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 hotplug instances for online & dead */
 142         struct hlist_node node;
 143         struct hlist_node node_dead;
 144
 145         /* Control VQ buffers: protected by the rtnl lock */
 146         struct virtio_net_ctrl_hdr ctrl_hdr;
 147         virtio_net_ctrl_ack ctrl_status;
 148         struct virtio_net_ctrl_mq ctrl_mq;
 149         u8 ctrl_promisc;
 150         u8 ctrl_allmulti;
 151         u16 ctrl_vid;
 152
 153         /* Ethtool settings */
 154         u8 duplex;
 155         u32 speed;
 156 };
 157
 158 struct padded_vnet_hdr {
 159         struct virtio_net_hdr_mrg_rxbuf hdr;
 160         /*
 161          * hdr is in a separate sg buffer, and data sg buffer shares same page
 162          * with this header sg. This padding makes next sg 16 byte aligned
 163          * after the header.
 164          */
 165         char padding[4];
 166 };
 167
 168 /* Converting between virtqueue no. and kernel tx/rx queue no.
 169  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 170  */
 171 static int vq2txq(struct virtqueue *vq)
 172 {
 173         return (vq->index - 1) / 2;
 174 }
 175
 176 static int txq2vq(int txq)
 177 {
 178         return txq * 2 + 1;
 179 }
 180
 181 static int vq2rxq(struct virtqueue *vq)
 182 {
 183         return vq->index / 2;
 184 }
 185
 186 static int rxq2vq(int rxq)
 187 {
 188         return rxq * 2;
 189 }
 190
 191 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 192 {
 193         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 194 }
 195
 196 /*
 197  * private is used to chain pages for big packets, put the whole
 198  * most recent used list in the beginning for reuse
 199  */
 200 static void give_pages(struct receive_queue *rq, struct page *page)
 201 {
 202         struct page *end;
 203
 204         /* Find end of list, sew whole thing into vi->rq.pages. */
 205         for (end = page; end->private; end = (struct page *)end->private);
 206         end->private = (unsigned long)rq->pages;
 207         rq->pages = page;
 208 }
 209
 210 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 211 {
 212         struct page *p = rq->pages;
 213
 214         if (p) {
 215                 rq->pages = (struct page *)p->private;
 216                 /* clear private here, it is used to chain pages */
 217                 p->private = 0;
 218         } else
 219                 p = alloc_page(gfp_mask);
 220         return p;
 221 }
 222
 223 static void skb_xmit_done(struct virtqueue *vq)
 224 {
 225         struct virtnet_info *vi = vq->vdev->priv;
 226
 227         /* Suppress further interrupts. */
 228         virtqueue_disable_cb(vq);
 229
 230         /* We were probably waiting for more output buffers. */
 231         netif_wake_subqueue(vi->dev, vq2txq(vq));
 232 }
 233
 234 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
 235 {
 236         unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
 237         return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
 238 }
 239
 240 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
 241 {
 242         return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
 243
 244 }
 245
 246 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
 247 {
 248         unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
 249         return (unsigned long)buf | (size - 1);
 250 }
 251
 252 /* Called from bottom half context */
 253 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 254                                    struct receive_queue *rq,
 255                                    struct page *page, unsigned int offset,
 256                                    unsigned int len, unsigned int truesize)
 257 {
 258         struct sk_buff *skb;
 259         struct virtio_net_hdr_mrg_rxbuf *hdr;
 260         unsigned int copy, hdr_len, hdr_padded_len;
 261         char *p;
 262
 263         p = page_address(page) + offset;
 264
 265         /* copy small packet so we can reuse these pages for small data */
 266         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 267         if (unlikely(!skb))
 268                 return NULL;
 269
 270         hdr = skb_vnet_hdr(skb);
 271
 272         hdr_len = vi->hdr_len;
 273         if (vi->mergeable_rx_bufs)
 274                 hdr_padded_len = sizeof *hdr;
 275         else
 276                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 277
 278         memcpy(hdr, p, hdr_len);
 279
 280         len -= hdr_len;
 281         offset += hdr_padded_len;
 282         p += hdr_padded_len;
 283
 284         copy = len;
 285         if (copy > skb_tailroom(skb))
 286                 copy = skb_tailroom(skb);
 287         memcpy(skb_put(skb, copy), p, copy);
 288
 289         len -= copy;
 290         offset += copy;
 291
 292         if (vi->mergeable_rx_bufs) {
 293                 if (len)
 294                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 295                 else
 296                         put_page(page);
 297                 return skb;
 298         }
 299
 300         /*
 301          * Verify that we can indeed put this data into a skb.
 302          * This is here to handle cases when the device erroneously
 303          * tries to receive more than is possible. This is usually
 304          * the case of a broken device.
 305          */
 306         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 307                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 308                 dev_kfree_skb(skb);
 309                 return NULL;
 310         }
 311         BUG_ON(offset >= PAGE_SIZE);
 312         while (len) {
 313                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 314                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 315                                 frag_size, truesize);
 316                 len -= frag_size;
 317                 page = (struct page *)page->private;
 318                 offset = 0;
 319         }
 320
 321         if (page)
 322                 give_pages(rq, page);
 323
 324         return skb;
 325 }
 326
 327 static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
 328 {
 329         struct sk_buff * skb = buf;
 330
 331         len -= vi->hdr_len;
 332         skb_trim(skb, len);
 333
 334         return skb;
 335 }
 336
 337 static struct sk_buff *receive_big(struct net_device *dev,
 338                                    struct virtnet_info *vi,
 339                                    struct receive_queue *rq,
 340                                    void *buf,
 341                                    unsigned int len)
 342 {
 343         struct page *page = buf;
 344         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 345
 346         if (unlikely(!skb))
 347                 goto err;
 348
 349         return skb;
 350
 351 err:
 352         dev->stats.rx_dropped++;
 353         give_pages(rq, page);
 354         return NULL;
 355 }
 356
 357 static struct sk_buff *receive_mergeable(struct net_device *dev,
 358                                          struct virtnet_info *vi,
 359                                          struct receive_queue *rq,
 360                                          unsigned long ctx,
 361                                          unsigned int len)
 362 {
 363         void *buf = mergeable_ctx_to_buf_address(ctx);
 364         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 365         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 366         struct page *page = virt_to_head_page(buf);
 367         int offset = buf - page_address(page);
 368         unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 369
 370         struct sk_buff *head_skb = page_to_skb(vi, rq, page, offset, len,
 371                                                truesize);
 372         struct sk_buff *curr_skb = head_skb;
 373
 374         if (unlikely(!curr_skb))
 375                 goto err_skb;
 376         while (--num_buf) {
 377                 int num_skb_frags;
 378
 379                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 380                 if (unlikely(!ctx)) {
 381                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 382                                  dev->name, num_buf,
 383                                  virtio16_to_cpu(vi->vdev,
 384                                                  hdr->num_buffers));
 385                         dev->stats.rx_length_errors++;
 386                         goto err_buf;
 387                 }
 388
 389                 buf = mergeable_ctx_to_buf_address(ctx);
 390                 page = virt_to_head_page(buf);
 391
 392                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 393                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 394                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 395
 396                         if (unlikely(!nskb))
 397                                 goto err_skb;
 398                         if (curr_skb == head_skb)
 399                                 skb_shinfo(curr_skb)->frag_list = nskb;
 400                         else
 401                                 curr_skb->next = nskb;
 402                         curr_skb = nskb;
 403                         head_skb->truesize += nskb->truesize;
 404                         num_skb_frags = 0;
 405                 }
 406                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 407                 if (curr_skb != head_skb) {
 408                         head_skb->data_len += len;
 409                         head_skb->len += len;
 410                         head_skb->truesize += truesize;
 411                 }
 412                 offset = buf - page_address(page);
 413                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 414                         put_page(page);
 415                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 416                                              len, truesize);
 417                 } else {
 418                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 419                                         offset, len, truesize);
 420                 }
 421         }
 422
 423         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 424         return head_skb;
 425
 426 err_skb:
 427         put_page(page);
 428         while (--num_buf) {
 429                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 430                 if (unlikely(!ctx)) {
 431                         pr_debug("%s: rx error: %d buffers missing\n",
 432                                  dev->name, num_buf);
 433                         dev->stats.rx_length_errors++;
 434                         break;
 435                 }
 436                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 437                 put_page(page);
 438         }
 439 err_buf:
 440         dev->stats.rx_dropped++;
 441         dev_kfree_skb(head_skb);
 442         return NULL;
 443 }
 444
 445 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 446                         void *buf, unsigned int len)
 447 {
 448         struct net_device *dev = vi->dev;
 449         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 450         struct sk_buff *skb;
 451         struct virtio_net_hdr_mrg_rxbuf *hdr;
 452
 453         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 454                 pr_debug("%s: short packet %i\n", dev->name, len);
 455                 dev->stats.rx_length_errors++;
 456                 if (vi->mergeable_rx_bufs) {
 457                         unsigned long ctx = (unsigned long)buf;
 458                         void *base = mergeable_ctx_to_buf_address(ctx);
 459                         put_page(virt_to_head_page(base));
 460                 } else if (vi->big_packets) {
 461                         give_pages(rq, buf);
 462                 } else {
 463                         dev_kfree_skb(buf);
 464                 }
 465                 return;
 466         }
 467
 468         if (vi->mergeable_rx_bufs)
 469                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
 470         else if (vi->big_packets)
 471                 skb = receive_big(dev, vi, rq, buf, len);
 472         else
 473                 skb = receive_small(vi, buf, len);
 474
 475         if (unlikely(!skb))
 476                 return;
 477
 478         hdr = skb_vnet_hdr(skb);
 479
 480         u64_stats_update_begin(&stats->rx_syncp);
 481         stats->rx_bytes += skb->len;
 482         stats->rx_packets++;
 483         u64_stats_update_end(&stats->rx_syncp);
 484
 485         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 486                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 487
 488         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 489                                   virtio_is_little_endian(vi->vdev))) {
 490                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 491                                      dev->name, hdr->hdr.gso_type,
 492                                      hdr->hdr.gso_size);
 493                 goto frame_err;
 494         }
 495
 496         skb->protocol = eth_type_trans(skb, dev);
 497         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 498                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 499
 500         napi_gro_receive(&rq->napi, skb);
 501         return;
 502
 503 frame_err:
 504         dev->stats.rx_frame_errors++;
 505         dev_kfree_skb(skb);
 506 }
 507
 508 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 509                              gfp_t gfp)
 510 {
 511         struct sk_buff *skb;
 512         struct virtio_net_hdr_mrg_rxbuf *hdr;
 513         int err;
 514
 515         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 516         if (unlikely(!skb))
 517                 return -ENOMEM;
 518
 519         skb_put(skb, GOOD_PACKET_LEN);
 520
 521         hdr = skb_vnet_hdr(skb);
 522         sg_init_table(rq->sg, 2);
 523         sg_set_buf(rq->sg, hdr, vi->hdr_len);
 524         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 525
 526         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 527         if (err < 0)
 528                 dev_kfree_skb(skb);
 529
 530         return err;
 531 }
 532
 533 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 534                            gfp_t gfp)
 535 {
 536         struct page *first, *list = NULL;
 537         char *p;
 538         int i, err, offset;
 539
 540         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 541
 542         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 543         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 544                 first = get_a_page(rq, gfp);
 545                 if (!first) {
 546                         if (list)
 547                                 give_pages(rq, list);
 548                         return -ENOMEM;
 549                 }
 550                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 551
 552                 /* chain new page in list head to match sg */
 553                 first->private = (unsigned long)list;
 554                 list = first;
 555         }
 556
 557         first = get_a_page(rq, gfp);
 558         if (!first) {
 559                 give_pages(rq, list);
 560                 return -ENOMEM;
 561         }
 562         p = page_address(first);
 563
 564         /* rq->sg[0], rq->sg[1] share the same page */
 565         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 566         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 567
 568         /* rq->sg[1] for data packet, from offset */
 569         offset = sizeof(struct padded_vnet_hdr);
 570         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 571
 572         /* chain first in list head */
 573         first->private = (unsigned long)list;
 574         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 575                                   first, gfp);
 576         if (err < 0)
 577                 give_pages(rq, first);
 578
 579         return err;
 580 }
 581
 582 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
 583 {
 584         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 585         unsigned int len;
 586
 587         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
 588                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 589         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 590 }
 591
 592 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 593 {
 594         struct page_frag *alloc_frag = &rq->alloc_frag;
 595         char *buf;
 596         unsigned long ctx;
 597         int err;
 598         unsigned int len, hole;
 599
 600         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 601         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 602                 return -ENOMEM;
 603
 604         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 605         ctx = mergeable_buf_to_ctx(buf, len);
 606         get_page(alloc_frag->page);
 607         alloc_frag->offset += len;
 608         hole = alloc_frag->size - alloc_frag->offset;
 609         if (hole < len) {
 610                 /* To avoid internal fragmentation, if there is very likely not
 611                  * enough space for another buffer, add the remaining space to
 612                  * the current buffer. This extra space is not included in
 613                  * the truesize stored in ctx.
 614                  */
 615                 len += hole;
 616                 alloc_frag->offset += hole;
 617         }
 618
 619         sg_init_one(rq->sg, buf, len);
 620         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 621         if (err < 0)
 622                 put_page(virt_to_head_page(buf));
 623
 624         return err;
 625 }
 626
 627 /*
 628  * Returns false if we couldn't fill entirely (OOM).
 629  *
 630  * Normally run in the receive path, but can also be run from ndo_open
 631  * before we're receiving packets, or from refill_work which is
 632  * careful to disable receiving (using napi_disable).
 633  */
 634 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 635                           gfp_t gfp)
 636 {
 637         int err;
 638         bool oom;
 639
 640         gfp |= __GFP_COLD;
 641         do {
 642                 if (vi->mergeable_rx_bufs)
 643                         err = add_recvbuf_mergeable(rq, gfp);
 644                 else if (vi->big_packets)
 645                         err = add_recvbuf_big(vi, rq, gfp);
 646                 else
 647                         err = add_recvbuf_small(vi, rq, gfp);
 648
 649                 oom = err == -ENOMEM;
 650                 if (err)
 651                         break;
 652         } while (rq->vq->num_free);
 653         virtqueue_kick(rq->vq);
 654         return !oom;
 655 }
 656
 657 static void skb_recv_done(struct virtqueue *rvq)
 658 {
 659         struct virtnet_info *vi = rvq->vdev->priv;
 660         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 661
 662         /* Schedule NAPI, Suppress further interrupts if successful. */
 663         if (napi_schedule_prep(&rq->napi)) {
 664                 virtqueue_disable_cb(rvq);
 665                 __napi_schedule(&rq->napi);
 666         }
 667 }
 668
 669 static void virtnet_napi_enable(struct receive_queue *rq)
 670 {
 671         napi_enable(&rq->napi);
 672
 673         /* If all buffers were filled by other side before we napi_enabled, we
 674          * won't get another interrupt, so process any outstanding packets
 675          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 676          * We synchronize against interrupts via NAPI_STATE_SCHED */
 677         if (napi_schedule_prep(&rq->napi)) {
 678                 virtqueue_disable_cb(rq->vq);
 679                 local_bh_disable();
 680                 __napi_schedule(&rq->napi);
 681                 local_bh_enable();
 682         }
 683 }
 684
 685 static void refill_work(struct work_struct *work)
 686 {
 687         struct virtnet_info *vi =
 688                 container_of(work, struct virtnet_info, refill.work);
 689         bool still_empty;
 690         int i;
 691
 692         for (i = 0; i < vi->curr_queue_pairs; i++) {
 693                 struct receive_queue *rq = &vi->rq[i];
 694
 695                 napi_disable(&rq->napi);
 696                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
 697                 virtnet_napi_enable(rq);
 698
 699                 /* In theory, this can happen: if we don't get any buffers in
 700                  * we will *never* try to fill again.
 701                  */
 702                 if (still_empty)
 703                         schedule_delayed_work(&vi->refill, HZ/2);
 704         }
 705 }
 706
 707 static int virtnet_receive(struct receive_queue *rq, int budget)
 708 {
 709         struct virtnet_info *vi = rq->vq->vdev->priv;
 710         unsigned int len, received = 0;
 711         void *buf;
 712
 713         while (received < budget &&
 714                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 715                 receive_buf(vi, rq, buf, len);
 716                 received++;
 717         }
 718
 719         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 720                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 721                         schedule_delayed_work(&vi->refill, 0);
 722         }
 723
 724         return received;
 725 }
 726
 727 static int virtnet_poll(struct napi_struct *napi, int budget)
 728 {
 729         struct receive_queue *rq =
 730                 container_of(napi, struct receive_queue, napi);
 731         unsigned int r, received;
 732
 733         received = virtnet_receive(rq, budget);
 734
 735         /* Out of packets? */
 736         if (received < budget) {
 737                 r = virtqueue_enable_cb_prepare(rq->vq);
 738                 napi_complete_done(napi, received);
 739                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 740                     napi_schedule_prep(napi)) {
 741                         virtqueue_disable_cb(rq->vq);
 742                         __napi_schedule(napi);
 743                 }
 744         }
 745
 746         return received;
 747 }
 748
 749 #ifdef CONFIG_NET_RX_BUSY_POLL
 750 /* must be called with local_bh_disable()d */
 751 static int virtnet_busy_poll(struct napi_struct *napi)
 752 {
 753         struct receive_queue *rq =
 754                 container_of(napi, struct receive_queue, napi);
 755         struct virtnet_info *vi = rq->vq->vdev->priv;
 756         int r, received = 0, budget = 4;
 757
 758         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
 759                 return LL_FLUSH_FAILED;
 760
 761         if (!napi_schedule_prep(napi))
 762                 return LL_FLUSH_BUSY;
 763
 764         virtqueue_disable_cb(rq->vq);
 765
 766 again:
 767         received += virtnet_receive(rq, budget);
 768
 769         r = virtqueue_enable_cb_prepare(rq->vq);
 770         clear_bit(NAPI_STATE_SCHED, &napi->state);
 771         if (unlikely(virtqueue_poll(rq->vq, r)) &&
 772             napi_schedule_prep(napi)) {
 773                 virtqueue_disable_cb(rq->vq);
 774                 if (received < budget) {
 775                         budget -= received;
 776                         goto again;
 777                 } else {
 778                         __napi_schedule(napi);
 779                 }
 780         }
 781
 782         return received;
 783 }
 784 #endif  /* CONFIG_NET_RX_BUSY_POLL */
 785
 786 static int virtnet_open(struct net_device *dev)
 787 {
 788         struct virtnet_info *vi = netdev_priv(dev);
 789         int i;
 790
 791         for (i = 0; i < vi->max_queue_pairs; i++) {
 792                 if (i < vi->curr_queue_pairs)
 793                         /* Make sure we have some buffers: if oom use wq. */
 794                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
 795                                 schedule_delayed_work(&vi->refill, 0);
 796                 virtnet_napi_enable(&vi->rq[i]);
 797         }
 798
 799         return 0;
 800 }
 801
 802 static void free_old_xmit_skbs(struct send_queue *sq)
 803 {
 804         struct sk_buff *skb;
 805         unsigned int len;
 806         struct virtnet_info *vi = sq->vq->vdev->priv;
 807         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 808
 809         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 810                 pr_debug("Sent skb %p\n", skb);
 811
 812                 u64_stats_update_begin(&stats->tx_syncp);
 813                 stats->tx_bytes += skb->len;
 814                 stats->tx_packets++;
 815                 u64_stats_update_end(&stats->tx_syncp);
 816
 817                 dev_kfree_skb_any(skb);
 818         }
 819 }
 820
 821 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 822 {
 823         struct virtio_net_hdr_mrg_rxbuf *hdr;
 824         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 825         struct virtnet_info *vi = sq->vq->vdev->priv;
 826         unsigned num_sg;
 827         unsigned hdr_len = vi->hdr_len;
 828         bool can_push;
 829
 830         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 831
 832         can_push = vi->any_header_sg &&
 833                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 834                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 835         /* Even if we can, don't push here yet as this would skew
 836          * csum_start offset below. */
 837         if (can_push)
 838                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
 839         else
 840                 hdr = skb_vnet_hdr(skb);
 841
 842         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
 843                                     virtio_is_little_endian(vi->vdev)))
 844                 BUG();
 845
 846         if (vi->mergeable_rx_bufs)
 847                 hdr->num_buffers = 0;
 848
 849         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
 850         if (can_push) {
 851                 __skb_push(skb, hdr_len);
 852                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 853                 /* Pull header back to avoid skew in tx bytes calculations. */
 854                 __skb_pull(skb, hdr_len);
 855         } else {
 856                 sg_set_buf(sq->sg, hdr, hdr_len);
 857                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 858         }
 859         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 860 }
 861
 862 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 863 {
 864         struct virtnet_info *vi = netdev_priv(dev);
 865         int qnum = skb_get_queue_mapping(skb);
 866         struct send_queue *sq = &vi->sq[qnum];
 867         int err;
 868         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
 869         bool kick = !skb->xmit_more;
 870
 871         /* Free up any pending old buffers before queueing new ones. */
 872         free_old_xmit_skbs(sq);
 873
 874         /* timestamp packet in software */
 875         skb_tx_timestamp(skb);
 876
 877         /* Try to transmit */
 878         err = xmit_skb(sq, skb);
 879
 880         /* This should not happen! */
 881         if (unlikely(err)) {
 882                 dev->stats.tx_fifo_errors++;
 883                 if (net_ratelimit())
 884                         dev_warn(&dev->dev,
 885                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 886                 dev->stats.tx_dropped++;
 887                 dev_kfree_skb_any(skb);
 888                 return NETDEV_TX_OK;
 889         }
 890
 891         /* Don't wait up for transmitted skbs to be freed. */
 892         skb_orphan(skb);
 893         nf_reset(skb);
 894
 895         /* If running out of space, stop queue to avoid getting packets that we
 896          * are then unable to transmit.
 897          * An alternative would be to force queuing layer to requeue the skb by
 898          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
 899          * returned in a normal path of operation: it means that driver is not
 900          * maintaining the TX queue stop/start state properly, and causes
 901          * the stack to do a non-trivial amount of useless work.
 902          * Since most packets only take 1 or 2 ring slots, stopping the queue
 903          * early means 16 slots are typically wasted.
 904          */
 905         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 906                 netif_stop_subqueue(dev, qnum);
 907                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 908                         /* More just got used, free them then recheck. */
 909                         free_old_xmit_skbs(sq);
 910                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 911                                 netif_start_subqueue(dev, qnum);
 912                                 virtqueue_disable_cb(sq->vq);
 913                         }
 914                 }
 915         }
 916
 917         if (kick || netif_xmit_stopped(txq))
 918                 virtqueue_kick(sq->vq);
 919
 920         return NETDEV_TX_OK;
 921 }
 922
 923 /*
 924  * Send command via the control virtqueue and check status.  Commands
 925  * supported by the hypervisor, as indicated by feature bits, should
 926  * never fail unless improperly formatted.
 927  */
 928 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 929                                  struct scatterlist *out)
 930 {
 931         struct scatterlist *sgs[4], hdr, stat;
 932         unsigned out_num = 0, tmp;
 933
 934         /* Caller should know better */
 935         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 936
 937         vi->ctrl_status = ~0;
 938         vi->ctrl_hdr.class = class;
 939         vi->ctrl_hdr.cmd = cmd;
 940         /* Add header */
 941         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
 942         sgs[out_num++] = &hdr;
 943
 944         if (out)
 945                 sgs[out_num++] = out;
 946
 947         /* Add return status. */
 948         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
 949         sgs[out_num] = &stat;
 950
 951         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
 952         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
 953
 954         if (unlikely(!virtqueue_kick(vi->cvq)))
 955                 return vi->ctrl_status == VIRTIO_NET_OK;
 956
 957         /* Spin for a response, the kick causes an ioport write, trapping
 958          * into the hypervisor, so the request should be handled immediately.
 959          */
 960         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
 961                !virtqueue_is_broken(vi->cvq))
 962                 cpu_relax();
 963
 964         return vi->ctrl_status == VIRTIO_NET_OK;
 965 }
 966
 967 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 968 {
 969         struct virtnet_info *vi = netdev_priv(dev);
 970         struct virtio_device *vdev = vi->vdev;
 971         int ret;
 972         struct sockaddr *addr = p;
 973         struct scatterlist sg;
 974
 975         ret = eth_prepare_mac_addr_change(dev, p);
 976         if (ret)
 977                 return ret;
 978
 979         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 980                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 981                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 982                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
 983                         dev_warn(&vdev->dev,
 984                                  "Failed to set mac address by vq command.\n");
 985                         return -EINVAL;
 986                 }
 987         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
 988                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
 989                 unsigned int i;
 990
 991                 /* Naturally, this has an atomicity problem. */
 992                 for (i = 0; i < dev->addr_len; i++)
 993                         virtio_cwrite8(vdev,
 994                                        offsetof(struct virtio_net_config, mac) +
 995                                        i, addr->sa_data[i]);
 996         }
 997
 998         eth_commit_mac_addr_change(dev, p);
 999
1000         return 0;
1001 }
1002
1003 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1004                                                struct rtnl_link_stats64 *tot)
1005 {
1006         struct virtnet_info *vi = netdev_priv(dev);
1007         int cpu;
1008         unsigned int start;
1009
1010         for_each_possible_cpu(cpu) {
1011                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1012                 u64 tpackets, tbytes, rpackets, rbytes;
1013
1014                 do {
1015                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1016                         tpackets = stats->tx_packets;
1017                         tbytes   = stats->tx_bytes;
1018                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1019
1020                 do {
1021                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1022                         rpackets = stats->rx_packets;
1023                         rbytes   = stats->rx_bytes;
1024                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1025
1026                 tot->rx_packets += rpackets;
1027                 tot->tx_packets += tpackets;
1028                 tot->rx_bytes   += rbytes;
1029                 tot->tx_bytes   += tbytes;
1030         }
1031
1032         tot->tx_dropped = dev->stats.tx_dropped;
1033         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1034         tot->rx_dropped = dev->stats.rx_dropped;
1035         tot->rx_length_errors = dev->stats.rx_length_errors;
1036         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1037
1038         return tot;
1039 }
1040
1041 #ifdef CONFIG_NET_POLL_CONTROLLER
1042 static void virtnet_netpoll(struct net_device *dev)
1043 {
1044         struct virtnet_info *vi = netdev_priv(dev);
1045         int i;
1046
1047         for (i = 0; i < vi->curr_queue_pairs; i++)
1048                 napi_schedule(&vi->rq[i].napi);
1049 }
1050 #endif
1051
1052 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1053 {
1054         rtnl_lock();
1055         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1056                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1057                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1058         rtnl_unlock();
1059 }
1060
1061 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1062 {
1063         struct scatterlist sg;
1064         struct net_device *dev = vi->dev;
1065
1066         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1067                 return 0;
1068
1069         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1070         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1071
1072         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1073                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1074                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1075                          queue_pairs);
1076                 return -EINVAL;
1077         } else {
1078                 vi->curr_queue_pairs = queue_pairs;
1079                 /* virtnet_open() will refill when device is going to up. */
1080                 if (dev->flags & IFF_UP)
1081                         schedule_delayed_work(&vi->refill, 0);
1082         }
1083
1084         return 0;
1085 }
1086
1087 static int virtnet_close(struct net_device *dev)
1088 {
1089         struct virtnet_info *vi = netdev_priv(dev);
1090         int i;
1091
1092         /* Make sure refill_work doesn't re-enable napi! */
1093         cancel_delayed_work_sync(&vi->refill);
1094
1095         for (i = 0; i < vi->max_queue_pairs; i++)
1096                 napi_disable(&vi->rq[i].napi);
1097
1098         return 0;
1099 }
1100
1101 static void virtnet_set_rx_mode(struct net_device *dev)
1102 {
1103         struct virtnet_info *vi = netdev_priv(dev);
1104         struct scatterlist sg[2];
1105         struct virtio_net_ctrl_mac *mac_data;
1106         struct netdev_hw_addr *ha;
1107         int uc_count;
1108         int mc_count;
1109         void *buf;
1110         int i;
1111
1112         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1113         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1114                 return;
1115
1116         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1117         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1118
1119         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1120
1121         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1122                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1123                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1124                          vi->ctrl_promisc ? "en" : "dis");
1125
1126         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1127
1128         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1129                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1130                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1131                          vi->ctrl_allmulti ? "en" : "dis");
1132
1133         uc_count = netdev_uc_count(dev);
1134         mc_count = netdev_mc_count(dev);
1135         /* MAC filter - use one buffer for both lists */
1136         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1137                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1138         mac_data = buf;
1139         if (!buf)
1140                 return;
1141
1142         sg_init_table(sg, 2);
1143
1144         /* Store the unicast list and count in the front of the buffer */
1145         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1146         i = 0;
1147         netdev_for_each_uc_addr(ha, dev)
1148                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1149
1150         sg_set_buf(&sg[0], mac_data,
1151                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1152
1153         /* multicast list and count fill the end */
1154         mac_data = (void *)&mac_data->macs[uc_count][0];
1155
1156         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1157         i = 0;
1158         netdev_for_each_mc_addr(ha, dev)
1159                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1160
1161         sg_set_buf(&sg[1], mac_data,
1162                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1163
1164         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1165                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1166                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1167
1168         kfree(buf);
1169 }
1170
1171 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1172                                    __be16 proto, u16 vid)
1173 {
1174         struct virtnet_info *vi = netdev_priv(dev);
1175         struct scatterlist sg;
1176
1177         vi->ctrl_vid = vid;
1178         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1179
1180         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1181                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1182                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1183         return 0;
1184 }
1185
1186 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1187                                     __be16 proto, u16 vid)
1188 {
1189         struct virtnet_info *vi = netdev_priv(dev);
1190         struct scatterlist sg;
1191
1192         vi->ctrl_vid = vid;
1193         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1194
1195         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1196                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1197                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1198         return 0;
1199 }
1200
1201 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1202 {
1203         int i;
1204
1205         if (vi->affinity_hint_set) {
1206                 for (i = 0; i < vi->max_queue_pairs; i++) {
1207                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1208                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1209                 }
1210
1211                 vi->affinity_hint_set = false;
1212         }
1213 }
1214
1215 static void virtnet_set_affinity(struct virtnet_info *vi)
1216 {
1217         int i;
1218         int cpu;
1219
1220         /* In multiqueue mode, when the number of cpu is equal to the number of
1221          * queue pairs, we let the queue pairs to be private to one cpu by
1222          * setting the affinity hint to eliminate the contention.
1223          */
1224         if (vi->curr_queue_pairs == 1 ||
1225             vi->max_queue_pairs != num_online_cpus()) {
1226                 virtnet_clean_affinity(vi, -1);
1227                 return;
1228         }
1229
1230         i = 0;
1231         for_each_online_cpu(cpu) {
1232                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1233                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1234                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1235                 i++;
1236         }
1237
1238         vi->affinity_hint_set = true;
1239 }
1240
1241 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1242 {
1243         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1244                                                    node);
1245         virtnet_set_affinity(vi);
1246         return 0;
1247 }
1248
1249 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1250 {
1251         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1252                                                    node_dead);
1253         virtnet_set_affinity(vi);
1254         return 0;
1255 }
1256
1257 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1258 {
1259         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1260                                                    node);
1261
1262         virtnet_clean_affinity(vi, cpu);
1263         return 0;
1264 }
1265
1266 static enum cpuhp_state virtionet_online;
1267
1268 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1269 {
1270         int ret;
1271
1272         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1273         if (ret)
1274                 return ret;
1275         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1276                                                &vi->node_dead);
1277         if (!ret)
1278                 return ret;
1279         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1280         return ret;
1281 }
1282
1283 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1284 {
1285         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1286         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1287                                             &vi->node_dead);
1288 }
1289
1290 static void virtnet_get_ringparam(struct net_device *dev,
1291                                 struct ethtool_ringparam *ring)
1292 {
1293         struct virtnet_info *vi = netdev_priv(dev);
1294
1295         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1296         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1297         ring->rx_pending = ring->rx_max_pending;
1298         ring->tx_pending = ring->tx_max_pending;
1299 }
1300
1301
1302 static void virtnet_get_drvinfo(struct net_device *dev,
1303                                 struct ethtool_drvinfo *info)
1304 {
1305         struct virtnet_info *vi = netdev_priv(dev);
1306         struct virtio_device *vdev = vi->vdev;
1307
1308         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1309         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1310         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1311
1312 }
1313
1314 /* TODO: Eliminate OOO packets during switching */
1315 static int virtnet_set_channels(struct net_device *dev,
1316                                 struct ethtool_channels *channels)
1317 {
1318         struct virtnet_info *vi = netdev_priv(dev);
1319         u16 queue_pairs = channels->combined_count;
1320         int err;
1321
1322         /* We don't support separate rx/tx channels.
1323          * We don't allow setting 'other' channels.
1324          */
1325         if (channels->rx_count || channels->tx_count || channels->other_count)
1326                 return -EINVAL;
1327
1328         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1329                 return -EINVAL;
1330
1331         get_online_cpus();
1332         err = virtnet_set_queues(vi, queue_pairs);
1333         if (!err) {
1334                 netif_set_real_num_tx_queues(dev, queue_pairs);
1335                 netif_set_real_num_rx_queues(dev, queue_pairs);
1336
1337                 virtnet_set_affinity(vi);
1338         }
1339         put_online_cpus();
1340
1341         return err;
1342 }
1343
1344 static void virtnet_get_channels(struct net_device *dev,
1345                                  struct ethtool_channels *channels)
1346 {
1347         struct virtnet_info *vi = netdev_priv(dev);
1348
1349         channels->combined_count = vi->curr_queue_pairs;
1350         channels->max_combined = vi->max_queue_pairs;
1351         channels->max_other = 0;
1352         channels->rx_count = 0;
1353         channels->tx_count = 0;
1354         channels->other_count = 0;
1355 }
1356
1357 /* Check if the user is trying to change anything besides speed/duplex */
1358 static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1359 {
1360         struct ethtool_cmd diff1 = *cmd;
1361         struct ethtool_cmd diff2 = {};
1362
1363         /* cmd is always set so we need to clear it, validate the port type
1364          * and also without autonegotiation we can ignore advertising
1365          */
1366         ethtool_cmd_speed_set(&diff1, 0);
1367         diff2.port = PORT_OTHER;
1368         diff1.advertising = 0;
1369         diff1.duplex = 0;
1370         diff1.cmd = 0;
1371
1372         return !memcmp(&diff1, &diff2, sizeof(diff1));
1373 }
1374
1375 static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1376 {
1377         struct virtnet_info *vi = netdev_priv(dev);
1378         u32 speed;
1379
1380         speed = ethtool_cmd_speed(cmd);
1381         /* don't allow custom speed and duplex */
1382         if (!ethtool_validate_speed(speed) ||
1383             !ethtool_validate_duplex(cmd->duplex) ||
1384             !virtnet_validate_ethtool_cmd(cmd))
1385                 return -EINVAL;
1386         vi->speed = speed;
1387         vi->duplex = cmd->duplex;
1388
1389         return 0;
1390 }
1391
1392 static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1393 {
1394         struct virtnet_info *vi = netdev_priv(dev);
1395
1396         ethtool_cmd_speed_set(cmd, vi->speed);
1397         cmd->duplex = vi->duplex;
1398         cmd->port = PORT_OTHER;
1399
1400         return 0;
1401 }
1402
1403 static void virtnet_init_settings(struct net_device *dev)
1404 {
1405         struct virtnet_info *vi = netdev_priv(dev);
1406
1407         vi->speed = SPEED_UNKNOWN;
1408         vi->duplex = DUPLEX_UNKNOWN;
1409 }
1410
1411 static const struct ethtool_ops virtnet_ethtool_ops = {
1412         .get_drvinfo = virtnet_get_drvinfo,
1413         .get_link = ethtool_op_get_link,
1414         .get_ringparam = virtnet_get_ringparam,
1415         .set_channels = virtnet_set_channels,
1416         .get_channels = virtnet_get_channels,
1417         .get_ts_info = ethtool_op_get_ts_info,
1418         .get_settings = virtnet_get_settings,
1419         .set_settings = virtnet_set_settings,
1420 };
1421
1422 #define MIN_MTU 68
1423 #define MAX_MTU 65535
1424
1425 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1426 {
1427         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1428                 return -EINVAL;
1429         dev->mtu = new_mtu;
1430         return 0;
1431 }
1432
1433 static const struct net_device_ops virtnet_netdev = {
1434         .ndo_open            = virtnet_open,
1435         .ndo_stop            = virtnet_close,
1436         .ndo_start_xmit      = start_xmit,
1437         .ndo_validate_addr   = eth_validate_addr,
1438         .ndo_set_mac_address = virtnet_set_mac_address,
1439         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1440         .ndo_change_mtu      = virtnet_change_mtu,
1441         .ndo_get_stats64     = virtnet_stats,
1442         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1443         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1444 #ifdef CONFIG_NET_POLL_CONTROLLER
1445         .ndo_poll_controller = virtnet_netpoll,
1446 #endif
1447 #ifdef CONFIG_NET_RX_BUSY_POLL
1448         .ndo_busy_poll          = virtnet_busy_poll,
1449 #endif
1450 };
1451
1452 static void virtnet_config_changed_work(struct work_struct *work)
1453 {
1454         struct virtnet_info *vi =
1455                 container_of(work, struct virtnet_info, config_work);
1456         u16 v;
1457
1458         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1459                                  struct virtio_net_config, status, &v) < 0)
1460                 return;
1461
1462         if (v & VIRTIO_NET_S_ANNOUNCE) {
1463                 netdev_notify_peers(vi->dev);
1464                 virtnet_ack_link_announce(vi);
1465         }
1466
1467         /* Ignore unknown (future) status bits */
1468         v &= VIRTIO_NET_S_LINK_UP;
1469
1470         if (vi->status == v)
1471                 return;
1472
1473         vi->status = v;
1474
1475         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1476                 netif_carrier_on(vi->dev);
1477                 netif_tx_wake_all_queues(vi->dev);
1478         } else {
1479                 netif_carrier_off(vi->dev);
1480                 netif_tx_stop_all_queues(vi->dev);
1481         }
1482 }
1483
1484 static void virtnet_config_changed(struct virtio_device *vdev)
1485 {
1486         struct virtnet_info *vi = vdev->priv;
1487
1488         schedule_work(&vi->config_work);
1489 }
1490
1491 static void virtnet_free_queues(struct virtnet_info *vi)
1492 {
1493         int i;
1494
1495         for (i = 0; i < vi->max_queue_pairs; i++) {
1496                 napi_hash_del(&vi->rq[i].napi);
1497                 netif_napi_del(&vi->rq[i].napi);
1498         }
1499
1500         kfree(vi->rq);
1501         kfree(vi->sq);
1502 }
1503
1504 static void free_receive_bufs(struct virtnet_info *vi)
1505 {
1506         int i;
1507
1508         for (i = 0; i < vi->max_queue_pairs; i++) {
1509                 while (vi->rq[i].pages)
1510                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1511         }
1512 }
1513
1514 static void free_receive_page_frags(struct virtnet_info *vi)
1515 {
1516         int i;
1517         for (i = 0; i < vi->max_queue_pairs; i++)
1518                 if (vi->rq[i].alloc_frag.page)
1519                         put_page(vi->rq[i].alloc_frag.page);
1520 }
1521
1522 static void free_unused_bufs(struct virtnet_info *vi)
1523 {
1524         void *buf;
1525         int i;
1526
1527         for (i = 0; i < vi->max_queue_pairs; i++) {
1528                 struct virtqueue *vq = vi->sq[i].vq;
1529                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1530                         dev_kfree_skb(buf);
1531         }
1532
1533         for (i = 0; i < vi->max_queue_pairs; i++) {
1534                 struct virtqueue *vq = vi->rq[i].vq;
1535
1536                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1537                         if (vi->mergeable_rx_bufs) {
1538                                 unsigned long ctx = (unsigned long)buf;
1539                                 void *base = mergeable_ctx_to_buf_address(ctx);
1540                                 put_page(virt_to_head_page(base));
1541                         } else if (vi->big_packets) {
1542                                 give_pages(&vi->rq[i], buf);
1543                         } else {
1544                                 dev_kfree_skb(buf);
1545                         }
1546                 }
1547         }
1548 }
1549
1550 static void virtnet_del_vqs(struct virtnet_info *vi)
1551 {
1552         struct virtio_device *vdev = vi->vdev;
1553
1554         virtnet_clean_affinity(vi, -1);
1555
1556         vdev->config->del_vqs(vdev);
1557
1558         virtnet_free_queues(vi);
1559 }
1560
1561 static int virtnet_find_vqs(struct virtnet_info *vi)
1562 {
1563         vq_callback_t **callbacks;
1564         struct virtqueue **vqs;
1565         int ret = -ENOMEM;
1566         int i, total_vqs;
1567         const char **names;
1568
1569         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1570          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1571          * possible control vq.
1572          */
1573         total_vqs = vi->max_queue_pairs * 2 +
1574                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1575
1576         /* Allocate space for find_vqs parameters */
1577         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1578         if (!vqs)
1579                 goto err_vq;
1580         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1581         if (!callbacks)
1582                 goto err_callback;
1583         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1584         if (!names)
1585                 goto err_names;
1586
1587         /* Parameters for control virtqueue, if any */
1588         if (vi->has_cvq) {
1589                 callbacks[total_vqs - 1] = NULL;
1590                 names[total_vqs - 1] = "control";
1591         }
1592
1593         /* Allocate/initialize parameters for send/receive virtqueues */
1594         for (i = 0; i < vi->max_queue_pairs; i++) {
1595                 callbacks[rxq2vq(i)] = skb_recv_done;
1596                 callbacks[txq2vq(i)] = skb_xmit_done;
1597                 sprintf(vi->rq[i].name, "input.%d", i);
1598                 sprintf(vi->sq[i].name, "output.%d", i);
1599                 names[rxq2vq(i)] = vi->rq[i].name;
1600                 names[txq2vq(i)] = vi->sq[i].name;
1601         }
1602
1603         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1604                                          names);
1605         if (ret)
1606                 goto err_find;
1607
1608         if (vi->has_cvq) {
1609                 vi->cvq = vqs[total_vqs - 1];
1610                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1611                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1612         }
1613
1614         for (i = 0; i < vi->max_queue_pairs; i++) {
1615                 vi->rq[i].vq = vqs[rxq2vq(i)];
1616                 vi->sq[i].vq = vqs[txq2vq(i)];
1617         }
1618
1619         kfree(names);
1620         kfree(callbacks);
1621         kfree(vqs);
1622
1623         return 0;
1624
1625 err_find:
1626         kfree(names);
1627 err_names:
1628         kfree(callbacks);
1629 err_callback:
1630         kfree(vqs);
1631 err_vq:
1632         return ret;
1633 }
1634
1635 static int virtnet_alloc_queues(struct virtnet_info *vi)
1636 {
1637         int i;
1638
1639         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1640         if (!vi->sq)
1641                 goto err_sq;
1642         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1643         if (!vi->rq)
1644                 goto err_rq;
1645
1646         INIT_DELAYED_WORK(&vi->refill, refill_work);
1647         for (i = 0; i < vi->max_queue_pairs; i++) {
1648                 vi->rq[i].pages = NULL;
1649                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1650                                napi_weight);
1651
1652                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1653                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
1654                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1655         }
1656
1657         return 0;
1658
1659 err_rq:
1660         kfree(vi->sq);
1661 err_sq:
1662         return -ENOMEM;
1663 }
1664
1665 static int init_vqs(struct virtnet_info *vi)
1666 {
1667         int ret;
1668
1669         /* Allocate send & receive queues */
1670         ret = virtnet_alloc_queues(vi);
1671         if (ret)
1672                 goto err;
1673
1674         ret = virtnet_find_vqs(vi);
1675         if (ret)
1676                 goto err_free;
1677
1678         get_online_cpus();
1679         virtnet_set_affinity(vi);
1680         put_online_cpus();
1681
1682         return 0;
1683
1684 err_free:
1685         virtnet_free_queues(vi);
1686 err:
1687         return ret;
1688 }
1689
1690 #ifdef CONFIG_SYSFS
1691 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1692                 struct rx_queue_attribute *attribute, char *buf)
1693 {
1694         struct virtnet_info *vi = netdev_priv(queue->dev);
1695         unsigned int queue_index = get_netdev_rx_queue_index(queue);
1696         struct ewma_pkt_len *avg;
1697
1698         BUG_ON(queue_index >= vi->max_queue_pairs);
1699         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1700         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1701 }
1702
1703 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1704         __ATTR_RO(mergeable_rx_buffer_size);
1705
1706 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1707         &mergeable_rx_buffer_size_attribute.attr,
1708         NULL
1709 };
1710
1711 static const struct attribute_group virtio_net_mrg_rx_group = {
1712         .name = "virtio_net",
1713         .attrs = virtio_net_mrg_rx_attrs
1714 };
1715 #endif
1716
1717 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1718                                     unsigned int fbit,
1719                                     const char *fname, const char *dname)
1720 {
1721         if (!virtio_has_feature(vdev, fbit))
1722                 return false;
1723
1724         dev_err(&vdev->dev, "device advertises feature %s but not %s",
1725                 fname, dname);
1726
1727         return true;
1728 }
1729
1730 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
1731         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1732
1733 static bool virtnet_validate_features(struct virtio_device *vdev)
1734 {
1735         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1736             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1737                              "VIRTIO_NET_F_CTRL_VQ") ||
1738              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1739                              "VIRTIO_NET_F_CTRL_VQ") ||
1740              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1741                              "VIRTIO_NET_F_CTRL_VQ") ||
1742              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1743              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1744                              "VIRTIO_NET_F_CTRL_VQ"))) {
1745                 return false;
1746         }
1747
1748         return true;
1749 }
1750
1751 static int virtnet_probe(struct virtio_device *vdev)
1752 {
1753         int i, err;
1754         struct net_device *dev;
1755         struct virtnet_info *vi;
1756         u16 max_queue_pairs;
1757         int mtu;
1758
1759         if (!vdev->config->get) {
1760                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
1761                         __func__);
1762                 return -EINVAL;
1763         }
1764
1765         if (!virtnet_validate_features(vdev))
1766                 return -EINVAL;
1767
1768         /* Find if host supports multiqueue virtio_net device */
1769         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1770                                    struct virtio_net_config,
1771                                    max_virtqueue_pairs, &max_queue_pairs);
1772
1773         /* We need at least 2 queue's */
1774         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1775             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1776             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1777                 max_queue_pairs = 1;
1778
1779         /* Allocate ourselves a network device with room for our info */
1780         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1781         if (!dev)
1782                 return -ENOMEM;
1783
1784         /* Set up network device as normal. */
1785         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1786         dev->netdev_ops = &virtnet_netdev;
1787         dev->features = NETIF_F_HIGHDMA;
1788
1789         dev->ethtool_ops = &virtnet_ethtool_ops;
1790         SET_NETDEV_DEV(dev, &vdev->dev);
1791
1792         /* Do we support "hardware" checksums? */
1793         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1794                 /* This opens up the world of extra features. */
1795                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1796                 if (csum)
1797                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1798
1799                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1800                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1801                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1802                 }
1803                 /* Individual feature bits: what can host handle? */
1804                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1805                         dev->hw_features |= NETIF_F_TSO;
1806                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1807                         dev->hw_features |= NETIF_F_TSO6;
1808                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1809                         dev->hw_features |= NETIF_F_TSO_ECN;
1810                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1811                         dev->hw_features |= NETIF_F_UFO;
1812
1813                 dev->features |= NETIF_F_GSO_ROBUST;
1814
1815                 if (gso)
1816                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1817                 /* (!csum && gso) case will be fixed by register_netdev() */
1818         }
1819         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1820                 dev->features |= NETIF_F_RXCSUM;
1821
1822         dev->vlan_features = dev->features;
1823
1824         /* Configuration may specify what MAC to use.  Otherwise random. */
1825         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1826                 virtio_cread_bytes(vdev,
1827                                    offsetof(struct virtio_net_config, mac),
1828                                    dev->dev_addr, dev->addr_len);
1829         else
1830                 eth_hw_addr_random(dev);
1831
1832         /* Set up our device-specific information */
1833         vi = netdev_priv(dev);
1834         vi->dev = dev;
1835         vi->vdev = vdev;
1836         vdev->priv = vi;
1837         vi->stats = alloc_percpu(struct virtnet_stats);
1838         err = -ENOMEM;
1839         if (vi->stats == NULL)
1840                 goto free;
1841
1842         for_each_possible_cpu(i) {
1843                 struct virtnet_stats *virtnet_stats;
1844                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1845                 u64_stats_init(&virtnet_stats->tx_syncp);
1846                 u64_stats_init(&virtnet_stats->rx_syncp);
1847         }
1848
1849         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1850
1851         /* If we can receive ANY GSO packets, we must allocate large ones. */
1852         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1853             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1854             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1855             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1856                 vi->big_packets = true;
1857
1858         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1859                 vi->mergeable_rx_bufs = true;
1860
1861         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
1862             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1863                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1864         else
1865                 vi->hdr_len = sizeof(struct virtio_net_hdr);
1866
1867         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
1868             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1869                 vi->any_header_sg = true;
1870
1871         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1872                 vi->has_cvq = true;
1873
1874         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
1875                 mtu = virtio_cread16(vdev,
1876                                      offsetof(struct virtio_net_config,
1877                                               mtu));
1878                 if (virtnet_change_mtu(dev, mtu))
1879                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
1880         }
1881
1882         if (vi->any_header_sg)
1883                 dev->needed_headroom = vi->hdr_len;
1884
1885         /* Use single tx/rx queue pair as default */
1886         vi->curr_queue_pairs = 1;
1887         vi->max_queue_pairs = max_queue_pairs;
1888
1889         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1890         err = init_vqs(vi);
1891         if (err)
1892                 goto free_stats;
1893
1894 #ifdef CONFIG_SYSFS
1895         if (vi->mergeable_rx_bufs)
1896                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1897 #endif
1898         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1899         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1900
1901         virtnet_init_settings(dev);
1902
1903         err = register_netdev(dev);
1904         if (err) {
1905                 pr_debug("virtio_net: registering device failed\n");
1906                 goto free_vqs;
1907         }
1908
1909         virtio_device_ready(vdev);
1910
1911         err = virtnet_cpu_notif_add(vi);
1912         if (err) {
1913                 pr_debug("virtio_net: registering cpu notifier failed\n");
1914                 goto free_unregister_netdev;
1915         }
1916
1917         /* Assume link up if device can't report link status,
1918            otherwise get link status from config. */
1919         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1920                 netif_carrier_off(dev);
1921                 schedule_work(&vi->config_work);
1922         } else {
1923                 vi->status = VIRTIO_NET_S_LINK_UP;
1924                 netif_carrier_on(dev);
1925         }
1926
1927         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1928                  dev->name, max_queue_pairs);
1929
1930         return 0;
1931
1932 free_unregister_netdev:
1933         vi->vdev->config->reset(vdev);
1934
1935         unregister_netdev(dev);
1936 free_vqs:
1937         cancel_delayed_work_sync(&vi->refill);
1938         free_receive_page_frags(vi);
1939         virtnet_del_vqs(vi);
1940 free_stats:
1941         free_percpu(vi->stats);
1942 free:
1943         free_netdev(dev);
1944         return err;
1945 }
1946
1947 static void remove_vq_common(struct virtnet_info *vi)
1948 {
1949         vi->vdev->config->reset(vi->vdev);
1950
1951         /* Free unused buffers in both send and recv, if any. */
1952         free_unused_bufs(vi);
1953
1954         free_receive_bufs(vi);
1955
1956         free_receive_page_frags(vi);
1957
1958         virtnet_del_vqs(vi);
1959 }
1960
1961 static void virtnet_remove(struct virtio_device *vdev)
1962 {
1963         struct virtnet_info *vi = vdev->priv;
1964
1965         virtnet_cpu_notif_remove(vi);
1966
1967         /* Make sure no work handler is accessing the device. */
1968         flush_work(&vi->config_work);
1969
1970         unregister_netdev(vi->dev);
1971
1972         remove_vq_common(vi);
1973
1974         free_percpu(vi->stats);
1975         free_netdev(vi->dev);
1976 }
1977
1978 #ifdef CONFIG_PM_SLEEP
1979 static int virtnet_freeze(struct virtio_device *vdev)
1980 {
1981         struct virtnet_info *vi = vdev->priv;
1982         int i;
1983
1984         virtnet_cpu_notif_remove(vi);
1985
1986         /* Make sure no work handler is accessing the device */
1987         flush_work(&vi->config_work);
1988
1989         netif_device_detach(vi->dev);
1990         cancel_delayed_work_sync(&vi->refill);
1991
1992         if (netif_running(vi->dev)) {
1993                 for (i = 0; i < vi->max_queue_pairs; i++)
1994                         napi_disable(&vi->rq[i].napi);
1995         }
1996
1997         remove_vq_common(vi);
1998
1999         return 0;
2000 }
2001
2002 static int virtnet_restore(struct virtio_device *vdev)
2003 {
2004         struct virtnet_info *vi = vdev->priv;
2005         int err, i;
2006
2007         err = init_vqs(vi);
2008         if (err)
2009                 return err;
2010
2011         virtio_device_ready(vdev);
2012
2013         if (netif_running(vi->dev)) {
2014                 for (i = 0; i < vi->curr_queue_pairs; i++)
2015                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2016                                 schedule_delayed_work(&vi->refill, 0);
2017
2018                 for (i = 0; i < vi->max_queue_pairs; i++)
2019                         virtnet_napi_enable(&vi->rq[i]);
2020         }
2021
2022         netif_device_attach(vi->dev);
2023
2024         rtnl_lock();
2025         virtnet_set_queues(vi, vi->curr_queue_pairs);
2026         rtnl_unlock();
2027
2028         err = virtnet_cpu_notif_add(vi);
2029         if (err)
2030                 return err;
2031
2032         return 0;
2033 }
2034 #endif
2035
2036 static struct virtio_device_id id_table[] = {
2037         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2038         { 0 },
2039 };
2040
2041 static unsigned int features[] = {
2042         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
2043         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
2044         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
2045         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
2046         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
2047         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
2048         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
2049         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
2050         VIRTIO_NET_F_CTRL_MAC_ADDR,
2051         VIRTIO_F_ANY_LAYOUT,
2052         VIRTIO_NET_F_MTU,
2053 };
2054
2055 static struct virtio_driver virtio_net_driver = {
2056         .feature_table = features,
2057         .feature_table_size = ARRAY_SIZE(features),
2058         .driver.name =  KBUILD_MODNAME,
2059         .driver.owner = THIS_MODULE,
2060         .id_table =     id_table,
2061         .probe =        virtnet_probe,
2062         .remove =       virtnet_remove,
2063         .config_changed = virtnet_config_changed,
2064 #ifdef CONFIG_PM_SLEEP
2065         .freeze =       virtnet_freeze,
2066         .restore =      virtnet_restore,
2067 #endif
2068 };
2069
2070 static __init int virtio_net_driver_init(void)
2071 {
2072         int ret;
2073
2074         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "AP_VIRT_NET_ONLINE",
2075                                       virtnet_cpu_online,
2076                                       virtnet_cpu_down_prep);
2077         if (ret < 0)
2078                 goto out;
2079         virtionet_online = ret;
2080         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "VIRT_NET_DEAD",
2081                                       NULL, virtnet_cpu_dead);
2082         if (ret)
2083                 goto err_dead;
2084
2085         ret = register_virtio_driver(&virtio_net_driver);
2086         if (ret)
2087                 goto err_virtio;
2088         return 0;
2089 err_virtio:
2090         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2091 err_dead:
2092         cpuhp_remove_multi_state(virtionet_online);
2093 out:
2094         return ret;
2095 }
2096 module_init(virtio_net_driver_init);
2097
2098 static __exit void virtio_net_driver_exit(void)
2099 {
2100         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2101         cpuhp_remove_multi_state(virtionet_online);
2102         unregister_virtio_driver(&virtio_net_driver);
2103 }
2104 module_exit(virtio_net_driver_exit);
2105
2106 MODULE_DEVICE_TABLE(virtio, id_table);
2107 MODULE_DESCRIPTION("Virtio network driver");
2108 MODULE_LICENSE("GPL");