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/bpf.h>
  26 #include <linux/scatterlist.h>
  27 #include <linux/if_vlan.h>
  28 #include <linux/slab.h>
  29 #include <linux/cpu.h>
  30 #include <linux/average.h>
  31 #include <net/busy_poll.h>
  32
  33 static int napi_weight = NAPI_POLL_WEIGHT;
  34 module_param(napi_weight, int, 0444);
  35
  36 static bool csum = true, gso = true;
  37 module_param(csum, bool, 0444);
  38 module_param(gso, bool, 0444);
  39
  40 /* FIXME: MTU in config. */
  41 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  42 #define GOOD_COPY_LEN   128
  43
  44 /* RX packet size EWMA. The average packet size is used to determine the packet
  45  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  46  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  47  * term, transient changes in packet size.
  48  */
  49 DECLARE_EWMA(pkt_len, 1, 64)
  50
  51 /* Minimum alignment for mergeable packet buffers. */
  52 #define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, 256)
  53
  54 #define VIRTNET_DRIVER_VERSION "1.0.0"
  55
  56 struct virtnet_stats {
  57         struct u64_stats_sync tx_syncp;
  58         struct u64_stats_sync rx_syncp;
  59         u64 tx_bytes;
  60         u64 tx_packets;
  61
  62         u64 rx_bytes;
  63         u64 rx_packets;
  64 };
  65
  66 /* Internal representation of a send virtqueue */
  67 struct send_queue {
  68         /* Virtqueue associated with this send _queue */
  69         struct virtqueue *vq;
  70
  71         /* TX: fragments + linear part + virtio header */
  72         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  73
  74         /* Name of the send queue: output.$index */
  75         char name[40];
  76 };
  77
  78 /* Internal representation of a receive virtqueue */
  79 struct receive_queue {
  80         /* Virtqueue associated with this receive_queue */
  81         struct virtqueue *vq;
  82
  83         struct napi_struct napi;
  84
  85         struct bpf_prog __rcu *xdp_prog;
  86
  87         /* Chain pages by the private ptr. */
  88         struct page *pages;
  89
  90         /* Average packet length for mergeable receive buffers. */
  91         struct ewma_pkt_len mrg_avg_pkt_len;
  92
  93         /* Page frag for packet buffer allocation. */
  94         struct page_frag alloc_frag;
  95
  96         /* RX: fragments + linear part + virtio header */
  97         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  98
  99         /* Name of this receive queue: input.$index */
 100         char name[40];
 101 };
 102
 103 struct virtnet_info {
 104         struct virtio_device *vdev;
 105         struct virtqueue *cvq;
 106         struct net_device *dev;
 107         struct send_queue *sq;
 108         struct receive_queue *rq;
 109         unsigned int status;
 110
 111         /* Max # of queue pairs supported by the device */
 112         u16 max_queue_pairs;
 113
 114         /* # of queue pairs currently used by the driver */
 115         u16 curr_queue_pairs;
 116
 117         /* # of XDP queue pairs currently used by the driver */
 118         u16 xdp_queue_pairs;
 119
 120         /* I like... big packets and I cannot lie! */
 121         bool big_packets;
 122
 123         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 124         bool mergeable_rx_bufs;
 125
 126         /* Has control virtqueue */
 127         bool has_cvq;
 128
 129         /* Host can handle any s/g split between our header and packet data */
 130         bool any_header_sg;
 131
 132         /* Packet virtio header size */
 133         u8 hdr_len;
 134
 135         /* Active statistics */
 136         struct virtnet_stats __percpu *stats;
 137
 138         /* Work struct for refilling if we run low on memory. */
 139         struct delayed_work refill;
 140
 141         /* Work struct for config space updates */
 142         struct work_struct config_work;
 143
 144         /* Does the affinity hint is set for virtqueues? */
 145         bool affinity_hint_set;
 146
 147         /* CPU hotplug instances for online & dead */
 148         struct hlist_node node;
 149         struct hlist_node node_dead;
 150
 151         /* Control VQ buffers: protected by the rtnl lock */
 152         struct virtio_net_ctrl_hdr ctrl_hdr;
 153         virtio_net_ctrl_ack ctrl_status;
 154         struct virtio_net_ctrl_mq ctrl_mq;
 155         u8 ctrl_promisc;
 156         u8 ctrl_allmulti;
 157         u16 ctrl_vid;
 158
 159         /* Ethtool settings */
 160         u8 duplex;
 161         u32 speed;
 162 };
 163
 164 struct padded_vnet_hdr {
 165         struct virtio_net_hdr_mrg_rxbuf hdr;
 166         /*
 167          * hdr is in a separate sg buffer, and data sg buffer shares same page
 168          * with this header sg. This padding makes next sg 16 byte aligned
 169          * after the header.
 170          */
 171         char padding[4];
 172 };
 173
 174 /* Converting between virtqueue no. and kernel tx/rx queue no.
 175  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 176  */
 177 static int vq2txq(struct virtqueue *vq)
 178 {
 179         return (vq->index - 1) / 2;
 180 }
 181
 182 static int txq2vq(int txq)
 183 {
 184         return txq * 2 + 1;
 185 }
 186
 187 static int vq2rxq(struct virtqueue *vq)
 188 {
 189         return vq->index / 2;
 190 }
 191
 192 static int rxq2vq(int rxq)
 193 {
 194         return rxq * 2;
 195 }
 196
 197 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 198 {
 199         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 200 }
 201
 202 /*
 203  * private is used to chain pages for big packets, put the whole
 204  * most recent used list in the beginning for reuse
 205  */
 206 static void give_pages(struct receive_queue *rq, struct page *page)
 207 {
 208         struct page *end;
 209
 210         /* Find end of list, sew whole thing into vi->rq.pages. */
 211         for (end = page; end->private; end = (struct page *)end->private);
 212         end->private = (unsigned long)rq->pages;
 213         rq->pages = page;
 214 }
 215
 216 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 217 {
 218         struct page *p = rq->pages;
 219
 220         if (p) {
 221                 rq->pages = (struct page *)p->private;
 222                 /* clear private here, it is used to chain pages */
 223                 p->private = 0;
 224         } else
 225                 p = alloc_page(gfp_mask);
 226         return p;
 227 }
 228
 229 static void skb_xmit_done(struct virtqueue *vq)
 230 {
 231         struct virtnet_info *vi = vq->vdev->priv;
 232
 233         /* Suppress further interrupts. */
 234         virtqueue_disable_cb(vq);
 235
 236         /* We were probably waiting for more output buffers. */
 237         netif_wake_subqueue(vi->dev, vq2txq(vq));
 238 }
 239
 240 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
 241 {
 242         unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
 243         return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
 244 }
 245
 246 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
 247 {
 248         return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
 249
 250 }
 251
 252 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
 253 {
 254         unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
 255         return (unsigned long)buf | (size - 1);
 256 }
 257
 258 /* Called from bottom half context */
 259 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 260                                    struct receive_queue *rq,
 261                                    struct page *page, unsigned int offset,
 262                                    unsigned int len, unsigned int truesize)
 263 {
 264         struct sk_buff *skb;
 265         struct virtio_net_hdr_mrg_rxbuf *hdr;
 266         unsigned int copy, hdr_len, hdr_padded_len;
 267         char *p;
 268
 269         p = page_address(page) + offset;
 270
 271         /* copy small packet so we can reuse these pages for small data */
 272         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 273         if (unlikely(!skb))
 274                 return NULL;
 275
 276         hdr = skb_vnet_hdr(skb);
 277
 278         hdr_len = vi->hdr_len;
 279         if (vi->mergeable_rx_bufs)
 280                 hdr_padded_len = sizeof *hdr;
 281         else
 282                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 283
 284         memcpy(hdr, p, hdr_len);
 285
 286         len -= hdr_len;
 287         offset += hdr_padded_len;
 288         p += hdr_padded_len;
 289
 290         copy = len;
 291         if (copy > skb_tailroom(skb))
 292                 copy = skb_tailroom(skb);
 293         memcpy(skb_put(skb, copy), p, copy);
 294
 295         len -= copy;
 296         offset += copy;
 297
 298         if (vi->mergeable_rx_bufs) {
 299                 if (len)
 300                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 301                 else
 302                         put_page(page);
 303                 return skb;
 304         }
 305
 306         /*
 307          * Verify that we can indeed put this data into a skb.
 308          * This is here to handle cases when the device erroneously
 309          * tries to receive more than is possible. This is usually
 310          * the case of a broken device.
 311          */
 312         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 313                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 314                 dev_kfree_skb(skb);
 315                 return NULL;
 316         }
 317         BUG_ON(offset >= PAGE_SIZE);
 318         while (len) {
 319                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 320                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 321                                 frag_size, truesize);
 322                 len -= frag_size;
 323                 page = (struct page *)page->private;
 324                 offset = 0;
 325         }
 326
 327         if (page)
 328                 give_pages(rq, page);
 329
 330         return skb;
 331 }
 332
 333 static void virtnet_xdp_xmit(struct virtnet_info *vi,
 334                              struct receive_queue *rq,
 335                              struct send_queue *sq,
 336                              struct xdp_buff *xdp)
 337 {
 338         struct page *page = virt_to_head_page(xdp->data);
 339         struct virtio_net_hdr_mrg_rxbuf *hdr;
 340         unsigned int num_sg, len;
 341         void *xdp_sent;
 342         int err;
 343
 344         /* Free up any pending old buffers before queueing new ones. */
 345         while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 346                 struct page *sent_page = virt_to_head_page(xdp_sent);
 347
 348                 if (vi->mergeable_rx_bufs)
 349                         put_page(sent_page);
 350                 else
 351                         give_pages(rq, sent_page);
 352         }
 353
 354         /* Zero header and leave csum up to XDP layers */
 355         hdr = xdp->data;
 356         memset(hdr, 0, vi->hdr_len);
 357
 358         num_sg = 1;
 359         sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
 360         err = virtqueue_add_outbuf(sq->vq, sq->sg, num_sg,
 361                                    xdp->data, GFP_ATOMIC);
 362         if (unlikely(err)) {
 363                 if (vi->mergeable_rx_bufs)
 364                         put_page(page);
 365                 else
 366                         give_pages(rq, page);
 367                 return; // On error abort to avoid unnecessary kick
 368         } else if (!vi->mergeable_rx_bufs) {
 369                 /* If not mergeable bufs must be big packets so cleanup pages */
 370                 give_pages(rq, (struct page *)page->private);
 371                 page->private = 0;
 372         }
 373
 374         virtqueue_kick(sq->vq);
 375 }
 376
 377 static u32 do_xdp_prog(struct virtnet_info *vi,
 378                        struct receive_queue *rq,
 379                        struct bpf_prog *xdp_prog,
 380                        struct page *page, int offset, int len)
 381 {
 382         int hdr_padded_len;
 383         struct xdp_buff xdp;
 384         unsigned int qp;
 385         u32 act;
 386         u8 *buf;
 387
 388         buf = page_address(page) + offset;
 389
 390         if (vi->mergeable_rx_bufs)
 391                 hdr_padded_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 392         else
 393                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 394
 395         xdp.data = buf + hdr_padded_len;
 396         xdp.data_end = xdp.data + (len - vi->hdr_len);
 397
 398         act = bpf_prog_run_xdp(xdp_prog, &xdp);
 399         switch (act) {
 400         case XDP_PASS:
 401                 return XDP_PASS;
 402         case XDP_TX:
 403                 qp = vi->curr_queue_pairs -
 404                         vi->xdp_queue_pairs +
 405                         smp_processor_id();
 406                 xdp.data = buf + (vi->mergeable_rx_bufs ? 0 : 4);
 407                 virtnet_xdp_xmit(vi, rq, &vi->sq[qp], &xdp);
 408                 return XDP_TX;
 409         default:
 410                 bpf_warn_invalid_xdp_action(act);
 411         case XDP_ABORTED:
 412         case XDP_DROP:
 413                 return XDP_DROP;
 414         }
 415 }
 416
 417 static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
 418 {
 419         struct sk_buff * skb = buf;
 420
 421         len -= vi->hdr_len;
 422         skb_trim(skb, len);
 423
 424         return skb;
 425 }
 426
 427 static struct sk_buff *receive_big(struct net_device *dev,
 428                                    struct virtnet_info *vi,
 429                                    struct receive_queue *rq,
 430                                    void *buf,
 431                                    unsigned int len)
 432 {
 433         struct bpf_prog *xdp_prog;
 434         struct page *page = buf;
 435         struct sk_buff *skb;
 436
 437         rcu_read_lock();
 438         xdp_prog = rcu_dereference(rq->xdp_prog);
 439         if (xdp_prog) {
 440                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 441                 u32 act;
 442
 443                 if (unlikely(hdr->hdr.gso_type))
 444                         goto err_xdp;
 445                 act = do_xdp_prog(vi, rq, xdp_prog, page, 0, len);
 446                 switch (act) {
 447                 case XDP_PASS:
 448                         break;
 449                 case XDP_TX:
 450                         rcu_read_unlock();
 451                         goto xdp_xmit;
 452                 case XDP_DROP:
 453                 default:
 454                         goto err_xdp;
 455                 }
 456         }
 457         rcu_read_unlock();
 458
 459         skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 460         if (unlikely(!skb))
 461                 goto err;
 462
 463         return skb;
 464
 465 err_xdp:
 466         rcu_read_unlock();
 467 err:
 468         dev->stats.rx_dropped++;
 469         give_pages(rq, page);
 470 xdp_xmit:
 471         return NULL;
 472 }
 473
 474 /* The conditions to enable XDP should preclude the underlying device from
 475  * sending packets across multiple buffers (num_buf > 1). However per spec
 476  * it does not appear to be illegal to do so but rather just against convention.
 477  * So in order to avoid making a system unresponsive the packets are pushed
 478  * into a page and the XDP program is run. This will be extremely slow and we
 479  * push a warning to the user to fix this as soon as possible. Fixing this may
 480  * require resolving the underlying hardware to determine why multiple buffers
 481  * are being received or simply loading the XDP program in the ingress stack
 482  * after the skb is built because there is no advantage to running it here
 483  * anymore.
 484  */
 485 static struct page *xdp_linearize_page(struct receive_queue *rq,
 486                                        u16 *num_buf,
 487                                        struct page *p,
 488                                        int offset,
 489                                        unsigned int *len)
 490 {
 491         struct page *page = alloc_page(GFP_ATOMIC);
 492         unsigned int page_off = 0;
 493
 494         if (!page)
 495                 return NULL;
 496
 497         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 498         page_off += *len;
 499
 500         while (--*num_buf) {
 501                 unsigned int buflen;
 502                 unsigned long ctx;
 503                 void *buf;
 504                 int off;
 505
 506                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &buflen);
 507                 if (unlikely(!ctx))
 508                         goto err_buf;
 509
 510                 buf = mergeable_ctx_to_buf_address(ctx);
 511                 p = virt_to_head_page(buf);
 512                 off = buf - page_address(p);
 513
 514                 /* guard against a misconfigured or uncooperative backend that
 515                  * is sending packet larger than the MTU.
 516                  */
 517                 if ((page_off + buflen) > PAGE_SIZE) {
 518                         put_page(p);
 519                         goto err_buf;
 520                 }
 521
 522                 memcpy(page_address(page) + page_off,
 523                        page_address(p) + off, buflen);
 524                 page_off += buflen;
 525                 put_page(p);
 526         }
 527
 528         *len = page_off;
 529         return page;
 530 err_buf:
 531         __free_pages(page, 0);
 532         return NULL;
 533 }
 534
 535 static struct sk_buff *receive_mergeable(struct net_device *dev,
 536                                          struct virtnet_info *vi,
 537                                          struct receive_queue *rq,
 538                                          unsigned long ctx,
 539                                          unsigned int len)
 540 {
 541         void *buf = mergeable_ctx_to_buf_address(ctx);
 542         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 543         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 544         struct page *page = virt_to_head_page(buf);
 545         int offset = buf - page_address(page);
 546         struct sk_buff *head_skb, *curr_skb;
 547         struct bpf_prog *xdp_prog;
 548         unsigned int truesize;
 549
 550         head_skb = NULL;
 551
 552         rcu_read_lock();
 553         xdp_prog = rcu_dereference(rq->xdp_prog);
 554         if (xdp_prog) {
 555                 struct page *xdp_page;
 556                 u32 act;
 557
 558                 /* This happens when rx buffer size is underestimated */
 559                 if (unlikely(num_buf > 1)) {
 560                         /* linearize data for XDP */
 561                         xdp_page = xdp_linearize_page(rq, &num_buf,
 562                                                       page, offset, &len);
 563                         if (!xdp_page)
 564                                 goto err_xdp;
 565                         offset = 0;
 566                 } else {
 567                         xdp_page = page;
 568                 }
 569
 570                 /* Transient failure which in theory could occur if
 571                  * in-flight packets from before XDP was enabled reach
 572                  * the receive path after XDP is loaded. In practice I
 573                  * was not able to create this condition.
 574                  */
 575                 if (unlikely(hdr->hdr.gso_type))
 576                         goto err_xdp;
 577
 578                 act = do_xdp_prog(vi, rq, xdp_prog, xdp_page, offset, len);
 579                 switch (act) {
 580                 case XDP_PASS:
 581                         /* We can only create skb based on xdp_page. */
 582                         if (unlikely(xdp_page != page)) {
 583                                 rcu_read_unlock();
 584                                 put_page(page);
 585                                 head_skb = page_to_skb(vi, rq, xdp_page,
 586                                                        0, len, PAGE_SIZE);
 587                                 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 588                                 return head_skb;
 589                         }
 590                         break;
 591                 case XDP_TX:
 592                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 593                         if (unlikely(xdp_page != page))
 594                                 goto err_xdp;
 595                         rcu_read_unlock();
 596                         goto xdp_xmit;
 597                 case XDP_DROP:
 598                 default:
 599                         if (unlikely(xdp_page != page))
 600                                 __free_pages(xdp_page, 0);
 601                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 602                         goto err_xdp;
 603                 }
 604         }
 605         rcu_read_unlock();
 606
 607         truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 608         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
 609         curr_skb = head_skb;
 610
 611         if (unlikely(!curr_skb))
 612                 goto err_skb;
 613         while (--num_buf) {
 614                 int num_skb_frags;
 615
 616                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 617                 if (unlikely(!ctx)) {
 618                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 619                                  dev->name, num_buf,
 620                                  virtio16_to_cpu(vi->vdev,
 621                                                  hdr->num_buffers));
 622                         dev->stats.rx_length_errors++;
 623                         goto err_buf;
 624                 }
 625
 626                 buf = mergeable_ctx_to_buf_address(ctx);
 627                 page = virt_to_head_page(buf);
 628
 629                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 630                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 631                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 632
 633                         if (unlikely(!nskb))
 634                                 goto err_skb;
 635                         if (curr_skb == head_skb)
 636                                 skb_shinfo(curr_skb)->frag_list = nskb;
 637                         else
 638                                 curr_skb->next = nskb;
 639                         curr_skb = nskb;
 640                         head_skb->truesize += nskb->truesize;
 641                         num_skb_frags = 0;
 642                 }
 643                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 644                 if (curr_skb != head_skb) {
 645                         head_skb->data_len += len;
 646                         head_skb->len += len;
 647                         head_skb->truesize += truesize;
 648                 }
 649                 offset = buf - page_address(page);
 650                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 651                         put_page(page);
 652                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 653                                              len, truesize);
 654                 } else {
 655                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 656                                         offset, len, truesize);
 657                 }
 658         }
 659
 660         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 661         return head_skb;
 662
 663 err_xdp:
 664         rcu_read_unlock();
 665 err_skb:
 666         put_page(page);
 667         while (--num_buf) {
 668                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 669                 if (unlikely(!ctx)) {
 670                         pr_debug("%s: rx error: %d buffers missing\n",
 671                                  dev->name, num_buf);
 672                         dev->stats.rx_length_errors++;
 673                         break;
 674                 }
 675                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 676                 put_page(page);
 677         }
 678 err_buf:
 679         dev->stats.rx_dropped++;
 680         dev_kfree_skb(head_skb);
 681 xdp_xmit:
 682         return NULL;
 683 }
 684
 685 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 686                         void *buf, unsigned int len)
 687 {
 688         struct net_device *dev = vi->dev;
 689         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 690         struct sk_buff *skb;
 691         struct virtio_net_hdr_mrg_rxbuf *hdr;
 692
 693         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 694                 pr_debug("%s: short packet %i\n", dev->name, len);
 695                 dev->stats.rx_length_errors++;
 696                 if (vi->mergeable_rx_bufs) {
 697                         unsigned long ctx = (unsigned long)buf;
 698                         void *base = mergeable_ctx_to_buf_address(ctx);
 699                         put_page(virt_to_head_page(base));
 700                 } else if (vi->big_packets) {
 701                         give_pages(rq, buf);
 702                 } else {
 703                         dev_kfree_skb(buf);
 704                 }
 705                 return;
 706         }
 707
 708         if (vi->mergeable_rx_bufs)
 709                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
 710         else if (vi->big_packets)
 711                 skb = receive_big(dev, vi, rq, buf, len);
 712         else
 713                 skb = receive_small(vi, buf, len);
 714
 715         if (unlikely(!skb))
 716                 return;
 717
 718         hdr = skb_vnet_hdr(skb);
 719
 720         u64_stats_update_begin(&stats->rx_syncp);
 721         stats->rx_bytes += skb->len;
 722         stats->rx_packets++;
 723         u64_stats_update_end(&stats->rx_syncp);
 724
 725         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 726                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 727
 728         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 729                                   virtio_is_little_endian(vi->vdev))) {
 730                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 731                                      dev->name, hdr->hdr.gso_type,
 732                                      hdr->hdr.gso_size);
 733                 goto frame_err;
 734         }
 735
 736         skb->protocol = eth_type_trans(skb, dev);
 737         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 738                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 739
 740         napi_gro_receive(&rq->napi, skb);
 741         return;
 742
 743 frame_err:
 744         dev->stats.rx_frame_errors++;
 745         dev_kfree_skb(skb);
 746 }
 747
 748 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 749                              gfp_t gfp)
 750 {
 751         struct sk_buff *skb;
 752         struct virtio_net_hdr_mrg_rxbuf *hdr;
 753         int err;
 754
 755         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 756         if (unlikely(!skb))
 757                 return -ENOMEM;
 758
 759         skb_put(skb, GOOD_PACKET_LEN);
 760
 761         hdr = skb_vnet_hdr(skb);
 762         sg_init_table(rq->sg, 2);
 763         sg_set_buf(rq->sg, hdr, vi->hdr_len);
 764         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 765
 766         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 767         if (err < 0)
 768                 dev_kfree_skb(skb);
 769
 770         return err;
 771 }
 772
 773 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 774                            gfp_t gfp)
 775 {
 776         struct page *first, *list = NULL;
 777         char *p;
 778         int i, err, offset;
 779
 780         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 781
 782         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 783         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 784                 first = get_a_page(rq, gfp);
 785                 if (!first) {
 786                         if (list)
 787                                 give_pages(rq, list);
 788                         return -ENOMEM;
 789                 }
 790                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 791
 792                 /* chain new page in list head to match sg */
 793                 first->private = (unsigned long)list;
 794                 list = first;
 795         }
 796
 797         first = get_a_page(rq, gfp);
 798         if (!first) {
 799                 give_pages(rq, list);
 800                 return -ENOMEM;
 801         }
 802         p = page_address(first);
 803
 804         /* rq->sg[0], rq->sg[1] share the same page */
 805         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 806         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 807
 808         /* rq->sg[1] for data packet, from offset */
 809         offset = sizeof(struct padded_vnet_hdr);
 810         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 811
 812         /* chain first in list head */
 813         first->private = (unsigned long)list;
 814         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 815                                   first, gfp);
 816         if (err < 0)
 817                 give_pages(rq, first);
 818
 819         return err;
 820 }
 821
 822 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
 823 {
 824         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 825         unsigned int len;
 826
 827         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
 828                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 829         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 830 }
 831
 832 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 833 {
 834         struct page_frag *alloc_frag = &rq->alloc_frag;
 835         char *buf;
 836         unsigned long ctx;
 837         int err;
 838         unsigned int len, hole;
 839
 840         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 841         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 842                 return -ENOMEM;
 843
 844         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 845         ctx = mergeable_buf_to_ctx(buf, len);
 846         get_page(alloc_frag->page);
 847         alloc_frag->offset += len;
 848         hole = alloc_frag->size - alloc_frag->offset;
 849         if (hole < len) {
 850                 /* To avoid internal fragmentation, if there is very likely not
 851                  * enough space for another buffer, add the remaining space to
 852                  * the current buffer. This extra space is not included in
 853                  * the truesize stored in ctx.
 854                  */
 855                 len += hole;
 856                 alloc_frag->offset += hole;
 857         }
 858
 859         sg_init_one(rq->sg, buf, len);
 860         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 861         if (err < 0)
 862                 put_page(virt_to_head_page(buf));
 863
 864         return err;
 865 }
 866
 867 /*
 868  * Returns false if we couldn't fill entirely (OOM).
 869  *
 870  * Normally run in the receive path, but can also be run from ndo_open
 871  * before we're receiving packets, or from refill_work which is
 872  * careful to disable receiving (using napi_disable).
 873  */
 874 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 875                           gfp_t gfp)
 876 {
 877         int err;
 878         bool oom;
 879
 880         gfp |= __GFP_COLD;
 881         do {
 882                 if (vi->mergeable_rx_bufs)
 883                         err = add_recvbuf_mergeable(rq, gfp);
 884                 else if (vi->big_packets)
 885                         err = add_recvbuf_big(vi, rq, gfp);
 886                 else
 887                         err = add_recvbuf_small(vi, rq, gfp);
 888
 889                 oom = err == -ENOMEM;
 890                 if (err)
 891                         break;
 892         } while (rq->vq->num_free);
 893         virtqueue_kick(rq->vq);
 894         return !oom;
 895 }
 896
 897 static void skb_recv_done(struct virtqueue *rvq)
 898 {
 899         struct virtnet_info *vi = rvq->vdev->priv;
 900         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 901
 902         /* Schedule NAPI, Suppress further interrupts if successful. */
 903         if (napi_schedule_prep(&rq->napi)) {
 904                 virtqueue_disable_cb(rvq);
 905                 __napi_schedule(&rq->napi);
 906         }
 907 }
 908
 909 static void virtnet_napi_enable(struct receive_queue *rq)
 910 {
 911         napi_enable(&rq->napi);
 912
 913         /* If all buffers were filled by other side before we napi_enabled, we
 914          * won't get another interrupt, so process any outstanding packets
 915          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 916          * We synchronize against interrupts via NAPI_STATE_SCHED */
 917         if (napi_schedule_prep(&rq->napi)) {
 918                 virtqueue_disable_cb(rq->vq);
 919                 local_bh_disable();
 920                 __napi_schedule(&rq->napi);
 921                 local_bh_enable();
 922         }
 923 }
 924
 925 static void refill_work(struct work_struct *work)
 926 {
 927         struct virtnet_info *vi =
 928                 container_of(work, struct virtnet_info, refill.work);
 929         bool still_empty;
 930         int i;
 931
 932         for (i = 0; i < vi->curr_queue_pairs; i++) {
 933                 struct receive_queue *rq = &vi->rq[i];
 934
 935                 napi_disable(&rq->napi);
 936                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
 937                 virtnet_napi_enable(rq);
 938
 939                 /* In theory, this can happen: if we don't get any buffers in
 940                  * we will *never* try to fill again.
 941                  */
 942                 if (still_empty)
 943                         schedule_delayed_work(&vi->refill, HZ/2);
 944         }
 945 }
 946
 947 static int virtnet_receive(struct receive_queue *rq, int budget)
 948 {
 949         struct virtnet_info *vi = rq->vq->vdev->priv;
 950         unsigned int len, received = 0;
 951         void *buf;
 952
 953         while (received < budget &&
 954                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 955                 receive_buf(vi, rq, buf, len);
 956                 received++;
 957         }
 958
 959         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 960                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 961                         schedule_delayed_work(&vi->refill, 0);
 962         }
 963
 964         return received;
 965 }
 966
 967 static int virtnet_poll(struct napi_struct *napi, int budget)
 968 {
 969         struct receive_queue *rq =
 970                 container_of(napi, struct receive_queue, napi);
 971         unsigned int r, received;
 972
 973         received = virtnet_receive(rq, budget);
 974
 975         /* Out of packets? */
 976         if (received < budget) {
 977                 r = virtqueue_enable_cb_prepare(rq->vq);
 978                 napi_complete_done(napi, received);
 979                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 980                     napi_schedule_prep(napi)) {
 981                         virtqueue_disable_cb(rq->vq);
 982                         __napi_schedule(napi);
 983                 }
 984         }
 985
 986         return received;
 987 }
 988
 989 #ifdef CONFIG_NET_RX_BUSY_POLL
 990 /* must be called with local_bh_disable()d */
 991 static int virtnet_busy_poll(struct napi_struct *napi)
 992 {
 993         struct receive_queue *rq =
 994                 container_of(napi, struct receive_queue, napi);
 995         struct virtnet_info *vi = rq->vq->vdev->priv;
 996         int r, received = 0, budget = 4;
 997
 998         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
 999                 return LL_FLUSH_FAILED;
1000
1001         if (!napi_schedule_prep(napi))
1002                 return LL_FLUSH_BUSY;
1003
1004         virtqueue_disable_cb(rq->vq);
1005
1006 again:
1007         received += virtnet_receive(rq, budget);
1008
1009         r = virtqueue_enable_cb_prepare(rq->vq);
1010         clear_bit(NAPI_STATE_SCHED, &napi->state);
1011         if (unlikely(virtqueue_poll(rq->vq, r)) &&
1012             napi_schedule_prep(napi)) {
1013                 virtqueue_disable_cb(rq->vq);
1014                 if (received < budget) {
1015                         budget -= received;
1016                         goto again;
1017                 } else {
1018                         __napi_schedule(napi);
1019                 }
1020         }
1021
1022         return received;
1023 }
1024 #endif  /* CONFIG_NET_RX_BUSY_POLL */
1025
1026 static int virtnet_open(struct net_device *dev)
1027 {
1028         struct virtnet_info *vi = netdev_priv(dev);
1029         int i;
1030
1031         for (i = 0; i < vi->max_queue_pairs; i++) {
1032                 if (i < vi->curr_queue_pairs)
1033                         /* Make sure we have some buffers: if oom use wq. */
1034                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1035                                 schedule_delayed_work(&vi->refill, 0);
1036                 virtnet_napi_enable(&vi->rq[i]);
1037         }
1038
1039         return 0;
1040 }
1041
1042 static void free_old_xmit_skbs(struct send_queue *sq)
1043 {
1044         struct sk_buff *skb;
1045         unsigned int len;
1046         struct virtnet_info *vi = sq->vq->vdev->priv;
1047         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1048
1049         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1050                 pr_debug("Sent skb %p\n", skb);
1051
1052                 u64_stats_update_begin(&stats->tx_syncp);
1053                 stats->tx_bytes += skb->len;
1054                 stats->tx_packets++;
1055                 u64_stats_update_end(&stats->tx_syncp);
1056
1057                 dev_kfree_skb_any(skb);
1058         }
1059 }
1060
1061 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1062 {
1063         struct virtio_net_hdr_mrg_rxbuf *hdr;
1064         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1065         struct virtnet_info *vi = sq->vq->vdev->priv;
1066         unsigned num_sg;
1067         unsigned hdr_len = vi->hdr_len;
1068         bool can_push;
1069
1070         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1071
1072         can_push = vi->any_header_sg &&
1073                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1074                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1075         /* Even if we can, don't push here yet as this would skew
1076          * csum_start offset below. */
1077         if (can_push)
1078                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1079         else
1080                 hdr = skb_vnet_hdr(skb);
1081
1082         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1083                                     virtio_is_little_endian(vi->vdev)))
1084                 BUG();
1085
1086         if (vi->mergeable_rx_bufs)
1087                 hdr->num_buffers = 0;
1088
1089         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1090         if (can_push) {
1091                 __skb_push(skb, hdr_len);
1092                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1093                 /* Pull header back to avoid skew in tx bytes calculations. */
1094                 __skb_pull(skb, hdr_len);
1095         } else {
1096                 sg_set_buf(sq->sg, hdr, hdr_len);
1097                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
1098         }
1099         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1100 }
1101
1102 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1103 {
1104         struct virtnet_info *vi = netdev_priv(dev);
1105         int qnum = skb_get_queue_mapping(skb);
1106         struct send_queue *sq = &vi->sq[qnum];
1107         int err;
1108         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1109         bool kick = !skb->xmit_more;
1110
1111         /* Free up any pending old buffers before queueing new ones. */
1112         free_old_xmit_skbs(sq);
1113
1114         /* timestamp packet in software */
1115         skb_tx_timestamp(skb);
1116
1117         /* Try to transmit */
1118         err = xmit_skb(sq, skb);
1119
1120         /* This should not happen! */
1121         if (unlikely(err)) {
1122                 dev->stats.tx_fifo_errors++;
1123                 if (net_ratelimit())
1124                         dev_warn(&dev->dev,
1125                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1126                 dev->stats.tx_dropped++;
1127                 dev_kfree_skb_any(skb);
1128                 return NETDEV_TX_OK;
1129         }
1130
1131         /* Don't wait up for transmitted skbs to be freed. */
1132         skb_orphan(skb);
1133         nf_reset(skb);
1134
1135         /* If running out of space, stop queue to avoid getting packets that we
1136          * are then unable to transmit.
1137          * An alternative would be to force queuing layer to requeue the skb by
1138          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1139          * returned in a normal path of operation: it means that driver is not
1140          * maintaining the TX queue stop/start state properly, and causes
1141          * the stack to do a non-trivial amount of useless work.
1142          * Since most packets only take 1 or 2 ring slots, stopping the queue
1143          * early means 16 slots are typically wasted.
1144          */
1145         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1146                 netif_stop_subqueue(dev, qnum);
1147                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1148                         /* More just got used, free them then recheck. */
1149                         free_old_xmit_skbs(sq);
1150                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1151                                 netif_start_subqueue(dev, qnum);
1152                                 virtqueue_disable_cb(sq->vq);
1153                         }
1154                 }
1155         }
1156
1157         if (kick || netif_xmit_stopped(txq))
1158                 virtqueue_kick(sq->vq);
1159
1160         return NETDEV_TX_OK;
1161 }
1162
1163 /*
1164  * Send command via the control virtqueue and check status.  Commands
1165  * supported by the hypervisor, as indicated by feature bits, should
1166  * never fail unless improperly formatted.
1167  */
1168 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1169                                  struct scatterlist *out)
1170 {
1171         struct scatterlist *sgs[4], hdr, stat;
1172         unsigned out_num = 0, tmp;
1173
1174         /* Caller should know better */
1175         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1176
1177         vi->ctrl_status = ~0;
1178         vi->ctrl_hdr.class = class;
1179         vi->ctrl_hdr.cmd = cmd;
1180         /* Add header */
1181         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1182         sgs[out_num++] = &hdr;
1183
1184         if (out)
1185                 sgs[out_num++] = out;
1186
1187         /* Add return status. */
1188         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1189         sgs[out_num] = &stat;
1190
1191         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1192         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1193
1194         if (unlikely(!virtqueue_kick(vi->cvq)))
1195                 return vi->ctrl_status == VIRTIO_NET_OK;
1196
1197         /* Spin for a response, the kick causes an ioport write, trapping
1198          * into the hypervisor, so the request should be handled immediately.
1199          */
1200         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1201                !virtqueue_is_broken(vi->cvq))
1202                 cpu_relax();
1203
1204         return vi->ctrl_status == VIRTIO_NET_OK;
1205 }
1206
1207 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1208 {
1209         struct virtnet_info *vi = netdev_priv(dev);
1210         struct virtio_device *vdev = vi->vdev;
1211         int ret;
1212         struct sockaddr *addr;
1213         struct scatterlist sg;
1214
1215         addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1216         if (!addr)
1217                 return -ENOMEM;
1218         memcpy(addr, p, sizeof(*addr));
1219
1220         ret = eth_prepare_mac_addr_change(dev, addr);
1221         if (ret)
1222                 goto out;
1223
1224         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1225                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1226                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1227                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1228                         dev_warn(&vdev->dev,
1229                                  "Failed to set mac address by vq command.\n");
1230                         ret = -EINVAL;
1231                         goto out;
1232                 }
1233         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1234                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1235                 unsigned int i;
1236
1237                 /* Naturally, this has an atomicity problem. */
1238                 for (i = 0; i < dev->addr_len; i++)
1239                         virtio_cwrite8(vdev,
1240                                        offsetof(struct virtio_net_config, mac) +
1241                                        i, addr->sa_data[i]);
1242         }
1243
1244         eth_commit_mac_addr_change(dev, p);
1245         ret = 0;
1246
1247 out:
1248         kfree(addr);
1249         return ret;
1250 }
1251
1252 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1253                                                struct rtnl_link_stats64 *tot)
1254 {
1255         struct virtnet_info *vi = netdev_priv(dev);
1256         int cpu;
1257         unsigned int start;
1258
1259         for_each_possible_cpu(cpu) {
1260                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1261                 u64 tpackets, tbytes, rpackets, rbytes;
1262
1263                 do {
1264                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1265                         tpackets = stats->tx_packets;
1266                         tbytes   = stats->tx_bytes;
1267                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1268
1269                 do {
1270                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1271                         rpackets = stats->rx_packets;
1272                         rbytes   = stats->rx_bytes;
1273                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1274
1275                 tot->rx_packets += rpackets;
1276                 tot->tx_packets += tpackets;
1277                 tot->rx_bytes   += rbytes;
1278                 tot->tx_bytes   += tbytes;
1279         }
1280
1281         tot->tx_dropped = dev->stats.tx_dropped;
1282         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1283         tot->rx_dropped = dev->stats.rx_dropped;
1284         tot->rx_length_errors = dev->stats.rx_length_errors;
1285         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1286
1287         return tot;
1288 }
1289
1290 #ifdef CONFIG_NET_POLL_CONTROLLER
1291 static void virtnet_netpoll(struct net_device *dev)
1292 {
1293         struct virtnet_info *vi = netdev_priv(dev);
1294         int i;
1295
1296         for (i = 0; i < vi->curr_queue_pairs; i++)
1297                 napi_schedule(&vi->rq[i].napi);
1298 }
1299 #endif
1300
1301 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1302 {
1303         rtnl_lock();
1304         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1305                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1306                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1307         rtnl_unlock();
1308 }
1309
1310 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1311 {
1312         struct scatterlist sg;
1313         struct net_device *dev = vi->dev;
1314
1315         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1316                 return 0;
1317
1318         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1319         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1320
1321         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1322                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1323                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1324                          queue_pairs);
1325                 return -EINVAL;
1326         } else {
1327                 vi->curr_queue_pairs = queue_pairs;
1328                 /* virtnet_open() will refill when device is going to up. */
1329                 if (dev->flags & IFF_UP)
1330                         schedule_delayed_work(&vi->refill, 0);
1331         }
1332
1333         return 0;
1334 }
1335
1336 static int virtnet_close(struct net_device *dev)
1337 {
1338         struct virtnet_info *vi = netdev_priv(dev);
1339         int i;
1340
1341         /* Make sure refill_work doesn't re-enable napi! */
1342         cancel_delayed_work_sync(&vi->refill);
1343
1344         for (i = 0; i < vi->max_queue_pairs; i++)
1345                 napi_disable(&vi->rq[i].napi);
1346
1347         return 0;
1348 }
1349
1350 static void virtnet_set_rx_mode(struct net_device *dev)
1351 {
1352         struct virtnet_info *vi = netdev_priv(dev);
1353         struct scatterlist sg[2];
1354         struct virtio_net_ctrl_mac *mac_data;
1355         struct netdev_hw_addr *ha;
1356         int uc_count;
1357         int mc_count;
1358         void *buf;
1359         int i;
1360
1361         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1362         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1363                 return;
1364
1365         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1366         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1367
1368         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1369
1370         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1371                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1372                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1373                          vi->ctrl_promisc ? "en" : "dis");
1374
1375         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1376
1377         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1378                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1379                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1380                          vi->ctrl_allmulti ? "en" : "dis");
1381
1382         uc_count = netdev_uc_count(dev);
1383         mc_count = netdev_mc_count(dev);
1384         /* MAC filter - use one buffer for both lists */
1385         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1386                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1387         mac_data = buf;
1388         if (!buf)
1389                 return;
1390
1391         sg_init_table(sg, 2);
1392
1393         /* Store the unicast list and count in the front of the buffer */
1394         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1395         i = 0;
1396         netdev_for_each_uc_addr(ha, dev)
1397                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1398
1399         sg_set_buf(&sg[0], mac_data,
1400                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1401
1402         /* multicast list and count fill the end */
1403         mac_data = (void *)&mac_data->macs[uc_count][0];
1404
1405         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1406         i = 0;
1407         netdev_for_each_mc_addr(ha, dev)
1408                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1409
1410         sg_set_buf(&sg[1], mac_data,
1411                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1412
1413         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1414                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1415                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1416
1417         kfree(buf);
1418 }
1419
1420 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1421                                    __be16 proto, u16 vid)
1422 {
1423         struct virtnet_info *vi = netdev_priv(dev);
1424         struct scatterlist sg;
1425
1426         vi->ctrl_vid = vid;
1427         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1428
1429         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1430                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1431                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1432         return 0;
1433 }
1434
1435 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1436                                     __be16 proto, u16 vid)
1437 {
1438         struct virtnet_info *vi = netdev_priv(dev);
1439         struct scatterlist sg;
1440
1441         vi->ctrl_vid = vid;
1442         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1443
1444         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1445                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1446                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1447         return 0;
1448 }
1449
1450 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1451 {
1452         int i;
1453
1454         if (vi->affinity_hint_set) {
1455                 for (i = 0; i < vi->max_queue_pairs; i++) {
1456                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1457                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1458                 }
1459
1460                 vi->affinity_hint_set = false;
1461         }
1462 }
1463
1464 static void virtnet_set_affinity(struct virtnet_info *vi)
1465 {
1466         int i;
1467         int cpu;
1468
1469         /* In multiqueue mode, when the number of cpu is equal to the number of
1470          * queue pairs, we let the queue pairs to be private to one cpu by
1471          * setting the affinity hint to eliminate the contention.
1472          */
1473         if (vi->curr_queue_pairs == 1 ||
1474             vi->max_queue_pairs != num_online_cpus()) {
1475                 virtnet_clean_affinity(vi, -1);
1476                 return;
1477         }
1478
1479         i = 0;
1480         for_each_online_cpu(cpu) {
1481                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1482                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1483                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1484                 i++;
1485         }
1486
1487         vi->affinity_hint_set = true;
1488 }
1489
1490 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1491 {
1492         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1493                                                    node);
1494         virtnet_set_affinity(vi);
1495         return 0;
1496 }
1497
1498 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1499 {
1500         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1501                                                    node_dead);
1502         virtnet_set_affinity(vi);
1503         return 0;
1504 }
1505
1506 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1507 {
1508         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1509                                                    node);
1510
1511         virtnet_clean_affinity(vi, cpu);
1512         return 0;
1513 }
1514
1515 static enum cpuhp_state virtionet_online;
1516
1517 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1518 {
1519         int ret;
1520
1521         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1522         if (ret)
1523                 return ret;
1524         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1525                                                &vi->node_dead);
1526         if (!ret)
1527                 return ret;
1528         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1529         return ret;
1530 }
1531
1532 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1533 {
1534         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1535         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1536                                             &vi->node_dead);
1537 }
1538
1539 static void virtnet_get_ringparam(struct net_device *dev,
1540                                 struct ethtool_ringparam *ring)
1541 {
1542         struct virtnet_info *vi = netdev_priv(dev);
1543
1544         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1545         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1546         ring->rx_pending = ring->rx_max_pending;
1547         ring->tx_pending = ring->tx_max_pending;
1548 }
1549
1550
1551 static void virtnet_get_drvinfo(struct net_device *dev,
1552                                 struct ethtool_drvinfo *info)
1553 {
1554         struct virtnet_info *vi = netdev_priv(dev);
1555         struct virtio_device *vdev = vi->vdev;
1556
1557         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1558         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1559         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1560
1561 }
1562
1563 /* TODO: Eliminate OOO packets during switching */
1564 static int virtnet_set_channels(struct net_device *dev,
1565                                 struct ethtool_channels *channels)
1566 {
1567         struct virtnet_info *vi = netdev_priv(dev);
1568         u16 queue_pairs = channels->combined_count;
1569         int err;
1570
1571         /* We don't support separate rx/tx channels.
1572          * We don't allow setting 'other' channels.
1573          */
1574         if (channels->rx_count || channels->tx_count || channels->other_count)
1575                 return -EINVAL;
1576
1577         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1578                 return -EINVAL;
1579
1580         /* For now we don't support modifying channels while XDP is loaded
1581          * also when XDP is loaded all RX queues have XDP programs so we only
1582          * need to check a single RX queue.
1583          */
1584         if (vi->rq[0].xdp_prog)
1585                 return -EINVAL;
1586
1587         get_online_cpus();
1588         err = virtnet_set_queues(vi, queue_pairs);
1589         if (!err) {
1590                 netif_set_real_num_tx_queues(dev, queue_pairs);
1591                 netif_set_real_num_rx_queues(dev, queue_pairs);
1592
1593                 virtnet_set_affinity(vi);
1594         }
1595         put_online_cpus();
1596
1597         return err;
1598 }
1599
1600 static void virtnet_get_channels(struct net_device *dev,
1601                                  struct ethtool_channels *channels)
1602 {
1603         struct virtnet_info *vi = netdev_priv(dev);
1604
1605         channels->combined_count = vi->curr_queue_pairs;
1606         channels->max_combined = vi->max_queue_pairs;
1607         channels->max_other = 0;
1608         channels->rx_count = 0;
1609         channels->tx_count = 0;
1610         channels->other_count = 0;
1611 }
1612
1613 /* Check if the user is trying to change anything besides speed/duplex */
1614 static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1615 {
1616         struct ethtool_cmd diff1 = *cmd;
1617         struct ethtool_cmd diff2 = {};
1618
1619         /* cmd is always set so we need to clear it, validate the port type
1620          * and also without autonegotiation we can ignore advertising
1621          */
1622         ethtool_cmd_speed_set(&diff1, 0);
1623         diff2.port = PORT_OTHER;
1624         diff1.advertising = 0;
1625         diff1.duplex = 0;
1626         diff1.cmd = 0;
1627
1628         return !memcmp(&diff1, &diff2, sizeof(diff1));
1629 }
1630
1631 static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1632 {
1633         struct virtnet_info *vi = netdev_priv(dev);
1634         u32 speed;
1635
1636         speed = ethtool_cmd_speed(cmd);
1637         /* don't allow custom speed and duplex */
1638         if (!ethtool_validate_speed(speed) ||
1639             !ethtool_validate_duplex(cmd->duplex) ||
1640             !virtnet_validate_ethtool_cmd(cmd))
1641                 return -EINVAL;
1642         vi->speed = speed;
1643         vi->duplex = cmd->duplex;
1644
1645         return 0;
1646 }
1647
1648 static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1649 {
1650         struct virtnet_info *vi = netdev_priv(dev);
1651
1652         ethtool_cmd_speed_set(cmd, vi->speed);
1653         cmd->duplex = vi->duplex;
1654         cmd->port = PORT_OTHER;
1655
1656         return 0;
1657 }
1658
1659 static void virtnet_init_settings(struct net_device *dev)
1660 {
1661         struct virtnet_info *vi = netdev_priv(dev);
1662
1663         vi->speed = SPEED_UNKNOWN;
1664         vi->duplex = DUPLEX_UNKNOWN;
1665 }
1666
1667 static const struct ethtool_ops virtnet_ethtool_ops = {
1668         .get_drvinfo = virtnet_get_drvinfo,
1669         .get_link = ethtool_op_get_link,
1670         .get_ringparam = virtnet_get_ringparam,
1671         .set_channels = virtnet_set_channels,
1672         .get_channels = virtnet_get_channels,
1673         .get_ts_info = ethtool_op_get_ts_info,
1674         .get_settings = virtnet_get_settings,
1675         .set_settings = virtnet_set_settings,
1676 };
1677
1678 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog)
1679 {
1680         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1681         struct virtnet_info *vi = netdev_priv(dev);
1682         struct bpf_prog *old_prog;
1683         u16 xdp_qp = 0, curr_qp;
1684         int i, err;
1685
1686         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1687             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1688             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
1689             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO)) {
1690                 netdev_warn(dev, "can't set XDP while host is implementing LRO, disable LRO first\n");
1691                 return -EOPNOTSUPP;
1692         }
1693
1694         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1695                 netdev_warn(dev, "XDP expects header/data in single page, any_header_sg required\n");
1696                 return -EINVAL;
1697         }
1698
1699         if (dev->mtu > max_sz) {
1700                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1701                 return -EINVAL;
1702         }
1703
1704         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1705         if (prog)
1706                 xdp_qp = nr_cpu_ids;
1707
1708         /* XDP requires extra queues for XDP_TX */
1709         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1710                 netdev_warn(dev, "request %i queues but max is %i\n",
1711                             curr_qp + xdp_qp, vi->max_queue_pairs);
1712                 return -ENOMEM;
1713         }
1714
1715         err = virtnet_set_queues(vi, curr_qp + xdp_qp);
1716         if (err) {
1717                 dev_warn(&dev->dev, "XDP Device queue allocation failure.\n");
1718                 return err;
1719         }
1720
1721         if (prog) {
1722                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1723                 if (IS_ERR(prog)) {
1724                         virtnet_set_queues(vi, curr_qp);
1725                         return PTR_ERR(prog);
1726                 }
1727         }
1728
1729         vi->xdp_queue_pairs = xdp_qp;
1730         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
1731
1732         for (i = 0; i < vi->max_queue_pairs; i++) {
1733                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1734                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
1735                 if (old_prog)
1736                         bpf_prog_put(old_prog);
1737         }
1738
1739         return 0;
1740 }
1741
1742 static bool virtnet_xdp_query(struct net_device *dev)
1743 {
1744         struct virtnet_info *vi = netdev_priv(dev);
1745         int i;
1746
1747         for (i = 0; i < vi->max_queue_pairs; i++) {
1748                 if (vi->rq[i].xdp_prog)
1749                         return true;
1750         }
1751         return false;
1752 }
1753
1754 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
1755 {
1756         switch (xdp->command) {
1757         case XDP_SETUP_PROG:
1758                 return virtnet_xdp_set(dev, xdp->prog);
1759         case XDP_QUERY_PROG:
1760                 xdp->prog_attached = virtnet_xdp_query(dev);
1761                 return 0;
1762         default:
1763                 return -EINVAL;
1764         }
1765 }
1766
1767 static const struct net_device_ops virtnet_netdev = {
1768         .ndo_open            = virtnet_open,
1769         .ndo_stop            = virtnet_close,
1770         .ndo_start_xmit      = start_xmit,
1771         .ndo_validate_addr   = eth_validate_addr,
1772         .ndo_set_mac_address = virtnet_set_mac_address,
1773         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1774         .ndo_get_stats64     = virtnet_stats,
1775         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1776         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1777 #ifdef CONFIG_NET_POLL_CONTROLLER
1778         .ndo_poll_controller = virtnet_netpoll,
1779 #endif
1780 #ifdef CONFIG_NET_RX_BUSY_POLL
1781         .ndo_busy_poll          = virtnet_busy_poll,
1782 #endif
1783         .ndo_xdp                = virtnet_xdp,
1784 };
1785
1786 static void virtnet_config_changed_work(struct work_struct *work)
1787 {
1788         struct virtnet_info *vi =
1789                 container_of(work, struct virtnet_info, config_work);
1790         u16 v;
1791
1792         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1793                                  struct virtio_net_config, status, &v) < 0)
1794                 return;
1795
1796         if (v & VIRTIO_NET_S_ANNOUNCE) {
1797                 netdev_notify_peers(vi->dev);
1798                 virtnet_ack_link_announce(vi);
1799         }
1800
1801         /* Ignore unknown (future) status bits */
1802         v &= VIRTIO_NET_S_LINK_UP;
1803
1804         if (vi->status == v)
1805                 return;
1806
1807         vi->status = v;
1808
1809         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1810                 netif_carrier_on(vi->dev);
1811                 netif_tx_wake_all_queues(vi->dev);
1812         } else {
1813                 netif_carrier_off(vi->dev);
1814                 netif_tx_stop_all_queues(vi->dev);
1815         }
1816 }
1817
1818 static void virtnet_config_changed(struct virtio_device *vdev)
1819 {
1820         struct virtnet_info *vi = vdev->priv;
1821
1822         schedule_work(&vi->config_work);
1823 }
1824
1825 static void virtnet_free_queues(struct virtnet_info *vi)
1826 {
1827         int i;
1828
1829         for (i = 0; i < vi->max_queue_pairs; i++) {
1830                 napi_hash_del(&vi->rq[i].napi);
1831                 netif_napi_del(&vi->rq[i].napi);
1832         }
1833
1834         /* We called napi_hash_del() before netif_napi_del(),
1835          * we need to respect an RCU grace period before freeing vi->rq
1836          */
1837         synchronize_net();
1838
1839         kfree(vi->rq);
1840         kfree(vi->sq);
1841 }
1842
1843 static void free_receive_bufs(struct virtnet_info *vi)
1844 {
1845         struct bpf_prog *old_prog;
1846         int i;
1847
1848         rtnl_lock();
1849         for (i = 0; i < vi->max_queue_pairs; i++) {
1850                 while (vi->rq[i].pages)
1851                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1852
1853                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1854                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
1855                 if (old_prog)
1856                         bpf_prog_put(old_prog);
1857         }
1858         rtnl_unlock();
1859 }
1860
1861 static void free_receive_page_frags(struct virtnet_info *vi)
1862 {
1863         int i;
1864         for (i = 0; i < vi->max_queue_pairs; i++)
1865                 if (vi->rq[i].alloc_frag.page)
1866                         put_page(vi->rq[i].alloc_frag.page);
1867 }
1868
1869 static bool is_xdp_queue(struct virtnet_info *vi, int q)
1870 {
1871         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1872                 return false;
1873         else if (q < vi->curr_queue_pairs)
1874                 return true;
1875         else
1876                 return false;
1877 }
1878
1879 static void free_unused_bufs(struct virtnet_info *vi)
1880 {
1881         void *buf;
1882         int i;
1883
1884         for (i = 0; i < vi->max_queue_pairs; i++) {
1885                 struct virtqueue *vq = vi->sq[i].vq;
1886                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1887                         if (!is_xdp_queue(vi, i))
1888                                 dev_kfree_skb(buf);
1889                         else
1890                                 put_page(virt_to_head_page(buf));
1891                 }
1892         }
1893
1894         for (i = 0; i < vi->max_queue_pairs; i++) {
1895                 struct virtqueue *vq = vi->rq[i].vq;
1896
1897                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1898                         if (vi->mergeable_rx_bufs) {
1899                                 unsigned long ctx = (unsigned long)buf;
1900                                 void *base = mergeable_ctx_to_buf_address(ctx);
1901                                 put_page(virt_to_head_page(base));
1902                         } else if (vi->big_packets) {
1903                                 give_pages(&vi->rq[i], buf);
1904                         } else {
1905                                 dev_kfree_skb(buf);
1906                         }
1907                 }
1908         }
1909 }
1910
1911 static void virtnet_del_vqs(struct virtnet_info *vi)
1912 {
1913         struct virtio_device *vdev = vi->vdev;
1914
1915         virtnet_clean_affinity(vi, -1);
1916
1917         vdev->config->del_vqs(vdev);
1918
1919         virtnet_free_queues(vi);
1920 }
1921
1922 static int virtnet_find_vqs(struct virtnet_info *vi)
1923 {
1924         vq_callback_t **callbacks;
1925         struct virtqueue **vqs;
1926         int ret = -ENOMEM;
1927         int i, total_vqs;
1928         const char **names;
1929
1930         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1931          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1932          * possible control vq.
1933          */
1934         total_vqs = vi->max_queue_pairs * 2 +
1935                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1936
1937         /* Allocate space for find_vqs parameters */
1938         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1939         if (!vqs)
1940                 goto err_vq;
1941         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1942         if (!callbacks)
1943                 goto err_callback;
1944         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1945         if (!names)
1946                 goto err_names;
1947
1948         /* Parameters for control virtqueue, if any */
1949         if (vi->has_cvq) {
1950                 callbacks[total_vqs - 1] = NULL;
1951                 names[total_vqs - 1] = "control";
1952         }
1953
1954         /* Allocate/initialize parameters for send/receive virtqueues */
1955         for (i = 0; i < vi->max_queue_pairs; i++) {
1956                 callbacks[rxq2vq(i)] = skb_recv_done;
1957                 callbacks[txq2vq(i)] = skb_xmit_done;
1958                 sprintf(vi->rq[i].name, "input.%d", i);
1959                 sprintf(vi->sq[i].name, "output.%d", i);
1960                 names[rxq2vq(i)] = vi->rq[i].name;
1961                 names[txq2vq(i)] = vi->sq[i].name;
1962         }
1963
1964         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1965                                          names);
1966         if (ret)
1967                 goto err_find;
1968
1969         if (vi->has_cvq) {
1970                 vi->cvq = vqs[total_vqs - 1];
1971                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1972                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1973         }
1974
1975         for (i = 0; i < vi->max_queue_pairs; i++) {
1976                 vi->rq[i].vq = vqs[rxq2vq(i)];
1977                 vi->sq[i].vq = vqs[txq2vq(i)];
1978         }
1979
1980         kfree(names);
1981         kfree(callbacks);
1982         kfree(vqs);
1983
1984         return 0;
1985
1986 err_find:
1987         kfree(names);
1988 err_names:
1989         kfree(callbacks);
1990 err_callback:
1991         kfree(vqs);
1992 err_vq:
1993         return ret;
1994 }
1995
1996 static int virtnet_alloc_queues(struct virtnet_info *vi)
1997 {
1998         int i;
1999
2000         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2001         if (!vi->sq)
2002                 goto err_sq;
2003         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2004         if (!vi->rq)
2005                 goto err_rq;
2006
2007         INIT_DELAYED_WORK(&vi->refill, refill_work);
2008         for (i = 0; i < vi->max_queue_pairs; i++) {
2009                 vi->rq[i].pages = NULL;
2010                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2011                                napi_weight);
2012
2013                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2014                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2015                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2016         }
2017
2018         return 0;
2019
2020 err_rq:
2021         kfree(vi->sq);
2022 err_sq:
2023         return -ENOMEM;
2024 }
2025
2026 static int init_vqs(struct virtnet_info *vi)
2027 {
2028         int ret;
2029
2030         /* Allocate send & receive queues */
2031         ret = virtnet_alloc_queues(vi);
2032         if (ret)
2033                 goto err;
2034
2035         ret = virtnet_find_vqs(vi);
2036         if (ret)
2037                 goto err_free;
2038
2039         get_online_cpus();
2040         virtnet_set_affinity(vi);
2041         put_online_cpus();
2042
2043         return 0;
2044
2045 err_free:
2046         virtnet_free_queues(vi);
2047 err:
2048         return ret;
2049 }
2050
2051 #ifdef CONFIG_SYSFS
2052 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2053                 struct rx_queue_attribute *attribute, char *buf)
2054 {
2055         struct virtnet_info *vi = netdev_priv(queue->dev);
2056         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2057         struct ewma_pkt_len *avg;
2058
2059         BUG_ON(queue_index >= vi->max_queue_pairs);
2060         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2061         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
2062 }
2063
2064 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2065         __ATTR_RO(mergeable_rx_buffer_size);
2066
2067 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2068         &mergeable_rx_buffer_size_attribute.attr,
2069         NULL
2070 };
2071
2072 static const struct attribute_group virtio_net_mrg_rx_group = {
2073         .name = "virtio_net",
2074         .attrs = virtio_net_mrg_rx_attrs
2075 };
2076 #endif
2077
2078 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2079                                     unsigned int fbit,
2080                                     const char *fname, const char *dname)
2081 {
2082         if (!virtio_has_feature(vdev, fbit))
2083                 return false;
2084
2085         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2086                 fname, dname);
2087
2088         return true;
2089 }
2090
2091 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2092         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2093
2094 static bool virtnet_validate_features(struct virtio_device *vdev)
2095 {
2096         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2097             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2098                              "VIRTIO_NET_F_CTRL_VQ") ||
2099              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2100                              "VIRTIO_NET_F_CTRL_VQ") ||
2101              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2102                              "VIRTIO_NET_F_CTRL_VQ") ||
2103              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2104              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2105                              "VIRTIO_NET_F_CTRL_VQ"))) {
2106                 return false;
2107         }
2108
2109         return true;
2110 }
2111
2112 #define MIN_MTU ETH_MIN_MTU
2113 #define MAX_MTU ETH_MAX_MTU
2114
2115 static int virtnet_probe(struct virtio_device *vdev)
2116 {
2117         int i, err;
2118         struct net_device *dev;
2119         struct virtnet_info *vi;
2120         u16 max_queue_pairs;
2121         int mtu;
2122
2123         if (!vdev->config->get) {
2124                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2125                         __func__);
2126                 return -EINVAL;
2127         }
2128
2129         if (!virtnet_validate_features(vdev))
2130                 return -EINVAL;
2131
2132         /* Find if host supports multiqueue virtio_net device */
2133         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2134                                    struct virtio_net_config,
2135                                    max_virtqueue_pairs, &max_queue_pairs);
2136
2137         /* We need at least 2 queue's */
2138         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2139             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2140             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2141                 max_queue_pairs = 1;
2142
2143         /* Allocate ourselves a network device with room for our info */
2144         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2145         if (!dev)
2146                 return -ENOMEM;
2147
2148         /* Set up network device as normal. */
2149         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2150         dev->netdev_ops = &virtnet_netdev;
2151         dev->features = NETIF_F_HIGHDMA;
2152
2153         dev->ethtool_ops = &virtnet_ethtool_ops;
2154         SET_NETDEV_DEV(dev, &vdev->dev);
2155
2156         /* Do we support "hardware" checksums? */
2157         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2158                 /* This opens up the world of extra features. */
2159                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2160                 if (csum)
2161                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2162
2163                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2164                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
2165                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2166                 }
2167                 /* Individual feature bits: what can host handle? */
2168                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2169                         dev->hw_features |= NETIF_F_TSO;
2170                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2171                         dev->hw_features |= NETIF_F_TSO6;
2172                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2173                         dev->hw_features |= NETIF_F_TSO_ECN;
2174                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
2175                         dev->hw_features |= NETIF_F_UFO;
2176
2177                 dev->features |= NETIF_F_GSO_ROBUST;
2178
2179                 if (gso)
2180                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
2181                 /* (!csum && gso) case will be fixed by register_netdev() */
2182         }
2183         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2184                 dev->features |= NETIF_F_RXCSUM;
2185
2186         dev->vlan_features = dev->features;
2187
2188         /* MTU range: 68 - 65535 */
2189         dev->min_mtu = MIN_MTU;
2190         dev->max_mtu = MAX_MTU;
2191
2192         /* Configuration may specify what MAC to use.  Otherwise random. */
2193         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2194                 virtio_cread_bytes(vdev,
2195                                    offsetof(struct virtio_net_config, mac),
2196                                    dev->dev_addr, dev->addr_len);
2197         else
2198                 eth_hw_addr_random(dev);
2199
2200         /* Set up our device-specific information */
2201         vi = netdev_priv(dev);
2202         vi->dev = dev;
2203         vi->vdev = vdev;
2204         vdev->priv = vi;
2205         vi->stats = alloc_percpu(struct virtnet_stats);
2206         err = -ENOMEM;
2207         if (vi->stats == NULL)
2208                 goto free;
2209
2210         for_each_possible_cpu(i) {
2211                 struct virtnet_stats *virtnet_stats;
2212                 virtnet_stats = per_cpu_ptr(vi->stats, i);
2213                 u64_stats_init(&virtnet_stats->tx_syncp);
2214                 u64_stats_init(&virtnet_stats->rx_syncp);
2215         }
2216
2217         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2218
2219         /* If we can receive ANY GSO packets, we must allocate large ones. */
2220         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2221             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2222             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2223             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2224                 vi->big_packets = true;
2225
2226         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2227                 vi->mergeable_rx_bufs = true;
2228
2229         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2230             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2231                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2232         else
2233                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2234
2235         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2236             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2237                 vi->any_header_sg = true;
2238
2239         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2240                 vi->has_cvq = true;
2241
2242         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2243                 mtu = virtio_cread16(vdev,
2244                                      offsetof(struct virtio_net_config,
2245                                               mtu));
2246                 if (mtu < dev->min_mtu) {
2247                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2248                 } else {
2249                         dev->mtu = mtu;
2250                         dev->max_mtu = mtu;
2251                 }
2252         }
2253
2254         if (vi->any_header_sg)
2255                 dev->needed_headroom = vi->hdr_len;
2256
2257         /* Enable multiqueue by default */
2258         if (num_online_cpus() >= max_queue_pairs)
2259                 vi->curr_queue_pairs = max_queue_pairs;
2260         else
2261                 vi->curr_queue_pairs = num_online_cpus();
2262         vi->max_queue_pairs = max_queue_pairs;
2263
2264         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2265         err = init_vqs(vi);
2266         if (err)
2267                 goto free_stats;
2268
2269 #ifdef CONFIG_SYSFS
2270         if (vi->mergeable_rx_bufs)
2271                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2272 #endif
2273         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2274         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2275
2276         virtnet_init_settings(dev);
2277
2278         err = register_netdev(dev);
2279         if (err) {
2280                 pr_debug("virtio_net: registering device failed\n");
2281                 goto free_vqs;
2282         }
2283
2284         virtio_device_ready(vdev);
2285
2286         err = virtnet_cpu_notif_add(vi);
2287         if (err) {
2288                 pr_debug("virtio_net: registering cpu notifier failed\n");
2289                 goto free_unregister_netdev;
2290         }
2291
2292         rtnl_lock();
2293         virtnet_set_queues(vi, vi->curr_queue_pairs);
2294         rtnl_unlock();
2295
2296         /* Assume link up if device can't report link status,
2297            otherwise get link status from config. */
2298         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2299                 netif_carrier_off(dev);
2300                 schedule_work(&vi->config_work);
2301         } else {
2302                 vi->status = VIRTIO_NET_S_LINK_UP;
2303                 netif_carrier_on(dev);
2304         }
2305
2306         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2307                  dev->name, max_queue_pairs);
2308
2309         return 0;
2310
2311 free_unregister_netdev:
2312         vi->vdev->config->reset(vdev);
2313
2314         unregister_netdev(dev);
2315 free_vqs:
2316         cancel_delayed_work_sync(&vi->refill);
2317         free_receive_page_frags(vi);
2318         virtnet_del_vqs(vi);
2319 free_stats:
2320         free_percpu(vi->stats);
2321 free:
2322         free_netdev(dev);
2323         return err;
2324 }
2325
2326 static void remove_vq_common(struct virtnet_info *vi)
2327 {
2328         vi->vdev->config->reset(vi->vdev);
2329
2330         /* Free unused buffers in both send and recv, if any. */
2331         free_unused_bufs(vi);
2332
2333         free_receive_bufs(vi);
2334
2335         free_receive_page_frags(vi);
2336
2337         virtnet_del_vqs(vi);
2338 }
2339
2340 static void virtnet_remove(struct virtio_device *vdev)
2341 {
2342         struct virtnet_info *vi = vdev->priv;
2343
2344         virtnet_cpu_notif_remove(vi);
2345
2346         /* Make sure no work handler is accessing the device. */
2347         flush_work(&vi->config_work);
2348
2349         unregister_netdev(vi->dev);
2350
2351         remove_vq_common(vi);
2352
2353         free_percpu(vi->stats);
2354         free_netdev(vi->dev);
2355 }
2356
2357 #ifdef CONFIG_PM_SLEEP
2358 static int virtnet_freeze(struct virtio_device *vdev)
2359 {
2360         struct virtnet_info *vi = vdev->priv;
2361         int i;
2362
2363         virtnet_cpu_notif_remove(vi);
2364
2365         /* Make sure no work handler is accessing the device */
2366         flush_work(&vi->config_work);
2367
2368         netif_device_detach(vi->dev);
2369         cancel_delayed_work_sync(&vi->refill);
2370
2371         if (netif_running(vi->dev)) {
2372                 for (i = 0; i < vi->max_queue_pairs; i++)
2373                         napi_disable(&vi->rq[i].napi);
2374         }
2375
2376         remove_vq_common(vi);
2377
2378         return 0;
2379 }
2380
2381 static int virtnet_restore(struct virtio_device *vdev)
2382 {
2383         struct virtnet_info *vi = vdev->priv;
2384         int err, i;
2385
2386         err = init_vqs(vi);
2387         if (err)
2388                 return err;
2389
2390         virtio_device_ready(vdev);
2391
2392         if (netif_running(vi->dev)) {
2393                 for (i = 0; i < vi->curr_queue_pairs; i++)
2394                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2395                                 schedule_delayed_work(&vi->refill, 0);
2396
2397                 for (i = 0; i < vi->max_queue_pairs; i++)
2398                         virtnet_napi_enable(&vi->rq[i]);
2399         }
2400
2401         netif_device_attach(vi->dev);
2402
2403         rtnl_lock();
2404         virtnet_set_queues(vi, vi->curr_queue_pairs);
2405         rtnl_unlock();
2406
2407         err = virtnet_cpu_notif_add(vi);
2408         if (err)
2409                 return err;
2410
2411         return 0;
2412 }
2413 #endif
2414
2415 static struct virtio_device_id id_table[] = {
2416         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2417         { 0 },
2418 };
2419
2420 #define VIRTNET_FEATURES \
2421         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2422         VIRTIO_NET_F_MAC, \
2423         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2424         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2425         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2426         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2427         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2428         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2429         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2430         VIRTIO_NET_F_MTU
2431
2432 static unsigned int features[] = {
2433         VIRTNET_FEATURES,
2434 };
2435
2436 static unsigned int features_legacy[] = {
2437         VIRTNET_FEATURES,
2438         VIRTIO_NET_F_GSO,
2439         VIRTIO_F_ANY_LAYOUT,
2440 };
2441
2442 static struct virtio_driver virtio_net_driver = {
2443         .feature_table = features,
2444         .feature_table_size = ARRAY_SIZE(features),
2445         .feature_table_legacy = features_legacy,
2446         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2447         .driver.name =  KBUILD_MODNAME,
2448         .driver.owner = THIS_MODULE,
2449         .id_table =     id_table,
2450         .probe =        virtnet_probe,
2451         .remove =       virtnet_remove,
2452         .config_changed = virtnet_config_changed,
2453 #ifdef CONFIG_PM_SLEEP
2454         .freeze =       virtnet_freeze,
2455         .restore =      virtnet_restore,
2456 #endif
2457 };
2458
2459 static __init int virtio_net_driver_init(void)
2460 {
2461         int ret;
2462
2463         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "AP_VIRT_NET_ONLINE",
2464                                       virtnet_cpu_online,
2465                                       virtnet_cpu_down_prep);
2466         if (ret < 0)
2467                 goto out;
2468         virtionet_online = ret;
2469         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "VIRT_NET_DEAD",
2470                                       NULL, virtnet_cpu_dead);
2471         if (ret)
2472                 goto err_dead;
2473
2474         ret = register_virtio_driver(&virtio_net_driver);
2475         if (ret)
2476                 goto err_virtio;
2477         return 0;
2478 err_virtio:
2479         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2480 err_dead:
2481         cpuhp_remove_multi_state(virtionet_online);
2482 out:
2483         return ret;
2484 }
2485 module_init(virtio_net_driver_init);
2486
2487 static __exit void virtio_net_driver_exit(void)
2488 {
2489         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2490         cpuhp_remove_multi_state(virtionet_online);
2491         unregister_virtio_driver(&virtio_net_driver);
2492 }
2493 module_exit(virtio_net_driver_exit);
2494
2495 MODULE_DEVICE_TABLE(virtio, id_table);
2496 MODULE_DESCRIPTION("Virtio network driver");
2497 MODULE_LICENSE("GPL");