnet_info *vi = vdev->priv;
1994
1995         unregister_hotcpu_notifier(&vi->nb);
1996
1997         /* Make sure no work handler is accessing the device. */
1998         flush_work(&vi->config_work);
1999
2000         unregister_netdev(vi->dev);
2001
2002         remove_vq_common(vi);
2003
2004         free_percpu(vi->stats);
2005         free_netdev(vi->dev);
2006 }
2007
2008 #ifdef CONFIG_PM_SLEEP
2009 static int virtnet_freeze(struct virtio_device *vdev)
2010 {
2011         struct virtnet_info *vi = vdev->priv;
2012         int i;
2013
2014         unregister_hotcpu_notifier(&vi->nb);
2015
2016         /* Make sure no work handler is accessing the device */
2017         flush_work(&vi->config_work);
2018
2019         netif_device_detach(vi->dev);
2020         cancel_delayed_work_sync(&vi->refill);
2021
2022         if (netif_running(vi->dev)) {
2023                 for (i = 0; i < vi->max_queue_pairs; i++)
2024                         napi_disable(&vi->rq[i].napi);
2025         }
2026
2027         remove_vq_common(vi);
2028
2029         return 0;
2030 }
2031
2032 static int virtnet_restore(struct virtio_device *vdev)
2033 {
2034         struct virtnet_info *vi = vdev->priv;
2035         int err, i;
2036
2037         err = init_vqs(vi);
2038         if (err)
2039                 return err;
2040
2041         virtio_device_ready(vdev);
2042
2043         if (netif_running(vi->dev)) {
2044                 for (i = 0; i < vi->curr_queue_pairs; i++)
2045                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2046                                 schedule_delayed_work(&vi->refill, 0);
2047
2048                 for (i = 0; i < vi->max_queue_pairs; i++)
2049                         virtnet_napi_enable(&vi->rq[i]);
2050         }
2051
2052         netif_device_attach(vi->dev);
2053
2054         rtnl_lock();
2055         virtnet_set_queues(vi, vi->curr_queue_pairs);
2056         rtnl_unlock();
2057
2058         err = register_hotcpu_notifier(&vi->nb);
2059         if (err)
2060                 return err;
2061
2062         return 0;
2063 }
2064 #endif
2065
2066 static struct virtio_device_id id_table[] = {
2067         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2068         { 0 },
2069 };
2070
2071 static unsigned int features[] = {
2072         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
2073         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
2074         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
2075         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
2076         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
2077         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
2078         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
2079         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
2080         VIRTIO_NET_F_CTRL_MAC_ADDR,
2081         VIRTIO_F_ANY_LAYOUT,
2082 };
2083
2084 static struct virtio_driver virtio_net_driver = {
2085         .feature_table = features,
2086         .feature_table_size = ARRAY_SIZE(features),
2087         .driver.name =  KBUILD_MODNAME,
2088         .driver.owner = THIS_MODULE,
2089         .id_table =     id_table,
2090         .probe =        virtnet_probe,
2091         .remove =       virtnet_remove,
2092         .config_changed = virtnet_config_changed,
2093 #ifdef CONFIG_PM_SLEEP
2094         .freeze =       virtnet_freeze,
2095         .restore =      virtnet_restore,
2096 #endif
2097 };
2098
2099 module_virtio_driver(virtio_net_driver);
2100
2101 MODULE_DEVICE_TABLE(virtio, id_table);
2102 MODULE_DESCRIPTION("Virtio network driver");
2103 MODULE_LICENSE("GPL");