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         /* enable config space updates */
 127         bool config_enable;
 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         /* Lock for config space updates */
 139         struct mutex config_lock;
 140
 141         /* Does the affinity hint is set for virtqueues? */
 142         bool affinity_hint_set;
 143
 144         /* CPU hot plug notifier */
 145         struct notifier_block nb;
 146 };
 147
 148 struct skb_vnet_hdr {
 149         union {
 150                 struct virtio_net_hdr hdr;
 151                 struct virtio_net_hdr_mrg_rxbuf mhdr;
 152         };
 153 };
 154
 155 struct padded_vnet_hdr {
 156         struct virtio_net_hdr hdr;
 157         /*
 158          * virtio_net_hdr should be in a separated sg buffer because of a
 159          * QEMU bug, and data sg buffer shares same page with this header sg.
 160          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
 161          */
 162         char padding[6];
 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 skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 189 {
 190         return (struct skb_vnet_hdr *)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 receive_queue *rq,
 251                                    struct page *page, unsigned int offset,
 252                                    unsigned int len, unsigned int truesize)
 253 {
 254         struct virtnet_info *vi = rq->vq->vdev->priv;
 255         struct sk_buff *skb;
 256         struct skb_vnet_hdr *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         if (vi->mergeable_rx_bufs) {
 270                 hdr_len = sizeof hdr->mhdr;
 271                 hdr_padded_len = sizeof hdr->mhdr;
 272         } else {
 273                 hdr_len = sizeof hdr->hdr;
 274                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 275         }
 276
 277         memcpy(hdr, p, hdr_len);
 278
 279         len -= hdr_len;
 280         offset += hdr_padded_len;
 281         p += hdr_padded_len;
 282
 283         copy = len;
 284         if (copy > skb_tailroom(skb))
 285                 copy = skb_tailroom(skb);
 286         memcpy(skb_put(skb, copy), p, copy);
 287
 288         len -= copy;
 289         offset += copy;
 290
 291         if (vi->mergeable_rx_bufs) {
 292                 if (len)
 293                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 294                 else
 295                         put_page(page);
 296                 return skb;
 297         }
 298
 299         /*
 300          * Verify that we can indeed put this data into a skb.
 301          * This is here to handle cases when the device erroneously
 302          * tries to receive more than is possible. This is usually
 303          * the case of a broken device.
 304          */
 305         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 306                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 307                 dev_kfree_skb(skb);
 308                 return NULL;
 309         }
 310         BUG_ON(offset >= PAGE_SIZE);
 311         while (len) {
 312                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 313                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 314                                 frag_size, truesize);
 315                 len -= frag_size;
 316                 page = (struct page *)page->private;
 317                 offset = 0;
 318         }
 319
 320         if (page)
 321                 give_pages(rq, page);
 322
 323         return skb;
 324 }
 325
 326 static struct sk_buff *receive_small(void *buf, unsigned int len)
 327 {
 328         struct sk_buff * skb = buf;
 329
 330         len -= sizeof(struct virtio_net_hdr);
 331         skb_trim(skb, len);
 332
 333         return skb;
 334 }
 335
 336 static struct sk_buff *receive_big(struct net_device *dev,
 337                                    struct receive_queue *rq,
 338                                    void *buf,
 339                                    unsigned int len)
 340 {
 341         struct page *page = buf;
 342         struct sk_buff *skb = page_to_skb(rq, page, 0, len, PAGE_SIZE);
 343
 344         if (unlikely(!skb))
 345                 goto err;
 346
 347         return skb;
 348
 349 err:
 350         dev->stats.rx_dropped++;
 351         give_pages(rq, page);
 352         return NULL;
 353 }
 354
 355 static struct sk_buff *receive_mergeable(struct net_device *dev,
 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 skb_vnet_hdr *hdr = buf;
 362         int num_buf = hdr->mhdr.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(rq, page, offset, len, truesize);
 368         struct sk_buff *curr_skb = head_skb;
 369
 370         if (unlikely(!curr_skb))
 371                 goto err_skb;
 372         while (--num_buf) {
 373                 int num_skb_frags;
 374
 375                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 376                 if (unlikely(!ctx)) {
 377                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 378                                  dev->name, num_buf, hdr->mhdr.num_buffers);
 379                         dev->stats.rx_length_errors++;
 380                         goto err_buf;
 381                 }
 382
 383                 buf = mergeable_ctx_to_buf_address(ctx);
 384                 page = virt_to_head_page(buf);
 385
 386                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 387                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 388                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 389
 390                         if (unlikely(!nskb))
 391                                 goto err_skb;
 392                         if (curr_skb == head_skb)
 393                                 skb_shinfo(curr_skb)->frag_list = nskb;
 394                         else
 395                                 curr_skb->next = nskb;
 396                         curr_skb = nskb;
 397                         head_skb->truesize += nskb->truesize;
 398                         num_skb_frags = 0;
 399                 }
 400                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 401                 if (curr_skb != head_skb) {
 402                         head_skb->data_len += len;
 403                         head_skb->len += len;
 404                         head_skb->truesize += truesize;
 405                 }
 406                 offset = buf - page_address(page);
 407                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 408                         put_page(page);
 409                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 410                                              len, truesize);
 411                 } else {
 412                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 413                                         offset, len, truesize);
 414                 }
 415         }
 416
 417         ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
 418         return head_skb;
 419
 420 err_skb:
 421         put_page(page);
 422         while (--num_buf) {
 423                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 424                 if (unlikely(!ctx)) {
 425                         pr_debug("%s: rx error: %d buffers missing\n",
 426                                  dev->name, num_buf);
 427                         dev->stats.rx_length_errors++;
 428                         break;
 429                 }
 430                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 431                 put_page(page);
 432         }
 433 err_buf:
 434         dev->stats.rx_dropped++;
 435         dev_kfree_skb(head_skb);
 436         return NULL;
 437 }
 438
 439 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
 440 {
 441         struct virtnet_info *vi = rq->vq->vdev->priv;
 442         struct net_device *dev = vi->dev;
 443         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 444         struct sk_buff *skb;
 445         struct skb_vnet_hdr *hdr;
 446
 447         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 448                 pr_debug("%s: short packet %i\n", dev->name, len);
 449                 dev->stats.rx_length_errors++;
 450                 if (vi->mergeable_rx_bufs) {
 451                         unsigned long ctx = (unsigned long)buf;
 452                         void *base = mergeable_ctx_to_buf_address(ctx);
 453                         put_page(virt_to_head_page(base));
 454                 } else if (vi->big_packets) {
 455                         give_pages(rq, buf);
 456                 } else {
 457                         dev_kfree_skb(buf);
 458                 }
 459                 return;
 460         }
 461
 462         if (vi->mergeable_rx_bufs)
 463                 skb = receive_mergeable(dev, rq, (unsigned long)buf, len);
 464         else if (vi->big_packets)
 465                 skb = receive_big(dev, rq, buf, len);
 466         else
 467                 skb = receive_small(buf, len);
 468
 469         if (unlikely(!skb))
 470                 return;
 471
 472         hdr = skb_vnet_hdr(skb);
 473
 474         u64_stats_update_begin(&stats->rx_syncp);
 475         stats->rx_bytes += skb->len;
 476         stats->rx_packets++;
 477         u64_stats_update_end(&stats->rx_syncp);
 478
 479         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 480                 pr_debug("Needs csum!\n");
 481                 if (!skb_partial_csum_set(skb,
 482                                           hdr->hdr.csum_start,
 483                                           hdr->hdr.csum_offset))
 484                         goto frame_err;
 485         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 486                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 487         }
 488
 489         skb->protocol = eth_type_trans(skb, dev);
 490         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 491                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 492
 493         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 494                 pr_debug("GSO!\n");
 495                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 496                 case VIRTIO_NET_HDR_GSO_TCPV4:
 497                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 498                         break;
 499                 case VIRTIO_NET_HDR_GSO_UDP:
 500                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 501                         break;
 502                 case VIRTIO_NET_HDR_GSO_TCPV6:
 503                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 504                         break;
 505                 default:
 506                         net_warn_ratelimited("%s: bad gso type %u.\n",
 507                                              dev->name, hdr->hdr.gso_type);
 508                         goto frame_err;
 509                 }
 510
 511                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 512                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 513
 514                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 515                 if (skb_shinfo(skb)->gso_size == 0) {
 516                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 517                         goto frame_err;
 518                 }
 519
 520                 /* Header must be checked, and gso_segs computed. */
 521                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 522                 skb_shinfo(skb)->gso_segs = 0;
 523         }
 524
 525         skb_mark_napi_id(skb, &rq->napi);
 526
 527         netif_receive_skb(skb);
 528         return;
 529
 530 frame_err:
 531         dev->stats.rx_frame_errors++;
 532         dev_kfree_skb(skb);
 533 }
 534
 535 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
 536 {
 537         struct virtnet_info *vi = rq->vq->vdev->priv;
 538         struct sk_buff *skb;
 539         struct skb_vnet_hdr *hdr;
 540         int err;
 541
 542         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 543         if (unlikely(!skb))
 544                 return -ENOMEM;
 545
 546         skb_put(skb, GOOD_PACKET_LEN);
 547
 548         hdr = skb_vnet_hdr(skb);
 549         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
 550
 551         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 552
 553         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 554         if (err < 0)
 555                 dev_kfree_skb(skb);
 556
 557         return err;
 558 }
 559
 560 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
 561 {
 562         struct page *first, *list = NULL;
 563         char *p;
 564         int i, err, offset;
 565
 566         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 567         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 568                 first = get_a_page(rq, gfp);
 569                 if (!first) {
 570                         if (list)
 571                                 give_pages(rq, list);
 572                         return -ENOMEM;
 573                 }
 574                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 575
 576                 /* chain new page in list head to match sg */
 577                 first->private = (unsigned long)list;
 578                 list = first;
 579         }
 580
 581         first = get_a_page(rq, gfp);
 582         if (!first) {
 583                 give_pages(rq, list);
 584                 return -ENOMEM;
 585         }
 586         p = page_address(first);
 587
 588         /* rq->sg[0], rq->sg[1] share the same page */
 589         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
 590         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
 591
 592         /* rq->sg[1] for data packet, from offset */
 593         offset = sizeof(struct padded_vnet_hdr);
 594         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 595
 596         /* chain first in list head */
 597         first->private = (unsigned long)list;
 598         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 599                                   first, gfp);
 600         if (err < 0)
 601                 give_pages(rq, first);
 602
 603         return err;
 604 }
 605
 606 static unsigned int get_mergeable_buf_len(struct ewma *avg_pkt_len)
 607 {
 608         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 609         unsigned int len;
 610
 611         len = hdr_len + clamp_t(unsigned int, ewma_read(avg_pkt_len),
 612                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 613         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 614 }
 615
 616 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 617 {
 618         struct page_frag *alloc_frag = &rq->alloc_frag;
 619         char *buf;
 620         unsigned long ctx;
 621         int err;
 622         unsigned int len, hole;
 623
 624         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 625         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 626                 return -ENOMEM;
 627
 628         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 629         ctx = mergeable_buf_to_ctx(buf, len);
 630         get_page(alloc_frag->page);
 631         alloc_frag->offset += len;
 632         hole = alloc_frag->size - alloc_frag->offset;
 633         if (hole < len) {
 634                 /* To avoid internal fragmentation, if there is very likely not
 635                  * enough space for another buffer, add the remaining space to
 636                  * the current buffer. This extra space is not included in
 637                  * the truesize stored in ctx.
 638                  */
 639                 len += hole;
 640                 alloc_frag->offset += hole;
 641         }
 642
 643         sg_init_one(rq->sg, buf, len);
 644         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 645         if (err < 0)
 646                 put_page(virt_to_head_page(buf));
 647
 648         return err;
 649 }
 650
 651 /*
 652  * Returns false if we couldn't fill entirely (OOM).
 653  *
 654  * Normally run in the receive path, but can also be run from ndo_open
 655  * before we're receiving packets, or from refill_work which is
 656  * careful to disable receiving (using napi_disable).
 657  */
 658 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
 659 {
 660         struct virtnet_info *vi = rq->vq->vdev->priv;
 661         int err;
 662         bool oom;
 663
 664         gfp |= __GFP_COLD;
 665         do {
 666                 if (vi->mergeable_rx_bufs)
 667                         err = add_recvbuf_mergeable(rq, gfp);
 668                 else if (vi->big_packets)
 669                         err = add_recvbuf_big(rq, gfp);
 670                 else
 671                         err = add_recvbuf_small(rq, gfp);
 672
 673                 oom = err == -ENOMEM;
 674                 if (err)
 675                         break;
 676         } while (rq->vq->num_free);
 677         virtqueue_kick(rq->vq);
 678         return !oom;
 679 }
 680
 681 static void skb_recv_done(struct virtqueue *rvq)
 682 {
 683         struct virtnet_info *vi = rvq->vdev->priv;
 684         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 685
 686         /* Schedule NAPI, Suppress further interrupts if successful. */
 687         if (napi_schedule_prep(&rq->napi)) {
 688                 virtqueue_disable_cb(rvq);
 689                 __napi_schedule(&rq->napi);
 690         }
 691 }
 692
 693 static void virtnet_napi_enable(struct receive_queue *rq)
 694 {
 695         napi_enable(&rq->napi);
 696
 697         /* If all buffers were filled by other side before we napi_enabled, we
 698          * won't get another interrupt, so process any outstanding packets
 699          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 700          * We synchronize against interrupts via NAPI_STATE_SCHED */
 701         if (napi_schedule_prep(&rq->napi)) {
 702                 virtqueue_disable_cb(rq->vq);
 703                 local_bh_disable();
 704                 __napi_schedule(&rq->napi);
 705                 local_bh_enable();
 706         }
 707 }
 708
 709 static void refill_work(struct work_struct *work)
 710 {
 711         struct virtnet_info *vi =
 712                 container_of(work, struct virtnet_info, refill.work);
 713         bool still_empty;
 714         int i;
 715
 716         for (i = 0; i < vi->curr_queue_pairs; i++) {
 717                 struct receive_queue *rq = &vi->rq[i];
 718
 719                 napi_disable(&rq->napi);
 720                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 721                 virtnet_napi_enable(rq);
 722
 723                 /* In theory, this can happen: if we don't get any buffers in
 724                  * we will *never* try to fill again.
 725                  */
 726                 if (still_empty)
 727                         schedule_delayed_work(&vi->refill, HZ/2);
 728         }
 729 }
 730
 731 static int virtnet_receive(struct receive_queue *rq, int budget)
 732 {
 733         struct virtnet_info *vi = rq->vq->vdev->priv;
 734         unsigned int len, received = 0;
 735         void *buf;
 736
 737         while (received < budget &&
 738                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 739                 receive_buf(rq, buf, len);
 740                 received++;
 741         }
 742
 743         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 744                 if (!try_fill_recv(rq, GFP_ATOMIC))
 745                         schedule_delayed_work(&vi->refill, 0);
 746         }
 747
 748         return received;
 749 }
 750
 751 static int virtnet_poll(struct napi_struct *napi, int budget)
 752 {
 753         struct receive_queue *rq =
 754                 container_of(napi, struct receive_queue, napi);
 755         unsigned int r, received = 0;
 756
 757 again:
 758         received += virtnet_receive(rq, budget - received);
 759
 760         /* Out of packets? */
 761         if (received < budget) {
 762                 r = virtqueue_enable_cb_prepare(rq->vq);
 763                 napi_complete(napi);
 764                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 765                     napi_schedule_prep(napi)) {
 766                         virtqueue_disable_cb(rq->vq);
 767                         __napi_schedule(napi);
 768                         goto again;
 769                 }
 770         }
 771
 772         return received;
 773 }
 774
 775 #ifdef CONFIG_NET_RX_BUSY_POLL
 776 /* must be called with local_bh_disable()d */
 777 static int virtnet_busy_poll(struct napi_struct *napi)
 778 {
 779         struct receive_queue *rq =
 780                 container_of(napi, struct receive_queue, napi);
 781         struct virtnet_info *vi = rq->vq->vdev->priv;
 782         int r, received = 0, budget = 4;
 783
 784         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
 785                 return LL_FLUSH_FAILED;
 786
 787         if (!napi_schedule_prep(napi))
 788                 return LL_FLUSH_BUSY;
 789
 790         virtqueue_disable_cb(rq->vq);
 791
 792 again:
 793         received += virtnet_receive(rq, budget);
 794
 795         r = virtqueue_enable_cb_prepare(rq->vq);
 796         clear_bit(NAPI_STATE_SCHED, &napi->state);
 797         if (unlikely(virtqueue_poll(rq->vq, r)) &&
 798             napi_schedule_prep(napi)) {
 799                 virtqueue_disable_cb(rq->vq);
 800                 if (received < budget) {
 801                         budget -= received;
 802                         goto again;
 803                 } else {
 804                         __napi_schedule(napi);
 805                 }
 806         }
 807
 808         return received;
 809 }
 810 #endif  /* CONFIG_NET_RX_BUSY_POLL */
 811
 812 static int virtnet_open(struct net_device *dev)
 813 {
 814         struct virtnet_info *vi = netdev_priv(dev);
 815         int i;
 816
 817         for (i = 0; i < vi->max_queue_pairs; i++) {
 818                 if (i < vi->curr_queue_pairs)
 819                         /* Make sure we have some buffers: if oom use wq. */
 820                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 821                                 schedule_delayed_work(&vi->refill, 0);
 822                 virtnet_napi_enable(&vi->rq[i]);
 823         }
 824
 825         return 0;
 826 }
 827
 828 static void free_old_xmit_skbs(struct send_queue *sq)
 829 {
 830         struct sk_buff *skb;
 831         unsigned int len;
 832         struct virtnet_info *vi = sq->vq->vdev->priv;
 833         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 834
 835         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 836                 pr_debug("Sent skb %p\n", skb);
 837
 838                 u64_stats_update_begin(&stats->tx_syncp);
 839                 stats->tx_bytes += skb->len;
 840                 stats->tx_packets++;
 841                 u64_stats_update_end(&stats->tx_syncp);
 842
 843                 dev_kfree_skb_any(skb);
 844         }
 845 }
 846
 847 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 848 {
 849         struct skb_vnet_hdr *hdr;
 850         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 851         struct virtnet_info *vi = sq->vq->vdev->priv;
 852         unsigned num_sg;
 853         unsigned hdr_len;
 854         bool can_push;
 855
 856         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 857         if (vi->mergeable_rx_bufs)
 858                 hdr_len = sizeof hdr->mhdr;
 859         else
 860                 hdr_len = sizeof hdr->hdr;
 861
 862         can_push = vi->any_header_sg &&
 863                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 864                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 865         /* Even if we can, don't push here yet as this would skew
 866          * csum_start offset below. */
 867         if (can_push)
 868                 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
 869         else
 870                 hdr = skb_vnet_hdr(skb);
 871
 872         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 873                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 874                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 875                 hdr->hdr.csum_offset = skb->csum_offset;
 876         } else {
 877                 hdr->hdr.flags = 0;
 878                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 879         }
 880
 881         if (skb_is_gso(skb)) {
 882                 hdr->hdr.hdr_len = skb_headlen(skb);
 883                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 884                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 885                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 886                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 887                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 888                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 889                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 890                 else
 891                         BUG();
 892                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 893                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 894         } else {
 895                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 896                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 897         }
 898
 899         if (vi->mergeable_rx_bufs)
 900                 hdr->mhdr.num_buffers = 0;
 901
 902         if (can_push) {
 903                 __skb_push(skb, hdr_len);
 904                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 905                 /* Pull header back to avoid skew in tx bytes calculations. */
 906                 __skb_pull(skb, hdr_len);
 907         } else {
 908                 sg_set_buf(sq->sg, hdr, hdr_len);
 909                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 910         }
 911         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 912 }
 913
 914 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 915 {
 916         struct virtnet_info *vi = netdev_priv(dev);
 917         int qnum = skb_get_queue_mapping(skb);
 918         struct send_queue *sq = &vi->sq[qnum];
 919         int err;
 920
 921         /* Free up any pending old buffers before queueing new ones. */
 922         free_old_xmit_skbs(sq);
 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         virtqueue_kick(sq->vq);
 938
 939         /* Don't wait up for transmitted skbs to be freed. */
 940         skb_orphan(skb);
 941         nf_reset(skb);
 942
 943         /* Apparently nice girls don't return TX_BUSY; stop the queue
 944          * before it gets out of hand.  Naturally, this wastes entries. */
 945         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 946                 netif_stop_subqueue(dev, qnum);
 947                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 948                         /* More just got used, free them then recheck. */
 949                         free_old_xmit_skbs(sq);
 950                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 951                                 netif_start_subqueue(dev, qnum);
 952                                 virtqueue_disable_cb(sq->vq);
 953                         }
 954                 }
 955         }
 956
 957         return NETDEV_TX_OK;
 958 }
 959
 960 /*
 961  * Send command via the control virtqueue and check status.  Commands
 962  * supported by the hypervisor, as indicated by feature bits, should
 963  * never fail unless improperly formatted.
 964  */
 965 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 966                                  struct scatterlist *out)
 967 {
 968         struct scatterlist *sgs[4], hdr, stat;
 969         struct virtio_net_ctrl_hdr ctrl;
 970         virtio_net_ctrl_ack status = ~0;
 971         unsigned out_num = 0, tmp;
 972
 973         /* Caller should know better */
 974         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 975
 976         ctrl.class = class;
 977         ctrl.cmd = cmd;
 978         /* Add header */
 979         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 980         sgs[out_num++] = &hdr;
 981
 982         if (out)
 983                 sgs[out_num++] = out;
 984
 985         /* Add return status. */
 986         sg_init_one(&stat, &status, sizeof(status));
 987         sgs[out_num] = &stat;
 988
 989         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
 990         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
 991
 992         if (unlikely(!virtqueue_kick(vi->cvq)))
 993                 return status == VIRTIO_NET_OK;
 994
 995         /* Spin for a response, the kick causes an ioport write, trapping
 996          * into the hypervisor, so the request should be handled immediately.
 997          */
 998         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
 999                !virtqueue_is_broken(vi->cvq))
1000                 cpu_relax();
1001
1002         return status == VIRTIO_NET_OK;
1003 }
1004
1005 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1006 {
1007         struct virtnet_info *vi = netdev_priv(dev);
1008         struct virtio_device *vdev = vi->vdev;
1009         int ret;
1010         struct sockaddr *addr = p;
1011         struct scatterlist sg;
1012
1013         ret = eth_prepare_mac_addr_change(dev, p);
1014         if (ret)
1015                 return ret;
1016
1017         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1018                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1019                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1020                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1021                         dev_warn(&vdev->dev,
1022                                  "Failed to set mac address by vq command.\n");
1023                         return -EINVAL;
1024                 }
1025         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
1026                 unsigned int i;
1027
1028                 /* Naturally, this has an atomicity problem. */
1029                 for (i = 0; i < dev->addr_len; i++)
1030                         virtio_cwrite8(vdev,
1031                                        offsetof(struct virtio_net_config, mac) +
1032                                        i, addr->sa_data[i]);
1033         }
1034
1035         eth_commit_mac_addr_change(dev, p);
1036
1037         return 0;
1038 }
1039
1040 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1041                                                struct rtnl_link_stats64 *tot)
1042 {
1043         struct virtnet_info *vi = netdev_priv(dev);
1044         int cpu;
1045         unsigned int start;
1046
1047         for_each_possible_cpu(cpu) {
1048                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1049                 u64 tpackets, tbytes, rpackets, rbytes;
1050
1051                 do {
1052                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1053                         tpackets = stats->tx_packets;
1054                         tbytes   = stats->tx_bytes;
1055                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1056
1057                 do {
1058                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1059                         rpackets = stats->rx_packets;
1060                         rbytes   = stats->rx_bytes;
1061                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1062
1063                 tot->rx_packets += rpackets;
1064                 tot->tx_packets += tpackets;
1065                 tot->rx_bytes   += rbytes;
1066                 tot->tx_bytes   += tbytes;
1067         }
1068
1069         tot->tx_dropped = dev->stats.tx_dropped;
1070         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1071         tot->rx_dropped = dev->stats.rx_dropped;
1072         tot->rx_length_errors = dev->stats.rx_length_errors;
1073         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1074
1075         return tot;
1076 }
1077
1078 #ifdef CONFIG_NET_POLL_CONTROLLER
1079 static void virtnet_netpoll(struct net_device *dev)
1080 {
1081         struct virtnet_info *vi = netdev_priv(dev);
1082         int i;
1083
1084         for (i = 0; i < vi->curr_queue_pairs; i++)
1085                 napi_schedule(&vi->rq[i].napi);
1086 }
1087 #endif
1088
1089 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1090 {
1091         rtnl_lock();
1092         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1093                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1094                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1095         rtnl_unlock();
1096 }
1097
1098 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1099 {
1100         struct scatterlist sg;
1101         struct virtio_net_ctrl_mq s;
1102         struct net_device *dev = vi->dev;
1103
1104         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1105                 return 0;
1106
1107         s.virtqueue_pairs = queue_pairs;
1108         sg_init_one(&sg, &s, sizeof(s));
1109
1110         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1111                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1112                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1113                          queue_pairs);
1114                 return -EINVAL;
1115         } else {
1116                 vi->curr_queue_pairs = queue_pairs;
1117                 /* virtnet_open() will refill when device is going to up. */
1118                 if (dev->flags & IFF_UP)
1119                         schedule_delayed_work(&vi->refill, 0);
1120         }
1121
1122         return 0;
1123 }
1124
1125 static int virtnet_close(struct net_device *dev)
1126 {
1127         struct virtnet_info *vi = netdev_priv(dev);
1128         int i;
1129
1130         /* Make sure refill_work doesn't re-enable napi! */
1131         cancel_delayed_work_sync(&vi->refill);
1132
1133         for (i = 0; i < vi->max_queue_pairs; i++)
1134                 napi_disable(&vi->rq[i].napi);
1135
1136         return 0;
1137 }
1138
1139 static void virtnet_set_rx_mode(struct net_device *dev)
1140 {
1141         struct virtnet_info *vi = netdev_priv(dev);
1142         struct scatterlist sg[2];
1143         u8 promisc, allmulti;
1144         struct virtio_net_ctrl_mac *mac_data;
1145         struct netdev_hw_addr *ha;
1146         int uc_count;
1147         int mc_count;
1148         void *buf;
1149         int i;
1150
1151         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1152         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1153                 return;
1154
1155         promisc = ((dev->flags & IFF_PROMISC) != 0);
1156         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1157
1158         sg_init_one(sg, &promisc, sizeof(promisc));
1159
1160         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1161                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1162                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1163                          promisc ? "en" : "dis");
1164
1165         sg_init_one(sg, &allmulti, sizeof(allmulti));
1166
1167         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1168                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1169                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1170                          allmulti ? "en" : "dis");
1171
1172         uc_count = netdev_uc_count(dev);
1173         mc_count = netdev_mc_count(dev);
1174         /* MAC filter - use one buffer for both lists */
1175         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1176                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1177         mac_data = buf;
1178         if (!buf)
1179                 return;
1180
1181         sg_init_table(sg, 2);
1182
1183         /* Store the unicast list and count in the front of the buffer */
1184         mac_data->entries = uc_count;
1185         i = 0;
1186         netdev_for_each_uc_addr(ha, dev)
1187                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1188
1189         sg_set_buf(&sg[0], mac_data,
1190                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1191
1192         /* multicast list and count fill the end */
1193         mac_data = (void *)&mac_data->macs[uc_count][0];
1194
1195         mac_data->entries = mc_count;
1196         i = 0;
1197         netdev_for_each_mc_addr(ha, dev)
1198                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1199
1200         sg_set_buf(&sg[1], mac_data,
1201                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1202
1203         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1204                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1205                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1206
1207         kfree(buf);
1208 }
1209
1210 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1211                                    __be16 proto, u16 vid)
1212 {
1213         struct virtnet_info *vi = netdev_priv(dev);
1214         struct scatterlist sg;
1215
1216         sg_init_one(&sg, &vid, sizeof(vid));
1217
1218         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1219                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1220                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1221         return 0;
1222 }
1223
1224 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1225                                     __be16 proto, u16 vid)
1226 {
1227         struct virtnet_info *vi = netdev_priv(dev);
1228         struct scatterlist sg;
1229
1230         sg_init_one(&sg, &vid, sizeof(vid));
1231
1232         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1233                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1234                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1235         return 0;
1236 }
1237
1238 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1239 {
1240         int i;
1241
1242         if (vi->affinity_hint_set) {
1243                 for (i = 0; i < vi->max_queue_pairs; i++) {
1244                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1245                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1246                 }
1247
1248                 vi->affinity_hint_set = false;
1249         }
1250 }
1251
1252 static void virtnet_set_affinity(struct virtnet_info *vi)
1253 {
1254         int i;
1255         int cpu;
1256
1257         /* In multiqueue mode, when the number of cpu is equal to the number of
1258          * queue pairs, we let the queue pairs to be private to one cpu by
1259          * setting the affinity hint to eliminate the contention.
1260          */
1261         if (vi->curr_queue_pairs == 1 ||
1262             vi->max_queue_pairs != num_online_cpus()) {
1263                 virtnet_clean_affinity(vi, -1);
1264                 return;
1265         }
1266
1267         i = 0;
1268         for_each_online_cpu(cpu) {
1269                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1270                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1271                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1272                 i++;
1273         }
1274
1275         vi->affinity_hint_set = true;
1276 }
1277
1278 static int virtnet_cpu_callback(struct notifier_block *nfb,
1279                                 unsigned long action, void *hcpu)
1280 {
1281         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1282
1283         switch(action & ~CPU_TASKS_FROZEN) {
1284         case CPU_ONLINE:
1285         case CPU_DOWN_FAILED:
1286         case CPU_DEAD:
1287                 virtnet_set_affinity(vi);
1288                 break;
1289         case CPU_DOWN_PREPARE:
1290                 virtnet_clean_affinity(vi, (long)hcpu);
1291                 break;
1292         default:
1293                 break;
1294         }
1295
1296         return NOTIFY_OK;
1297 }
1298
1299 static void virtnet_get_ringparam(struct net_device *dev,
1300                                 struct ethtool_ringparam *ring)
1301 {
1302         struct virtnet_info *vi = netdev_priv(dev);
1303
1304         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1305         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1306         ring->rx_pending = ring->rx_max_pending;
1307         ring->tx_pending = ring->tx_max_pending;
1308 }
1309
1310
1311 static void virtnet_get_drvinfo(struct net_device *dev,
1312                                 struct ethtool_drvinfo *info)
1313 {
1314         struct virtnet_info *vi = netdev_priv(dev);
1315         struct virtio_device *vdev = vi->vdev;
1316
1317         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1318         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1319         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1320
1321 }
1322
1323 /* TODO: Eliminate OOO packets during switching */
1324 static int virtnet_set_channels(struct net_device *dev,
1325                                 struct ethtool_channels *channels)
1326 {
1327         struct virtnet_info *vi = netdev_priv(dev);
1328         u16 queue_pairs = channels->combined_count;
1329         int err;
1330
1331         /* We don't support separate rx/tx channels.
1332          * We don't allow setting 'other' channels.
1333          */
1334         if (channels->rx_count || channels->tx_count || channels->other_count)
1335                 return -EINVAL;
1336
1337         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1338                 return -EINVAL;
1339
1340         get_online_cpus();
1341         err = virtnet_set_queues(vi, queue_pairs);
1342         if (!err) {
1343                 netif_set_real_num_tx_queues(dev, queue_pairs);
1344                 netif_set_real_num_rx_queues(dev, queue_pairs);
1345
1346                 virtnet_set_affinity(vi);
1347         }
1348         put_online_cpus();
1349
1350         return err;
1351 }
1352
1353 static void virtnet_get_channels(struct net_device *dev,
1354                                  struct ethtool_channels *channels)
1355 {
1356         struct virtnet_info *vi = netdev_priv(dev);
1357
1358         channels->combined_count = vi->curr_queue_pairs;
1359         channels->max_combined = vi->max_queue_pairs;
1360         channels->max_other = 0;
1361         channels->rx_count = 0;
1362         channels->tx_count = 0;
1363         channels->other_count = 0;
1364 }
1365
1366 static const struct ethtool_ops virtnet_ethtool_ops = {
1367         .get_drvinfo = virtnet_get_drvinfo,
1368         .get_link = ethtool_op_get_link,
1369         .get_ringparam = virtnet_get_ringparam,
1370         .set_channels = virtnet_set_channels,
1371         .get_channels = virtnet_get_channels,
1372 };
1373
1374 #define MIN_MTU 68
1375 #define MAX_MTU 65535
1376
1377 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1378 {
1379         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1380                 return -EINVAL;
1381         dev->mtu = new_mtu;
1382         return 0;
1383 }
1384
1385 static const struct net_device_ops virtnet_netdev = {
1386         .ndo_open            = virtnet_open,
1387         .ndo_stop            = virtnet_close,
1388         .ndo_start_xmit      = start_xmit,
1389         .ndo_validate_addr   = eth_validate_addr,
1390         .ndo_set_mac_address = virtnet_set_mac_address,
1391         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1392         .ndo_change_mtu      = virtnet_change_mtu,
1393         .ndo_get_stats64     = virtnet_stats,
1394         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1395         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1396 #ifdef CONFIG_NET_POLL_CONTROLLER
1397         .ndo_poll_controller = virtnet_netpoll,
1398 #endif
1399 #ifdef CONFIG_NET_RX_BUSY_POLL
1400         .ndo_busy_poll          = virtnet_busy_poll,
1401 #endif
1402 };
1403
1404 static void virtnet_config_changed_work(struct work_struct *work)
1405 {
1406         struct virtnet_info *vi =
1407                 container_of(work, struct virtnet_info, config_work);
1408         u16 v;
1409
1410         mutex_lock(&vi->config_lock);
1411         if (!vi->config_enable)
1412                 goto done;
1413
1414         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1415                                  struct virtio_net_config, status, &v) < 0)
1416                 goto done;
1417
1418         if (v & VIRTIO_NET_S_ANNOUNCE) {
1419                 netdev_notify_peers(vi->dev);
1420                 virtnet_ack_link_announce(vi);
1421         }
1422
1423         /* Ignore unknown (future) status bits */
1424         v &= VIRTIO_NET_S_LINK_UP;
1425
1426         if (vi->status == v)
1427                 goto done;
1428
1429         vi->status = v;
1430
1431         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1432                 netif_carrier_on(vi->dev);
1433                 netif_tx_wake_all_queues(vi->dev);
1434         } else {
1435                 netif_carrier_off(vi->dev);
1436                 netif_tx_stop_all_queues(vi->dev);
1437         }
1438 done:
1439         mutex_unlock(&vi->config_lock);
1440 }
1441
1442 static void virtnet_config_changed(struct virtio_device *vdev)
1443 {
1444         struct virtnet_info *vi = vdev->priv;
1445
1446         schedule_work(&vi->config_work);
1447 }
1448
1449 static void virtnet_free_queues(struct virtnet_info *vi)
1450 {
1451         int i;
1452
1453         for (i = 0; i < vi->max_queue_pairs; i++)
1454                 netif_napi_del(&vi->rq[i].napi);
1455
1456         kfree(vi->rq);
1457         kfree(vi->sq);
1458 }
1459
1460 static void free_receive_bufs(struct virtnet_info *vi)
1461 {
1462         int i;
1463
1464         for (i = 0; i < vi->max_queue_pairs; i++) {
1465                 while (vi->rq[i].pages)
1466                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1467         }
1468 }
1469
1470 static void free_receive_page_frags(struct virtnet_info *vi)
1471 {
1472         int i;
1473         for (i = 0; i < vi->max_queue_pairs; i++)
1474                 if (vi->rq[i].alloc_frag.page)
1475                         put_page(vi->rq[i].alloc_frag.page);
1476 }
1477
1478 static void free_unused_bufs(struct virtnet_info *vi)
1479 {
1480         void *buf;
1481         int i;
1482
1483         for (i = 0; i < vi->max_queue_pairs; i++) {
1484                 struct virtqueue *vq = vi->sq[i].vq;
1485                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1486                         dev_kfree_skb(buf);
1487         }
1488
1489         for (i = 0; i < vi->max_queue_pairs; i++) {
1490                 struct virtqueue *vq = vi->rq[i].vq;
1491
1492                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1493                         if (vi->mergeable_rx_bufs) {
1494                                 unsigned long ctx = (unsigned long)buf;
1495                                 void *base = mergeable_ctx_to_buf_address(ctx);
1496                                 put_page(virt_to_head_page(base));
1497                         } else if (vi->big_packets) {
1498                                 give_pages(&vi->rq[i], buf);
1499                         } else {
1500                                 dev_kfree_skb(buf);
1501                         }
1502                 }
1503         }
1504 }
1505
1506 static void virtnet_del_vqs(struct virtnet_info *vi)
1507 {
1508         struct virtio_device *vdev = vi->vdev;
1509
1510         virtnet_clean_affinity(vi, -1);
1511
1512         vdev->config->del_vqs(vdev);
1513
1514         virtnet_free_queues(vi);
1515 }
1516
1517 static int virtnet_find_vqs(struct virtnet_info *vi)
1518 {
1519         vq_callback_t **callbacks;
1520         struct virtqueue **vqs;
1521         int ret = -ENOMEM;
1522         int i, total_vqs;
1523         const char **names;
1524
1525         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1526          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1527          * possible control vq.
1528          */
1529         total_vqs = vi->max_queue_pairs * 2 +
1530                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1531
1532         /* Allocate space for find_vqs parameters */
1533         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1534         if (!vqs)
1535                 goto err_vq;
1536         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1537         if (!callbacks)
1538                 goto err_callback;
1539         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1540         if (!names)
1541                 goto err_names;
1542
1543         /* Parameters for control virtqueue, if any */
1544         if (vi->has_cvq) {
1545                 callbacks[total_vqs - 1] = NULL;
1546                 names[total_vqs - 1] = "control";
1547         }
1548
1549         /* Allocate/initialize parameters for send/receive virtqueues */
1550         for (i = 0; i < vi->max_queue_pairs; i++) {
1551                 callbacks[rxq2vq(i)] = skb_recv_done;
1552                 callbacks[txq2vq(i)] = skb_xmit_done;
1553                 sprintf(vi->rq[i].name, "input.%d", i);
1554                 sprintf(vi->sq[i].name, "output.%d", i);
1555                 names[rxq2vq(i)] = vi->rq[i].name;
1556                 names[txq2vq(i)] = vi->sq[i].name;
1557         }
1558
1559         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1560                                          names);
1561         if (ret)
1562                 goto err_find;
1563
1564         if (vi->has_cvq) {
1565                 vi->cvq = vqs[total_vqs - 1];
1566                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1567                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1568         }
1569
1570         for (i = 0; i < vi->max_queue_pairs; i++) {
1571                 vi->rq[i].vq = vqs[rxq2vq(i)];
1572                 vi->sq[i].vq = vqs[txq2vq(i)];
1573         }
1574
1575         kfree(names);
1576         kfree(callbacks);
1577         kfree(vqs);
1578
1579         return 0;
1580
1581 err_find:
1582         kfree(names);
1583 err_names:
1584         kfree(callbacks);
1585 err_callback:
1586         kfree(vqs);
1587 err_vq:
1588         return ret;
1589 }
1590
1591 static int virtnet_alloc_queues(struct virtnet_info *vi)
1592 {
1593         int i;
1594
1595         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1596         if (!vi->sq)
1597                 goto err_sq;
1598         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1599         if (!vi->rq)
1600                 goto err_rq;
1601
1602         INIT_DELAYED_WORK(&vi->refill, refill_work);
1603         for (i = 0; i < vi->max_queue_pairs; i++) {
1604                 vi->rq[i].pages = NULL;
1605                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1606                                napi_weight);
1607                 napi_hash_add(&vi->rq[i].napi);
1608
1609                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1610                 ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
1611                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1612         }
1613
1614         return 0;
1615
1616 err_rq:
1617         kfree(vi->sq);
1618 err_sq:
1619         return -ENOMEM;
1620 }
1621
1622 static int init_vqs(struct virtnet_info *vi)
1623 {
1624         int ret;
1625
1626         /* Allocate send & receive queues */
1627         ret = virtnet_alloc_queues(vi);
1628         if (ret)
1629                 goto err;
1630
1631         ret = virtnet_find_vqs(vi);
1632         if (ret)
1633                 goto err_free;
1634
1635         get_online_cpus();
1636         virtnet_set_affinity(vi);
1637         put_online_cpus();
1638
1639         return 0;
1640
1641 err_free:
1642         virtnet_free_queues(vi);
1643 err:
1644         return ret;
1645 }
1646
1647 #ifdef CONFIG_SYSFS
1648 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1649                 struct rx_queue_attribute *attribute, char *buf)
1650 {
1651         struct virtnet_info *vi = netdev_priv(queue->dev);
1652         unsigned int queue_index = get_netdev_rx_queue_index(queue);
1653         struct ewma *avg;
1654
1655         BUG_ON(queue_index >= vi->max_queue_pairs);
1656         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1657         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1658 }
1659
1660 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1661         __ATTR_RO(mergeable_rx_buffer_size);
1662
1663 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1664         &mergeable_rx_buffer_size_attribute.attr,
1665         NULL
1666 };
1667
1668 static const struct attribute_group virtio_net_mrg_rx_group = {
1669         .name = "virtio_net",
1670         .attrs = virtio_net_mrg_rx_attrs
1671 };
1672 #endif
1673
1674 static int virtnet_probe(struct virtio_device *vdev)
1675 {
1676         int i, err;
1677         struct net_device *dev;
1678         struct virtnet_info *vi;
1679         u16 max_queue_pairs;
1680
1681         /* Find if host supports multiqueue virtio_net device */
1682         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1683                                    struct virtio_net_config,
1684                                    max_virtqueue_pairs, &max_queue_pairs);
1685
1686         /* We need at least 2 queue's */
1687         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1688             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1689             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1690                 max_queue_pairs = 1;
1691
1692         /* Allocate ourselves a network device with room for our info */
1693         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1694         if (!dev)
1695                 return -ENOMEM;
1696
1697         /* Set up network device as normal. */
1698         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1699         dev->netdev_ops = &virtnet_netdev;
1700         dev->features = NETIF_F_HIGHDMA;
1701
1702         dev->ethtool_ops = &virtnet_ethtool_ops;
1703         SET_NETDEV_DEV(dev, &vdev->dev);
1704
1705         /* Do we support "hardware" checksums? */
1706         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1707                 /* This opens up the world of extra features. */
1708                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1709                 if (csum)
1710                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1711
1712                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1713                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1714                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1715                 }
1716                 /* Individual feature bits: what can host handle? */
1717                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1718                         dev->hw_features |= NETIF_F_TSO;
1719                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1720                         dev->hw_features |= NETIF_F_TSO6;
1721                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1722                         dev->hw_features |= NETIF_F_TSO_ECN;
1723                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1724                         dev->hw_features |= NETIF_F_UFO;
1725
1726                 if (gso)
1727                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1728                 /* (!csum && gso) case will be fixed by register_netdev() */
1729         }
1730         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1731                 dev->features |= NETIF_F_RXCSUM;
1732
1733         dev->vlan_features = dev->features;
1734
1735         /* Configuration may specify what MAC to use.  Otherwise random. */
1736         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1737                 virtio_cread_bytes(vdev,
1738                                    offsetof(struct virtio_net_config, mac),
1739                                    dev->dev_addr, dev->addr_len);
1740         else
1741                 eth_hw_addr_random(dev);
1742
1743         /* Set up our device-specific information */
1744         vi = netdev_priv(dev);
1745         vi->dev = dev;
1746         vi->vdev = vdev;
1747         vdev->priv = vi;
1748         vi->stats = alloc_percpu(struct virtnet_stats);
1749         err = -ENOMEM;
1750         if (vi->stats == NULL)
1751                 goto free;
1752
1753         for_each_possible_cpu(i) {
1754                 struct virtnet_stats *virtnet_stats;
1755                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1756                 u64_stats_init(&virtnet_stats->tx_syncp);
1757                 u64_stats_init(&virtnet_stats->rx_syncp);
1758         }
1759
1760         mutex_init(&vi->config_lock);
1761         vi->config_enable = true;
1762         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1763
1764         /* If we can receive ANY GSO packets, we must allocate large ones. */
1765         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1766             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1767             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1768             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1769                 vi->big_packets = true;
1770
1771         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1772                 vi->mergeable_rx_bufs = true;
1773
1774         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1775                 vi->any_header_sg = true;
1776
1777         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1778                 vi->has_cvq = true;
1779
1780         if (vi->any_header_sg) {
1781                 if (vi->mergeable_rx_bufs)
1782                         dev->needed_headroom = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1783                 else
1784                         dev->needed_headroom = sizeof(struct virtio_net_hdr);
1785         }
1786
1787         /* Use single tx/rx queue pair as default */
1788         vi->curr_queue_pairs = 1;
1789         vi->max_queue_pairs = max_queue_pairs;
1790
1791         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1792         err = init_vqs(vi);
1793         if (err)
1794                 goto free_stats;
1795
1796 #ifdef CONFIG_SYSFS
1797         if (vi->mergeable_rx_bufs)
1798                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1799 #endif
1800         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1801         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1802
1803         err = register_netdev(dev);
1804         if (err) {
1805                 pr_debug("virtio_net: registering device failed\n");
1806                 goto free_vqs;
1807         }
1808
1809         /* Last of all, set up some receive buffers. */
1810         for (i = 0; i < vi->curr_queue_pairs; i++) {
1811                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1812
1813                 /* If we didn't even get one input buffer, we're useless. */
1814                 if (vi->rq[i].vq->num_free ==
1815                     virtqueue_get_vring_size(vi->rq[i].vq)) {
1816                         free_unused_bufs(vi);
1817                         err = -ENOMEM;
1818                         goto free_recv_bufs;
1819                 }
1820         }
1821
1822         vi->nb.notifier_call = &virtnet_cpu_callback;
1823         err = register_hotcpu_notifier(&vi->nb);
1824         if (err) {
1825                 pr_debug("virtio_net: registering cpu notifier failed\n");
1826                 goto free_recv_bufs;
1827         }
1828
1829         /* Assume link up if device can't report link status,
1830            otherwise get link status from config. */
1831         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1832                 netif_carrier_off(dev);
1833                 schedule_work(&vi->config_work);
1834         } else {
1835                 vi->status = VIRTIO_NET_S_LINK_UP;
1836                 netif_carrier_on(dev);
1837         }
1838
1839         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1840                  dev->name, max_queue_pairs);
1841
1842         return 0;
1843
1844 free_recv_bufs:
1845         free_receive_bufs(vi);
1846         unregister_netdev(dev);
1847 free_vqs:
1848         cancel_delayed_work_sync(&vi->refill);
1849         free_receive_page_frags(vi);
1850         virtnet_del_vqs(vi);
1851 free_stats:
1852         free_percpu(vi->stats);
1853 free:
1854         free_netdev(dev);
1855         return err;
1856 }
1857
1858 static void remove_vq_common(struct virtnet_info *vi)
1859 {
1860         vi->vdev->config->reset(vi->vdev);
1861
1862         /* Free unused buffers in both send and recv, if any. */
1863         free_unused_bufs(vi);
1864
1865         free_receive_bufs(vi);
1866
1867         free_receive_page_frags(vi);
1868
1869         virtnet_del_vqs(vi);
1870 }
1871
1872 static void virtnet_remove(struct virtio_device *vdev)
1873 {
1874         struct virtnet_info *vi = vdev->priv;
1875
1876         unregister_hotcpu_notifier(&vi->nb);
1877
1878         /* Prevent config work handler from accessing the device. */
1879         mutex_lock(&vi->config_lock);
1880         vi->config_enable = false;
1881         mutex_unlock(&vi->config_lock);
1882
1883         unregister_netdev(vi->dev);
1884
1885         remove_vq_common(vi);
1886
1887         flush_work(&vi->config_work);
1888
1889         free_percpu(vi->stats);
1890         free_netdev(vi->dev);
1891 }
1892
1893 #ifdef CONFIG_PM_SLEEP
1894 static int virtnet_freeze(struct virtio_device *vdev)
1895 {
1896         struct virtnet_info *vi = vdev->priv;
1897         int i;
1898
1899         unregister_hotcpu_notifier(&vi->nb);
1900
1901         /* Prevent config work handler from accessing the device */
1902         mutex_lock(&vi->config_lock);
1903         vi->config_enable = false;
1904         mutex_unlock(&vi->config_lock);
1905
1906         netif_device_detach(vi->dev);
1907         cancel_delayed_work_sync(&vi->refill);
1908
1909         if (netif_running(vi->dev)) {
1910                 for (i = 0; i < vi->max_queue_pairs; i++) {
1911                         napi_disable(&vi->rq[i].napi);
1912                         napi_hash_del(&vi->rq[i].napi);
1913                         netif_napi_del(&vi->rq[i].napi);
1914                 }
1915         }
1916
1917         remove_vq_common(vi);
1918
1919         flush_work(&vi->config_work);
1920
1921         return 0;
1922 }
1923
1924 static int virtnet_restore(struct virtio_device *vdev)
1925 {
1926         struct virtnet_info *vi = vdev->priv;
1927         int err, i;
1928
1929         err = init_vqs(vi);
1930         if (err)
1931                 return err;
1932
1933         if (netif_running(vi->dev)) {
1934                 for (i = 0; i < vi->curr_queue_pairs; i++)
1935                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1936                                 schedule_delayed_work(&vi->refill, 0);
1937
1938                 for (i = 0; i < vi->max_queue_pairs; i++)
1939                         virtnet_napi_enable(&vi->rq[i]);
1940         }
1941
1942         netif_device_attach(vi->dev);
1943
1944         mutex_lock(&vi->config_lock);
1945         vi->config_enable = true;
1946         mutex_unlock(&vi->config_lock);
1947
1948         rtnl_lock();
1949         virtnet_set_queues(vi, vi->curr_queue_pairs);
1950         rtnl_unlock();
1951
1952         err = register_hotcpu_notifier(&vi->nb);
1953         if (err)
1954                 return err;
1955
1956         return 0;
1957 }
1958 #endif
1959
1960 static struct virtio_device_id id_table[] = {
1961         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1962         { 0 },
1963 };
1964
1965 static unsigned int features[] = {
1966         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1967         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1968         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1969         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1970         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1971         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1972         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1973         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1974         VIRTIO_NET_F_CTRL_MAC_ADDR,
1975         VIRTIO_F_ANY_LAYOUT,
1976 };
1977
1978 static struct virtio_driver virtio_net_driver = {
1979         .feature_table = features,
1980         .feature_table_size = ARRAY_SIZE(features),
1981         .driver.name =  KBUILD_MODNAME,
1982         .driver.owner = THIS_MODULE,
1983         .id_table =     id_table,
1984         .probe =        virtnet_probe,
1985         .remove =       virtnet_remove,
1986         .config_changed = virtnet_config_changed,
1987 #ifdef CONFIG_PM_SLEEP
1988         .freeze =       virtnet_freeze,
1989         .restore =      virtnet_restore,
1990 #endif
1991 };
1992
1993 module_virtio_driver(virtio_net_driver);
1994
1995 MODULE_DEVICE_TABLE(virtio, id_table);
1996 MODULE_DESCRIPTION("Virtio network driver");
1997 MODULE_LICENSE("GPL");