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