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
 938         /* Don't wait up for transmitted skbs to be freed. */
 939         skb_orphan(skb);
 940         nf_reset(skb);
 941
 942         /* Apparently nice girls don't return TX_BUSY; stop the queue
 943          * before it gets out of hand.  Naturally, this wastes entries. */
 944         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 945                 netif_stop_subqueue(dev, qnum);
 946                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 947                         /* More just got used, free them then recheck. */
 948                         free_old_xmit_skbs(sq);
 949                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 950                                 netif_start_subqueue(dev, qnum);
 951                                 virtqueue_disable_cb(sq->vq);
 952                         }
 953                 }
 954         }
 955
 956         if (__netif_subqueue_stopped(dev, qnum) || !skb->xmit_more)
 957                 virtqueue_kick(sq->vq);
 958
 959         return NETDEV_TX_OK;
 960 }
 961
 962 /*
 963  * Send command via the control virtqueue and check status.  Commands
 964  * supported by the hypervisor, as indicated by feature bits, should
 965  * never fail unless improperly formatted.
 966  */
 967 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 968                                  struct scatterlist *out)
 969 {
 970         struct scatterlist *sgs[4], hdr, stat;
 971         struct virtio_net_ctrl_hdr ctrl;
 972         virtio_net_ctrl_ack status = ~0;
 973         unsigned out_num = 0, tmp;
 974
 975         /* Caller should know better */
 976         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 977
 978         ctrl.class = class;
 979         ctrl.cmd = cmd;
 980         /* Add header */
 981         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 982         sgs[out_num++] = &hdr;
 983
 984         if (out)
 985                 sgs[out_num++] = out;
 986
 987         /* Add return status. */
 988         sg_init_one(&stat, &status, sizeof(status));
 989         sgs[out_num] = &stat;
 990
 991         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
 992         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
 993
 994         if (unlikely(!virtqueue_kick(vi->cvq)))
 995                 return status == VIRTIO_NET_OK;
 996
 997         /* Spin for a response, the kick causes an ioport write, trapping
 998          * into the hypervisor, so the request should be handled immediately.
 999          */
1000         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1001                !virtqueue_is_broken(vi->cvq))
1002                 cpu_relax();
1003
1004         return status == VIRTIO_NET_OK;
1005 }
1006
1007 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1008 {
1009         struct virtnet_info *vi = netdev_priv(dev);
1010         struct virtio_device *vdev = vi->vdev;
1011         int ret;
1012         struct sockaddr *addr = p;
1013         struct scatterlist sg;
1014
1015         ret = eth_prepare_mac_addr_change(dev, p);
1016         if (ret)
1017                 return ret;
1018
1019         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1020                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1021                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1022                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1023                         dev_warn(&vdev->dev,
1024                                  "Failed to set mac address by vq command.\n");
1025                         return -EINVAL;
1026                 }
1027         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
1028                 unsigned int i;
1029
1030                 /* Naturally, this has an atomicity problem. */
1031                 for (i = 0; i < dev->addr_len; i++)
1032                         virtio_cwrite8(vdev,
1033                                        offsetof(struct virtio_net_config, mac) +
1034                                        i, addr->sa_data[i]);
1035         }
1036
1037         eth_commit_mac_addr_change(dev, p);
1038
1039         return 0;
1040 }
1041
1042 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1043                                                struct rtnl_link_stats64 *tot)
1044 {
1045         struct virtnet_info *vi = netdev_priv(dev);
1046         int cpu;
1047         unsigned int start;
1048
1049         for_each_possible_cpu(cpu) {
1050                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1051                 u64 tpackets, tbytes, rpackets, rbytes;
1052
1053                 do {
1054                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1055                         tpackets = stats->tx_packets;
1056                         tbytes   = stats->tx_bytes;
1057                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1058
1059                 do {
1060                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1061                         rpackets = stats->rx_packets;
1062                         rbytes   = stats->rx_bytes;
1063                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1064
1065                 tot->rx_packets += rpackets;
1066                 tot->tx_packets += tpackets;
1067                 tot->rx_bytes   += rbytes;
1068                 tot->tx_bytes   += tbytes;
1069         }
1070
1071         tot->tx_dropped = dev->stats.tx_dropped;
1072         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1073         tot->rx_dropped = dev->stats.rx_dropped;
1074         tot->rx_length_errors = dev->stats.rx_length_errors;
1075         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1076
1077         return tot;
1078 }
1079
1080 #ifdef CONFIG_NET_POLL_CONTROLLER
1081 static void virtnet_netpoll(struct net_device *dev)
1082 {
1083         struct virtnet_info *vi = netdev_priv(dev);
1084         int i;
1085
1086         for (i = 0; i < vi->curr_queue_pairs; i++)
1087                 napi_schedule(&vi->rq[i].napi);
1088 }
1089 #endif
1090
1091 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1092 {
1093         rtnl_lock();
1094         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1095                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1096                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1097         rtnl_unlock();
1098 }
1099
1100 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1101 {
1102         struct scatterlist sg;
1103         struct virtio_net_ctrl_mq s;
1104         struct net_device *dev = vi->dev;
1105
1106         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1107                 return 0;
1108
1109         s.virtqueue_pairs = queue_pairs;
1110         sg_init_one(&sg, &s, sizeof(s));
1111
1112         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1113                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1114                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1115                          queue_pairs);
1116                 return -EINVAL;
1117         } else {
1118                 vi->curr_queue_pairs = queue_pairs;
1119                 /* virtnet_open() will refill when device is going to up. */
1120                 if (dev->flags & IFF_UP)
1121                         schedule_delayed_work(&vi->refill, 0);
1122         }
1123
1124         return 0;
1125 }
1126
1127 static int virtnet_close(struct net_device *dev)
1128 {
1129         struct virtnet_info *vi = netdev_priv(dev);
1130         int i;
1131
1132         /* Make sure refill_work doesn't re-enable napi! */
1133         cancel_delayed_work_sync(&vi->refill);
1134
1135         for (i = 0; i < vi->max_queue_pairs; i++)
1136                 napi_disable(&vi->rq[i].napi);
1137
1138         return 0;
1139 }
1140
1141 static void virtnet_set_rx_mode(struct net_device *dev)
1142 {
1143         struct virtnet_info *vi = netdev_priv(dev);
1144         struct scatterlist sg[2];
1145         u8 promisc, allmulti;
1146         struct virtio_net_ctrl_mac *mac_data;
1147         struct netdev_hw_addr *ha;
1148         int uc_count;
1149         int mc_count;
1150         void *buf;
1151         int i;
1152
1153         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1154         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1155                 return;
1156
1157         promisc = ((dev->flags & IFF_PROMISC) != 0);
1158         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1159
1160         sg_init_one(sg, &promisc, sizeof(promisc));
1161
1162         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1163                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1164                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1165                          promisc ? "en" : "dis");
1166
1167         sg_init_one(sg, &allmulti, sizeof(allmulti));
1168
1169         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1170                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1171                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1172                          allmulti ? "en" : "dis");
1173
1174         uc_count = netdev_uc_count(dev);
1175         mc_count = netdev_mc_count(dev);
1176         /* MAC filter - use one buffer for both lists */
1177         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1178                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1179         mac_data = buf;
1180         if (!buf)
1181                 return;
1182
1183         sg_init_table(sg, 2);
1184
1185         /* Store the unicast list and count in the front of the buffer */
1186         mac_data->entries = uc_count;
1187         i = 0;
1188         netdev_for_each_uc_addr(ha, dev)
1189                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1190
1191         sg_set_buf(&sg[0], mac_data,
1192                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1193
1194         /* multicast list and count fill the end */
1195         mac_data = (void *)&mac_data->macs[uc_count][0];
1196
1197         mac_data->entries = mc_count;
1198         i = 0;
1199         netdev_for_each_mc_addr(ha, dev)
1200                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1201
1202         sg_set_buf(&sg[1], mac_data,
1203                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1204
1205         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1206                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1207                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1208
1209         kfree(buf);
1210 }
1211
1212 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1213                                    __be16 proto, u16 vid)
1214 {
1215         struct virtnet_info *vi = netdev_priv(dev);
1216         struct scatterlist sg;
1217
1218         sg_init_one(&sg, &vid, sizeof(vid));
1219
1220         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1221                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1222                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1223         return 0;
1224 }
1225
1226 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1227                                     __be16 proto, u16 vid)
1228 {
1229         struct virtnet_info *vi = netdev_priv(dev);
1230         struct scatterlist sg;
1231
1232         sg_init_one(&sg, &vid, sizeof(vid));
1233
1234         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1235                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1236                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1237         return 0;
1238 }
1239
1240 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1241 {
1242         int i;
1243
1244         if (vi->affinity_hint_set) {
1245                 for (i = 0; i < vi->max_queue_pairs; i++) {
1246                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1247                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1248                 }
1249
1250                 vi->affinity_hint_set = false;
1251         }
1252 }
1253
1254 static void virtnet_set_affinity(struct virtnet_info *vi)
1255 {
1256         int i;
1257         int cpu;
1258
1259         /* In multiqueue mode, when the number of cpu is equal to the number of
1260          * queue pairs, we let the queue pairs to be private to one cpu by
1261          * setting the affinity hint to eliminate the contention.
1262          */
1263         if (vi->curr_queue_pairs == 1 ||
1264             vi->max_queue_pairs != num_online_cpus()) {
1265                 virtnet_clean_affinity(vi, -1);
1266                 return;
1267         }
1268
1269         i = 0;
1270         for_each_online_cpu(cpu) {
1271                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1272                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1273                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1274                 i++;
1275         }
1276
1277         vi->affinity_hint_set = true;
1278 }
1279
1280 static int virtnet_cpu_callback(struct notifier_block *nfb,
1281                                 unsigned long action, void *hcpu)
1282 {
1283         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1284
1285         switch(action & ~CPU_TASKS_FROZEN) {
1286         case CPU_ONLINE:
1287         case CPU_DOWN_FAILED:
1288         case CPU_DEAD:
1289                 virtnet_set_affinity(vi);
1290                 break;
1291         case CPU_DOWN_PREPARE:
1292                 virtnet_clean_affinity(vi, (long)hcpu);
1293                 break;
1294         default:
1295                 break;
1296         }
1297
1298         return NOTIFY_OK;
1299 }
1300
1301 static void virtnet_get_ringparam(struct net_device *dev,
1302                                 struct ethtool_ringparam *ring)
1303 {
1304         struct virtnet_info *vi = netdev_priv(dev);
1305
1306         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1307         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1308         ring->rx_pending = ring->rx_max_pending;
1309         ring->tx_pending = ring->tx_max_pending;
1310 }
1311
1312
1313 static void virtnet_get_drvinfo(struct net_device *dev,
1314                                 struct ethtool_drvinfo *info)
1315 {
1316         struct virtnet_info *vi = netdev_priv(dev);
1317         struct virtio_device *vdev = vi->vdev;
1318
1319         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1320         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1321         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1322
1323 }
1324
1325 /* TODO: Eliminate OOO packets during switching */
1326 static int virtnet_set_channels(struct net_device *dev,
1327                                 struct ethtool_channels *channels)
1328 {
1329         struct virtnet_info *vi = netdev_priv(dev);
1330         u16 queue_pairs = channels->combined_count;
1331         int err;
1332
1333         /* We don't support separate rx/tx channels.
1334          * We don't allow setting 'other' channels.
1335          */
1336         if (channels->rx_count || channels->tx_count || channels->other_count)
1337                 return -EINVAL;
1338
1339         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1340                 return -EINVAL;
1341
1342         get_online_cpus();
1343         err = virtnet_set_queues(vi, queue_pairs);
1344         if (!err) {
1345                 netif_set_real_num_tx_queues(dev, queue_pairs);
1346                 netif_set_real_num_rx_queues(dev, queue_pairs);
1347
1348                 virtnet_set_affinity(vi);
1349         }
1350         put_online_cpus();
1351
1352         return err;
1353 }
1354
1355 static void virtnet_get_channels(struct net_device *dev,
1356                                  struct ethtool_channels *channels)
1357 {
1358         struct virtnet_info *vi = netdev_priv(dev);
1359
1360         channels->combined_count = vi->curr_queue_pairs;
1361         channels->max_combined = vi->max_queue_pairs;
1362         channels->max_other = 0;
1363         channels->rx_count = 0;
1364         channels->tx_count = 0;
1365         channels->other_count = 0;
1366 }
1367
1368 static const struct ethtool_ops virtnet_ethtool_ops = {
1369         .get_drvinfo = virtnet_get_drvinfo,
1370         .get_link = ethtool_op_get_link,
1371         .get_ringparam = virtnet_get_ringparam,
1372         .set_channels = virtnet_set_channels,
1373         .get_channels = virtnet_get_channels,
1374 };
1375
1376 #define MIN_MTU 68
1377 #define MAX_MTU 65535
1378
1379 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1380 {
1381         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1382                 return -EINVAL;
1383         dev->mtu = new_mtu;
1384         return 0;
1385 }
1386
1387 static const struct net_device_ops virtnet_netdev = {
1388         .ndo_open            = virtnet_open,
1389         .ndo_stop            = virtnet_close,
1390         .ndo_start_xmit      = start_xmit,
1391         .ndo_validate_addr   = eth_validate_addr,
1392         .ndo_set_mac_address = virtnet_set_mac_address,
1393         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1394         .ndo_change_mtu      = virtnet_change_mtu,
1395         .ndo_get_stats64     = virtnet_stats,
1396         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1397         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1398 #ifdef CONFIG_NET_POLL_CONTROLLER
1399         .ndo_poll_controller = virtnet_netpoll,
1400 #endif
1401 #ifdef CONFIG_NET_RX_BUSY_POLL
1402         .ndo_busy_poll          = virtnet_busy_poll,
1403 #endif
1404 };
1405
1406 static void virtnet_config_changed_work(struct work_struct *work)
1407 {
1408         struct virtnet_info *vi =
1409                 container_of(work, struct virtnet_info, config_work);
1410         u16 v;
1411
1412         mutex_lock(&vi->config_lock);
1413         if (!vi->config_enable)
1414                 goto done;
1415
1416         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1417                                  struct virtio_net_config, status, &v) < 0)
1418                 goto done;
1419
1420         if (v & VIRTIO_NET_S_ANNOUNCE) {
1421                 netdev_notify_peers(vi->dev);
1422                 virtnet_ack_link_announce(vi);
1423         }
1424
1425         /* Ignore unknown (future) status bits */
1426         v &= VIRTIO_NET_S_LINK_UP;
1427
1428         if (vi->status == v)
1429                 goto done;
1430
1431         vi->status = v;
1432
1433         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1434                 netif_carrier_on(vi->dev);
1435                 netif_tx_wake_all_queues(vi->dev);
1436         } else {
1437                 netif_carrier_off(vi->dev);
1438                 netif_tx_stop_all_queues(vi->dev);
1439         }
1440 done:
1441         mutex_unlock(&vi->config_lock);
1442 }
1443
1444 static void virtnet_config_changed(struct virtio_device *vdev)
1445 {
1446         struct virtnet_info *vi = vdev->priv;
1447
1448         schedule_work(&vi->config_work);
1449 }
1450
1451 static void virtnet_free_queues(struct virtnet_info *vi)
1452 {
1453         int i;
1454
1455         for (i = 0; i < vi->max_queue_pairs; i++)
1456                 netif_napi_del(&vi->rq[i].napi);
1457
1458         kfree(vi->rq);
1459         kfree(vi->sq);
1460 }
1461
1462 static void free_receive_bufs(struct virtnet_info *vi)
1463 {
1464         int i;
1465
1466         for (i = 0; i < vi->max_queue_pairs; i++) {
1467                 while (vi->rq[i].pages)
1468                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1469         }
1470 }
1471
1472 static void free_receive_page_frags(struct virtnet_info *vi)
1473 {
1474         int i;
1475         for (i = 0; i < vi->max_queue_pairs; i++)
1476                 if (vi->rq[i].alloc_frag.page)
1477                         put_page(vi->rq[i].alloc_frag.page);
1478 }
1479
1480 static void free_unused_bufs(struct virtnet_info *vi)
1481 {
1482         void *buf;
1483         int i;
1484
1485         for (i = 0; i < vi->max_queue_pairs; i++) {
1486                 struct virtqueue *vq = vi->sq[i].vq;
1487                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1488                         dev_kfree_skb(buf);
1489         }
1490
1491         for (i = 0; i < vi->max_queue_pairs; i++) {
1492                 struct virtqueue *vq = vi->rq[i].vq;
1493
1494                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1495                         if (vi->mergeable_rx_bufs) {
1496                                 unsigned long ctx = (unsigned long)buf;
1497                                 void *base = mergeable_ctx_to_buf_address(ctx);
1498                                 put_page(virt_to_head_page(base));
1499                         } else if (vi->big_packets) {
1500                                 give_pages(&vi->rq[i], buf);
1501                         } else {
1502                                 dev_kfree_skb(buf);
1503                         }
1504                 }
1505         }
1506 }
1507
1508 static void virtnet_del_vqs(struct virtnet_info *vi)
1509 {
1510         struct virtio_device *vdev = vi->vdev;
1511
1512         virtnet_clean_affinity(vi, -1);
1513
1514         vdev->config->del_vqs(vdev);
1515
1516         virtnet_free_queues(vi);
1517 }
1518
1519 static int virtnet_find_vqs(struct virtnet_info *vi)
1520 {
1521         vq_callback_t **callbacks;
1522         struct virtqueue **vqs;
1523         int ret = -ENOMEM;
1524         int i, total_vqs;
1525         const char **names;
1526
1527         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1528          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1529          * possible control vq.
1530          */
1531         total_vqs = vi->max_queue_pairs * 2 +
1532                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1533
1534         /* Allocate space for find_vqs parameters */
1535         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1536         if (!vqs)
1537                 goto err_vq;
1538         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1539         if (!callbacks)
1540                 goto err_callback;
1541         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1542         if (!names)
1543                 goto err_names;
1544
1545         /* Parameters for control virtqueue, if any */
1546         if (vi->has_cvq) {
1547                 callbacks[total_vqs - 1] = NULL;
1548                 names[total_vqs - 1] = "control";
1549         }
1550
1551         /* Allocate/initialize parameters for send/receive virtqueues */
1552         for (i = 0; i < vi->max_queue_pairs; i++) {
1553                 callbacks[rxq2vq(i)] = skb_recv_done;
1554                 callbacks[txq2vq(i)] = skb_xmit_done;
1555                 sprintf(vi->rq[i].name, "input.%d", i);
1556                 sprintf(vi->sq[i].name, "output.%d", i);
1557                 names[rxq2vq(i)] = vi->rq[i].name;
1558                 names[txq2vq(i)] = vi->sq[i].name;
1559         }
1560
1561         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1562                                          names);
1563         if (ret)
1564                 goto err_find;
1565
1566         if (vi->has_cvq) {
1567                 vi->cvq = vqs[total_vqs - 1];
1568                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1569                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1570         }
1571
1572         for (i = 0; i < vi->max_queue_pairs; i++) {
1573                 vi->rq[i].vq = vqs[rxq2vq(i)];
1574                 vi->sq[i].vq = vqs[txq2vq(i)];
1575         }
1576
1577         kfree(names);
1578         kfree(callbacks);
1579         kfree(vqs);
1580
1581         return 0;
1582
1583 err_find:
1584         kfree(names);
1585 err_names:
1586         kfree(callbacks);
1587 err_callback:
1588         kfree(vqs);
1589 err_vq:
1590         return ret;
1591 }
1592
1593 static int virtnet_alloc_queues(struct virtnet_info *vi)
1594 {
1595         int i;
1596
1597         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1598         if (!vi->sq)
1599                 goto err_sq;
1600         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1601         if (!vi->rq)
1602                 goto err_rq;
1603
1604         INIT_DELAYED_WORK(&vi->refill, refill_work);
1605         for (i = 0; i < vi->max_queue_pairs; i++) {
1606                 vi->rq[i].pages = NULL;
1607                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1608                                napi_weight);
1609                 napi_hash_add(&vi->rq[i].napi);
1610
1611                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1612                 ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
1613                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1614         }
1615
1616         return 0;
1617
1618 err_rq:
1619         kfree(vi->sq);
1620 err_sq:
1621         return -ENOMEM;
1622 }
1623
1624 static int init_vqs(struct virtnet_info *vi)
1625 {
1626         int ret;
1627
1628         /* Allocate send & receive queues */
1629         ret = virtnet_alloc_queues(vi);
1630         if (ret)
1631                 goto err;
1632
1633         ret = virtnet_find_vqs(vi);
1634         if (ret)
1635                 goto err_free;
1636
1637         get_online_cpus();
1638         virtnet_set_affinity(vi);
1639         put_online_cpus();
1640
1641         return 0;
1642
1643 err_free:
1644         virtnet_free_queues(vi);
1645 err:
1646         return ret;
1647 }
1648
1649 #ifdef CONFIG_SYSFS
1650 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1651                 struct rx_queue_attribute *attribute, char *buf)
1652 {
1653         struct virtnet_info *vi = netdev_priv(queue->dev);
1654         unsigned int queue_index = get_netdev_rx_queue_index(queue);
1655         struct ewma *avg;
1656
1657         BUG_ON(queue_index >= vi->max_queue_pairs);
1658         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1659         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1660 }
1661
1662 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1663         __ATTR_RO(mergeable_rx_buffer_size);
1664
1665 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1666         &mergeable_rx_buffer_size_attribute.attr,
1667         NULL
1668 };
1669
1670 static const struct attribute_group virtio_net_mrg_rx_group = {
1671         .name = "virtio_net",
1672         .attrs = virtio_net_mrg_rx_attrs
1673 };
1674 #endif
1675
1676 static int virtnet_probe(struct virtio_device *vdev)
1677 {
1678         int i, err;
1679         struct net_device *dev;
1680         struct virtnet_info *vi;
1681         u16 max_queue_pairs;
1682
1683         /* Find if host supports multiqueue virtio_net device */
1684         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1685                                    struct virtio_net_config,
1686                                    max_virtqueue_pairs, &max_queue_pairs);
1687
1688         /* We need at least 2 queue's */
1689         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1690             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1691             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1692                 max_queue_pairs = 1;
1693
1694         /* Allocate ourselves a network device with room for our info */
1695         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1696         if (!dev)
1697                 return -ENOMEM;
1698
1699         /* Set up network device as normal. */
1700         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1701         dev->netdev_ops = &virtnet_netdev;
1702         dev->features = NETIF_F_HIGHDMA;
1703
1704         dev->ethtool_ops = &virtnet_ethtool_ops;
1705         SET_NETDEV_DEV(dev, &vdev->dev);
1706
1707         /* Do we support "hardware" checksums? */
1708         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1709                 /* This opens up the world of extra features. */
1710                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1711                 if (csum)
1712                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1713
1714                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1715                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1716                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1717                 }
1718                 /* Individual feature bits: what can host handle? */
1719                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1720                         dev->hw_features |= NETIF_F_TSO;
1721                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1722                         dev->hw_features |= NETIF_F_TSO6;
1723                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1724                         dev->hw_features |= NETIF_F_TSO_ECN;
1725                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1726                         dev->hw_features |= NETIF_F_UFO;
1727
1728                 if (gso)
1729                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1730                 /* (!csum && gso) case will be fixed by register_netdev() */
1731         }
1732         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1733                 dev->features |= NETIF_F_RXCSUM;
1734
1735         dev->vlan_features = dev->features;
1736
1737         /* Configuration may specify what MAC to use.  Otherwise random. */
1738         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1739                 virtio_cread_bytes(vdev,
1740                                    offsetof(struct virtio_net_config, mac),
1741                                    dev->dev_addr, dev->addr_len);
1742         else
1743                 eth_hw_addr_random(dev);
1744
1745         /* Set up our device-specific information */
1746         vi = netdev_priv(dev);
1747         vi->dev = dev;
1748         vi->vdev = vdev;
1749         vdev->priv = vi;
1750         vi->stats = alloc_percpu(struct virtnet_stats);
1751         err = -ENOMEM;
1752         if (vi->stats == NULL)
1753                 goto free;
1754
1755         for_each_possible_cpu(i) {
1756                 struct virtnet_stats *virtnet_stats;
1757                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1758                 u64_stats_init(&virtnet_stats->tx_syncp);
1759                 u64_stats_init(&virtnet_stats->rx_syncp);
1760         }
1761
1762         mutex_init(&vi->config_lock);
1763         vi->config_enable = true;
1764         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1765
1766         /* If we can receive ANY GSO packets, we must allocate large ones. */
1767         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1768             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1769             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1770             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1771                 vi->big_packets = true;
1772
1773         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1774                 vi->mergeable_rx_bufs = true;
1775
1776         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1777                 vi->any_header_sg = true;
1778
1779         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1780                 vi->has_cvq = true;
1781
1782         if (vi->any_header_sg) {
1783                 if (vi->mergeable_rx_bufs)
1784                         dev->needed_headroom = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1785                 else
1786                         dev->needed_headroom = sizeof(struct virtio_net_hdr);
1787         }
1788
1789         /* Use single tx/rx queue pair as default */
1790         vi->curr_queue_pairs = 1;
1791         vi->max_queue_pairs = max_queue_pairs;
1792
1793         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1794         err = init_vqs(vi);
1795         if (err)
1796                 goto free_stats;
1797
1798 #ifdef CONFIG_SYSFS
1799         if (vi->mergeable_rx_bufs)
1800                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1801 #endif
1802         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1803         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1804
1805         err = register_netdev(dev);
1806         if (err) {
1807                 pr_debug("virtio_net: registering device failed\n");
1808                 goto free_vqs;
1809         }
1810
1811         /* Last of all, set up some receive buffers. */
1812         for (i = 0; i < vi->curr_queue_pairs; i++) {
1813                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1814
1815                 /* If we didn't even get one input buffer, we're useless. */
1816                 if (vi->rq[i].vq->num_free ==
1817                     virtqueue_get_vring_size(vi->rq[i].vq)) {
1818                         free_unused_bufs(vi);
1819                         err = -ENOMEM;
1820                         goto free_recv_bufs;
1821                 }
1822         }
1823
1824         vi->nb.notifier_call = &virtnet_cpu_callback;
1825         err = register_hotcpu_notifier(&vi->nb);
1826         if (err) {
1827                 pr_debug("virtio_net: registering cpu notifier failed\n");
1828                 goto free_recv_bufs;
1829         }
1830
1831         /* Assume link up if device can't report link status,
1832            otherwise get link status from config. */
1833         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1834                 netif_carrier_off(dev);
1835                 schedule_work(&vi->config_work);
1836         } else {
1837                 vi->status = VIRTIO_NET_S_LINK_UP;
1838                 netif_carrier_on(dev);
1839         }
1840
1841         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1842                  dev->name, max_queue_pairs);
1843
1844         return 0;
1845
1846 free_recv_bufs:
1847         free_receive_bufs(vi);
1848         unregister_netdev(dev);
1849 free_vqs:
1850         cancel_delayed_work_sync(&vi->refill);
1851         free_receive_page_frags(vi);
1852         virtnet_del_vqs(vi);
1853 free_stats:
1854         free_percpu(vi->stats);
1855 free:
1856         free_netdev(dev);
1857         return err;
1858 }
1859
1860 static void remove_vq_common(struct virtnet_info *vi)
1861 {
1862         vi->vdev->config->reset(vi->vdev);
1863
1864         /* Free unused buffers in both send and recv, if any. */
1865         free_unused_bufs(vi);
1866
1867         free_receive_bufs(vi);
1868
1869         free_receive_page_frags(vi);
1870
1871         virtnet_del_vqs(vi);
1872 }
1873
1874 static void virtnet_remove(struct virtio_device *vdev)
1875 {
1876         struct virtnet_info *vi = vdev->priv;
1877
1878         unregister_hotcpu_notifier(&vi->nb);
1879
1880         /* Prevent config work handler from accessing the device. */
1881         mutex_lock(&vi->config_lock);
1882         vi->config_enable = false;
1883         mutex_unlock(&vi->config_lock);
1884
1885         unregister_netdev(vi->dev);
1886
1887         remove_vq_common(vi);
1888
1889         flush_work(&vi->config_work);
1890
1891         free_percpu(vi->stats);
1892         free_netdev(vi->dev);
1893 }
1894
1895 #ifdef CONFIG_PM_SLEEP
1896 static int virtnet_freeze(struct virtio_device *vdev)
1897 {
1898         struct virtnet_info *vi = vdev->priv;
1899         int i;
1900
1901         unregister_hotcpu_notifier(&vi->nb);
1902
1903         /* Prevent config work handler from accessing the device */
1904         mutex_lock(&vi->config_lock);
1905         vi->config_enable = false;
1906         mutex_unlock(&vi->config_lock);
1907
1908         netif_device_detach(vi->dev);
1909         cancel_delayed_work_sync(&vi->refill);
1910
1911         if (netif_running(vi->dev)) {
1912                 for (i = 0; i < vi->max_queue_pairs; i++) {
1913                         napi_disable(&vi->rq[i].napi);
1914                         napi_hash_del(&vi->rq[i].napi);
1915                         netif_napi_del(&vi->rq[i].napi);
1916                 }
1917         }
1918
1919         remove_vq_common(vi);
1920
1921         flush_work(&vi->config_work);
1922
1923         return 0;
1924 }
1925
1926 static int virtnet_restore(struct virtio_device *vdev)
1927 {
1928         struct virtnet_info *vi = vdev->priv;
1929         int err, i;
1930
1931         err = init_vqs(vi);
1932         if (err)
1933                 return err;
1934
1935         if (netif_running(vi->dev)) {
1936                 for (i = 0; i < vi->curr_queue_pairs; i++)
1937                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1938                                 schedule_delayed_work(&vi->refill, 0);
1939
1940                 for (i = 0; i < vi->max_queue_pairs; i++)
1941                         virtnet_napi_enable(&vi->rq[i]);
1942         }
1943
1944         netif_device_attach(vi->dev);
1945
1946         mutex_lock(&vi->config_lock);
1947         vi->config_enable = true;
1948         mutex_unlock(&vi->config_lock);
1949
1950         rtnl_lock();
1951         virtnet_set_queues(vi, vi->curr_queue_pairs);
1952         rtnl_unlock();
1953
1954         err = register_hotcpu_notifier(&vi->nb);
1955         if (err)
1956                 return err;
1957
1958         return 0;
1959 }
1960 #endif
1961
1962 static struct virtio_device_id id_table[] = {
1963         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1964         { 0 },
1965 };
1966
1967 static unsigned int features[] = {
1968         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1969         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1970         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1971         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1972         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1973         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1974         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1975         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1976         VIRTIO_NET_F_CTRL_MAC_ADDR,
1977         VIRTIO_F_ANY_LAYOUT,
1978 };
1979
1980 static struct virtio_driver virtio_net_driver = {
1981         .feature_table = features,
1982         .feature_table_size = ARRAY_SIZE(features),
1983         .driver.name =  KBUILD_MODNAME,
1984         .driver.owner = THIS_MODULE,
1985         .id_table =     id_table,
1986         .probe =        virtnet_probe,
1987         .remove =       virtnet_remove,
1988         .config_changed = virtnet_config_changed,
1989 #ifdef CONFIG_PM_SLEEP
1990         .freeze =       virtnet_freeze,
1991         .restore =      virtnet_restore,
1992 #endif
1993 };
1994
1995 module_virtio_driver(virtio_net_driver);
1996
1997 MODULE_DEVICE_TABLE(virtio, id_table);
1998 MODULE_DESCRIPTION("Virtio network driver");
1999 MODULE_LICENSE("GPL");