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                              void *data)
 338 {
 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                 if (vi->mergeable_rx_bufs) {
 347                         struct page *sent_page = virt_to_head_page(xdp_sent);
 348
 349                         put_page(sent_page);
 350                 } else { /* small buffer */
 351                         struct sk_buff *skb = xdp_sent;
 352
 353                         kfree_skb(skb);
 354                 }
 355         }
 356
 357         if (vi->mergeable_rx_bufs) {
 358                 /* Zero header and leave csum up to XDP layers */
 359                 hdr = xdp->data;
 360                 memset(hdr, 0, vi->hdr_len);
 361
 362                 num_sg = 1;
 363                 sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
 364         } else { /* small buffer */
 365                 struct sk_buff *skb = data;
 366
 367                 /* Zero header and leave csum up to XDP layers */
 368                 hdr = skb_vnet_hdr(skb);
 369                 memset(hdr, 0, vi->hdr_len);
 370
 371                 num_sg = 2;
 372                 sg_init_table(sq->sg, 2);
 373                 sg_set_buf(sq->sg, hdr, vi->hdr_len);
 374                 skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
 375         }
 376         err = virtqueue_add_outbuf(sq->vq, sq->sg, num_sg,
 377                                    data, GFP_ATOMIC);
 378         if (unlikely(err)) {
 379                 if (vi->mergeable_rx_bufs) {
 380                         struct page *page = virt_to_head_page(xdp->data);
 381
 382                         put_page(page);
 383                 } else /* small buffer */
 384                         kfree_skb(data);
 385                 return; // On error abort to avoid unnecessary kick
 386         }
 387
 388         virtqueue_kick(sq->vq);
 389 }
 390
 391 static u32 do_xdp_prog(struct virtnet_info *vi,
 392                        struct receive_queue *rq,
 393                        struct bpf_prog *xdp_prog,
 394                        void *data, int len)
 395 {
 396         int hdr_padded_len;
 397         struct xdp_buff xdp;
 398         void *buf;
 399         unsigned int qp;
 400         u32 act;
 401
 402         if (vi->mergeable_rx_bufs) {
 403                 hdr_padded_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 404                 xdp.data = data + hdr_padded_len;
 405                 xdp.data_end = xdp.data + (len - vi->hdr_len);
 406                 buf = data;
 407         } else { /* small buffers */
 408                 struct sk_buff *skb = data;
 409
 410                 xdp.data = skb->data;
 411                 xdp.data_end = xdp.data + len;
 412                 buf = skb->data;
 413         }
 414
 415         act = bpf_prog_run_xdp(xdp_prog, &xdp);
 416         switch (act) {
 417         case XDP_PASS:
 418                 return XDP_PASS;
 419         case XDP_TX:
 420                 qp = vi->curr_queue_pairs -
 421                         vi->xdp_queue_pairs +
 422                         smp_processor_id();
 423                 xdp.data = buf;
 424                 virtnet_xdp_xmit(vi, rq, &vi->sq[qp], &xdp, data);
 425                 return XDP_TX;
 426         default:
 427                 bpf_warn_invalid_xdp_action(act);
 428         case XDP_ABORTED:
 429         case XDP_DROP:
 430                 return XDP_DROP;
 431         }
 432 }
 433
 434 static struct sk_buff *receive_small(struct net_device *dev,
 435                                      struct virtnet_info *vi,
 436                                      struct receive_queue *rq,
 437                                      void *buf, unsigned int len)
 438 {
 439         struct sk_buff * skb = buf;
 440         struct bpf_prog *xdp_prog;
 441
 442         len -= vi->hdr_len;
 443         skb_trim(skb, len);
 444
 445         rcu_read_lock();
 446         xdp_prog = rcu_dereference(rq->xdp_prog);
 447         if (xdp_prog) {
 448                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 449                 u32 act;
 450
 451                 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
 452                         goto err_xdp;
 453                 act = do_xdp_prog(vi, rq, xdp_prog, skb, len);
 454                 switch (act) {
 455                 case XDP_PASS:
 456                         break;
 457                 case XDP_TX:
 458                         rcu_read_unlock();
 459                         goto xdp_xmit;
 460                 case XDP_DROP:
 461                 default:
 462                         goto err_xdp;
 463                 }
 464         }
 465         rcu_read_unlock();
 466
 467         return skb;
 468
 469 err_xdp:
 470         rcu_read_unlock();
 471         dev->stats.rx_dropped++;
 472         kfree_skb(skb);
 473 xdp_xmit:
 474         return NULL;
 475 }
 476
 477 static struct sk_buff *receive_big(struct net_device *dev,
 478                                    struct virtnet_info *vi,
 479                                    struct receive_queue *rq,
 480                                    void *buf,
 481                                    unsigned int len)
 482 {
 483         struct page *page = buf;
 484         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 485
 486         if (unlikely(!skb))
 487                 goto err;
 488
 489         return skb;
 490
 491 err:
 492         dev->stats.rx_dropped++;
 493         give_pages(rq, page);
 494         return NULL;
 495 }
 496
 497 /* The conditions to enable XDP should preclude the underlying device from
 498  * sending packets across multiple buffers (num_buf > 1). However per spec
 499  * it does not appear to be illegal to do so but rather just against convention.
 500  * So in order to avoid making a system unresponsive the packets are pushed
 501  * into a page and the XDP program is run. This will be extremely slow and we
 502  * push a warning to the user to fix this as soon as possible. Fixing this may
 503  * require resolving the underlying hardware to determine why multiple buffers
 504  * are being received or simply loading the XDP program in the ingress stack
 505  * after the skb is built because there is no advantage to running it here
 506  * anymore.
 507  */
 508 static struct page *xdp_linearize_page(struct receive_queue *rq,
 509                                        u16 *num_buf,
 510                                        struct page *p,
 511                                        int offset,
 512                                        unsigned int *len)
 513 {
 514         struct page *page = alloc_page(GFP_ATOMIC);
 515         unsigned int page_off = 0;
 516
 517         if (!page)
 518                 return NULL;
 519
 520         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 521         page_off += *len;
 522
 523         while (--*num_buf) {
 524                 unsigned int buflen;
 525                 unsigned long ctx;
 526                 void *buf;
 527                 int off;
 528
 529                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &buflen);
 530                 if (unlikely(!ctx))
 531                         goto err_buf;
 532
 533                 buf = mergeable_ctx_to_buf_address(ctx);
 534                 p = virt_to_head_page(buf);
 535                 off = buf - page_address(p);
 536
 537                 /* guard against a misconfigured or uncooperative backend that
 538                  * is sending packet larger than the MTU.
 539                  */
 540                 if ((page_off + buflen) > PAGE_SIZE) {
 541                         put_page(p);
 542                         goto err_buf;
 543                 }
 544
 545                 memcpy(page_address(page) + page_off,
 546                        page_address(p) + off, buflen);
 547                 page_off += buflen;
 548                 put_page(p);
 549         }
 550
 551         *len = page_off;
 552         return page;
 553 err_buf:
 554         __free_pages(page, 0);
 555         return NULL;
 556 }
 557
 558 static struct sk_buff *receive_mergeable(struct net_device *dev,
 559                                          struct virtnet_info *vi,
 560                                          struct receive_queue *rq,
 561                                          unsigned long ctx,
 562                                          unsigned int len)
 563 {
 564         void *buf = mergeable_ctx_to_buf_address(ctx);
 565         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 566         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 567         struct page *page = virt_to_head_page(buf);
 568         int offset = buf - page_address(page);
 569         struct sk_buff *head_skb, *curr_skb;
 570         struct bpf_prog *xdp_prog;
 571         unsigned int truesize;
 572
 573         head_skb = NULL;
 574
 575         rcu_read_lock();
 576         xdp_prog = rcu_dereference(rq->xdp_prog);
 577         if (xdp_prog) {
 578                 struct page *xdp_page;
 579                 u32 act;
 580
 581                 /* This happens when rx buffer size is underestimated */
 582                 if (unlikely(num_buf > 1)) {
 583                         /* linearize data for XDP */
 584                         xdp_page = xdp_linearize_page(rq, &num_buf,
 585                                                       page, offset, &len);
 586                         if (!xdp_page)
 587                                 goto err_xdp;
 588                         offset = 0;
 589                 } else {
 590                         xdp_page = page;
 591                 }
 592
 593                 /* Transient failure which in theory could occur if
 594                  * in-flight packets from before XDP was enabled reach
 595                  * the receive path after XDP is loaded. In practice I
 596                  * was not able to create this condition.
 597                  */
 598                 if (unlikely(hdr->hdr.gso_type))
 599                         goto err_xdp;
 600
 601                 act = do_xdp_prog(vi, rq, xdp_prog,
 602                                   page_address(xdp_page) + offset, len);
 603                 switch (act) {
 604                 case XDP_PASS:
 605                         /* We can only create skb based on xdp_page. */
 606                         if (unlikely(xdp_page != page)) {
 607                                 rcu_read_unlock();
 608                                 put_page(page);
 609                                 head_skb = page_to_skb(vi, rq, xdp_page,
 610                                                        0, len, PAGE_SIZE);
 611                                 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 612                                 return head_skb;
 613                         }
 614                         break;
 615                 case XDP_TX:
 616                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 617                         if (unlikely(xdp_page != page))
 618                                 goto err_xdp;
 619                         rcu_read_unlock();
 620                         goto xdp_xmit;
 621                 case XDP_DROP:
 622                 default:
 623                         if (unlikely(xdp_page != page))
 624                                 __free_pages(xdp_page, 0);
 625                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 626                         goto err_xdp;
 627                 }
 628         }
 629         rcu_read_unlock();
 630
 631         truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 632         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
 633         curr_skb = head_skb;
 634
 635         if (unlikely(!curr_skb))
 636                 goto err_skb;
 637         while (--num_buf) {
 638                 int num_skb_frags;
 639
 640                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 641                 if (unlikely(!ctx)) {
 642                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 643                                  dev->name, num_buf,
 644                                  virtio16_to_cpu(vi->vdev,
 645                                                  hdr->num_buffers));
 646                         dev->stats.rx_length_errors++;
 647                         goto err_buf;
 648                 }
 649
 650                 buf = mergeable_ctx_to_buf_address(ctx);
 651                 page = virt_to_head_page(buf);
 652
 653                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 654                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 655                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 656
 657                         if (unlikely(!nskb))
 658                                 goto err_skb;
 659                         if (curr_skb == head_skb)
 660                                 skb_shinfo(curr_skb)->frag_list = nskb;
 661                         else
 662                                 curr_skb->next = nskb;
 663                         curr_skb = nskb;
 664                         head_skb->truesize += nskb->truesize;
 665                         num_skb_frags = 0;
 666                 }
 667                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 668                 if (curr_skb != head_skb) {
 669                         head_skb->data_len += len;
 670                         head_skb->len += len;
 671                         head_skb->truesize += truesize;
 672                 }
 673                 offset = buf - page_address(page);
 674                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 675                         put_page(page);
 676                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 677                                              len, truesize);
 678                 } else {
 679                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 680                                         offset, len, truesize);
 681                 }
 682         }
 683
 684         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 685         return head_skb;
 686
 687 err_xdp:
 688         rcu_read_unlock();
 689 err_skb:
 690         put_page(page);
 691         while (--num_buf) {
 692                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 693                 if (unlikely(!ctx)) {
 694                         pr_debug("%s: rx error: %d buffers missing\n",
 695                                  dev->name, num_buf);
 696                         dev->stats.rx_length_errors++;
 697                         break;
 698                 }
 699                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 700                 put_page(page);
 701         }
 702 err_buf:
 703         dev->stats.rx_dropped++;
 704         dev_kfree_skb(head_skb);
 705 xdp_xmit:
 706         return NULL;
 707 }
 708
 709 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 710                         void *buf, unsigned int len)
 711 {
 712         struct net_device *dev = vi->dev;
 713         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 714         struct sk_buff *skb;
 715         struct virtio_net_hdr_mrg_rxbuf *hdr;
 716
 717         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 718                 pr_debug("%s: short packet %i\n", dev->name, len);
 719                 dev->stats.rx_length_errors++;
 720                 if (vi->mergeable_rx_bufs) {
 721                         unsigned long ctx = (unsigned long)buf;
 722                         void *base = mergeable_ctx_to_buf_address(ctx);
 723                         put_page(virt_to_head_page(base));
 724                 } else if (vi->big_packets) {
 725                         give_pages(rq, buf);
 726                 } else {
 727                         dev_kfree_skb(buf);
 728                 }
 729                 return;
 730         }
 731
 732         if (vi->mergeable_rx_bufs)
 733                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
 734         else if (vi->big_packets)
 735                 skb = receive_big(dev, vi, rq, buf, len);
 736         else
 737                 skb = receive_small(dev, vi, rq, buf, len);
 738
 739         if (unlikely(!skb))
 740                 return;
 741
 742         hdr = skb_vnet_hdr(skb);
 743
 744         u64_stats_update_begin(&stats->rx_syncp);
 745         stats->rx_bytes += skb->len;
 746         stats->rx_packets++;
 747         u64_stats_update_end(&stats->rx_syncp);
 748
 749         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 750                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 751
 752         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 753                                   virtio_is_little_endian(vi->vdev))) {
 754                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 755                                      dev->name, hdr->hdr.gso_type,
 756                                      hdr->hdr.gso_size);
 757                 goto frame_err;
 758         }
 759
 760         skb->protocol = eth_type_trans(skb, dev);
 761         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 762                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 763
 764         napi_gro_receive(&rq->napi, skb);
 765         return;
 766
 767 frame_err:
 768         dev->stats.rx_frame_errors++;
 769         dev_kfree_skb(skb);
 770 }
 771
 772 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 773                              gfp_t gfp)
 774 {
 775         struct sk_buff *skb;
 776         struct virtio_net_hdr_mrg_rxbuf *hdr;
 777         int err;
 778
 779         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 780         if (unlikely(!skb))
 781                 return -ENOMEM;
 782
 783         skb_put(skb, GOOD_PACKET_LEN);
 784
 785         hdr = skb_vnet_hdr(skb);
 786         sg_init_table(rq->sg, 2);
 787         sg_set_buf(rq->sg, hdr, vi->hdr_len);
 788         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 789
 790         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 791         if (err < 0)
 792                 dev_kfree_skb(skb);
 793
 794         return err;
 795 }
 796
 797 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 798                            gfp_t gfp)
 799 {
 800         struct page *first, *list = NULL;
 801         char *p;
 802         int i, err, offset;
 803
 804         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 805
 806         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 807         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 808                 first = get_a_page(rq, gfp);
 809                 if (!first) {
 810                         if (list)
 811                                 give_pages(rq, list);
 812                         return -ENOMEM;
 813                 }
 814                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 815
 816                 /* chain new page in list head to match sg */
 817                 first->private = (unsigned long)list;
 818                 list = first;
 819         }
 820
 821         first = get_a_page(rq, gfp);
 822         if (!first) {
 823                 give_pages(rq, list);
 824                 return -ENOMEM;
 825         }
 826         p = page_address(first);
 827
 828         /* rq->sg[0], rq->sg[1] share the same page */
 829         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 830         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 831
 832         /* rq->sg[1] for data packet, from offset */
 833         offset = sizeof(struct padded_vnet_hdr);
 834         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 835
 836         /* chain first in list head */
 837         first->private = (unsigned long)list;
 838         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 839                                   first, gfp);
 840         if (err < 0)
 841                 give_pages(rq, first);
 842
 843         return err;
 844 }
 845
 846 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
 847 {
 848         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 849         unsigned int len;
 850
 851         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
 852                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 853         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 854 }
 855
 856 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 857 {
 858         struct page_frag *alloc_frag = &rq->alloc_frag;
 859         char *buf;
 860         unsigned long ctx;
 861         int err;
 862         unsigned int len, hole;
 863
 864         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 865         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 866                 return -ENOMEM;
 867
 868         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 869         ctx = mergeable_buf_to_ctx(buf, len);
 870         get_page(alloc_frag->page);
 871         alloc_frag->offset += len;
 872         hole = alloc_frag->size - alloc_frag->offset;
 873         if (hole < len) {
 874                 /* To avoid internal fragmentation, if there is very likely not
 875                  * enough space for another buffer, add the remaining space to
 876                  * the current buffer. This extra space is not included in
 877                  * the truesize stored in ctx.
 878                  */
 879                 len += hole;
 880                 alloc_frag->offset += hole;
 881         }
 882
 883         sg_init_one(rq->sg, buf, len);
 884         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 885         if (err < 0)
 886                 put_page(virt_to_head_page(buf));
 887
 888         return err;
 889 }
 890
 891 /*
 892  * Returns false if we couldn't fill entirely (OOM).
 893  *
 894  * Normally run in the receive path, but can also be run from ndo_open
 895  * before we're receiving packets, or from refill_work which is
 896  * careful to disable receiving (using napi_disable).
 897  */
 898 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 899                           gfp_t gfp)
 900 {
 901         int err;
 902         bool oom;
 903
 904         gfp |= __GFP_COLD;
 905         do {
 906                 if (vi->mergeable_rx_bufs)
 907                         err = add_recvbuf_mergeable(rq, gfp);
 908                 else if (vi->big_packets)
 909                         err = add_recvbuf_big(vi, rq, gfp);
 910                 else
 911                         err = add_recvbuf_small(vi, rq, gfp);
 912
 913                 oom = err == -ENOMEM;
 914                 if (err)
 915                         break;
 916         } while (rq->vq->num_free);
 917         virtqueue_kick(rq->vq);
 918         return !oom;
 919 }
 920
 921 static void skb_recv_done(struct virtqueue *rvq)
 922 {
 923         struct virtnet_info *vi = rvq->vdev->priv;
 924         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 925
 926         /* Schedule NAPI, Suppress further interrupts if successful. */
 927         if (napi_schedule_prep(&rq->napi)) {
 928                 virtqueue_disable_cb(rvq);
 929                 __napi_schedule(&rq->napi);
 930         }
 931 }
 932
 933 static void virtnet_napi_enable(struct receive_queue *rq)
 934 {
 935         napi_enable(&rq->napi);
 936
 937         /* If all buffers were filled by other side before we napi_enabled, we
 938          * won't get another interrupt, so process any outstanding packets
 939          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 940          * We synchronize against interrupts via NAPI_STATE_SCHED */
 941         if (napi_schedule_prep(&rq->napi)) {
 942                 virtqueue_disable_cb(rq->vq);
 943                 local_bh_disable();
 944                 __napi_schedule(&rq->napi);
 945                 local_bh_enable();
 946         }
 947 }
 948
 949 static void refill_work(struct work_struct *work)
 950 {
 951         struct virtnet_info *vi =
 952                 container_of(work, struct virtnet_info, refill.work);
 953         bool still_empty;
 954         int i;
 955
 956         for (i = 0; i < vi->curr_queue_pairs; i++) {
 957                 struct receive_queue *rq = &vi->rq[i];
 958
 959                 napi_disable(&rq->napi);
 960                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
 961                 virtnet_napi_enable(rq);
 962
 963                 /* In theory, this can happen: if we don't get any buffers in
 964                  * we will *never* try to fill again.
 965                  */
 966                 if (still_empty)
 967                         schedule_delayed_work(&vi->refill, HZ/2);
 968         }
 969 }
 970
 971 static int virtnet_receive(struct receive_queue *rq, int budget)
 972 {
 973         struct virtnet_info *vi = rq->vq->vdev->priv;
 974         unsigned int len, received = 0;
 975         void *buf;
 976
 977         while (received < budget &&
 978                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 979                 receive_buf(vi, rq, buf, len);
 980                 received++;
 981         }
 982
 983         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 984                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 985                         schedule_delayed_work(&vi->refill, 0);
 986         }
 987
 988         return received;
 989 }
 990
 991 static int virtnet_poll(struct napi_struct *napi, int budget)
 992 {
 993         struct receive_queue *rq =
 994                 container_of(napi, struct receive_queue, napi);
 995         unsigned int r, received;
 996
 997         received = virtnet_receive(rq, budget);
 998
 999         /* Out of packets? */
1000         if (received < budget) {
1001                 r = virtqueue_enable_cb_prepare(rq->vq);
1002                 napi_complete_done(napi, received);
1003                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
1004                     napi_schedule_prep(napi)) {
1005                         virtqueue_disable_cb(rq->vq);
1006                         __napi_schedule(napi);
1007                 }
1008         }
1009
1010         return received;
1011 }
1012
1013 #ifdef CONFIG_NET_RX_BUSY_POLL
1014 /* must be called with local_bh_disable()d */
1015 static int virtnet_busy_poll(struct napi_struct *napi)
1016 {
1017         struct receive_queue *rq =
1018                 container_of(napi, struct receive_queue, napi);
1019         struct virtnet_info *vi = rq->vq->vdev->priv;
1020         int r, received = 0, budget = 4;
1021
1022         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
1023                 return LL_FLUSH_FAILED;
1024
1025         if (!napi_schedule_prep(napi))
1026                 return LL_FLUSH_BUSY;
1027
1028         virtqueue_disable_cb(rq->vq);
1029
1030 again:
1031         received += virtnet_receive(rq, budget);
1032
1033         r = virtqueue_enable_cb_prepare(rq->vq);
1034         clear_bit(NAPI_STATE_SCHED, &napi->state);
1035         if (unlikely(virtqueue_poll(rq->vq, r)) &&
1036             napi_schedule_prep(napi)) {
1037                 virtqueue_disable_cb(rq->vq);
1038                 if (received < budget) {
1039                         budget -= received;
1040                         goto again;
1041                 } else {
1042                         __napi_schedule(napi);
1043                 }
1044         }
1045
1046         return received;
1047 }
1048 #endif  /* CONFIG_NET_RX_BUSY_POLL */
1049
1050 static int virtnet_open(struct net_device *dev)
1051 {
1052         struct virtnet_info *vi = netdev_priv(dev);
1053         int i;
1054
1055         for (i = 0; i < vi->max_queue_pairs; i++) {
1056                 if (i < vi->curr_queue_pairs)
1057                         /* Make sure we have some buffers: if oom use wq. */
1058                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1059                                 schedule_delayed_work(&vi->refill, 0);
1060                 virtnet_napi_enable(&vi->rq[i]);
1061         }
1062
1063         return 0;
1064 }
1065
1066 static void free_old_xmit_skbs(struct send_queue *sq)
1067 {
1068         struct sk_buff *skb;
1069         unsigned int len;
1070         struct virtnet_info *vi = sq->vq->vdev->priv;
1071         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1072
1073         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1074                 pr_debug("Sent skb %p\n", skb);
1075
1076                 u64_stats_update_begin(&stats->tx_syncp);
1077                 stats->tx_bytes += skb->len;
1078                 stats->tx_packets++;
1079                 u64_stats_update_end(&stats->tx_syncp);
1080
1081                 dev_kfree_skb_any(skb);
1082         }
1083 }
1084
1085 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1086 {
1087         struct virtio_net_hdr_mrg_rxbuf *hdr;
1088         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1089         struct virtnet_info *vi = sq->vq->vdev->priv;
1090         unsigned num_sg;
1091         unsigned hdr_len = vi->hdr_len;
1092         bool can_push;
1093
1094         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1095
1096         can_push = vi->any_header_sg &&
1097                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1098                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1099         /* Even if we can, don't push here yet as this would skew
1100          * csum_start offset below. */
1101         if (can_push)
1102                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1103         else
1104                 hdr = skb_vnet_hdr(skb);
1105
1106         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1107                                     virtio_is_little_endian(vi->vdev)))
1108                 BUG();
1109
1110         if (vi->mergeable_rx_bufs)
1111                 hdr->num_buffers = 0;
1112
1113         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1114         if (can_push) {
1115                 __skb_push(skb, hdr_len);
1116                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1117                 /* Pull header back to avoid skew in tx bytes calculations. */
1118                 __skb_pull(skb, hdr_len);
1119         } else {
1120                 sg_set_buf(sq->sg, hdr, hdr_len);
1121                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
1122         }
1123         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1124 }
1125
1126 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1127 {
1128         struct virtnet_info *vi = netdev_priv(dev);
1129         int qnum = skb_get_queue_mapping(skb);
1130         struct send_queue *sq = &vi->sq[qnum];
1131         int err;
1132         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1133         bool kick = !skb->xmit_more;
1134
1135         /* Free up any pending old buffers before queueing new ones. */
1136         free_old_xmit_skbs(sq);
1137
1138         /* timestamp packet in software */
1139         skb_tx_timestamp(skb);
1140
1141         /* Try to transmit */
1142         err = xmit_skb(sq, skb);
1143
1144         /* This should not happen! */
1145         if (unlikely(err)) {
1146                 dev->stats.tx_fifo_errors++;
1147                 if (net_ratelimit())
1148                         dev_warn(&dev->dev,
1149                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1150                 dev->stats.tx_dropped++;
1151                 dev_kfree_skb_any(skb);
1152                 return NETDEV_TX_OK;
1153         }
1154
1155         /* Don't wait up for transmitted skbs to be freed. */
1156         skb_orphan(skb);
1157         nf_reset(skb);
1158
1159         /* If running out of space, stop queue to avoid getting packets that we
1160          * are then unable to transmit.
1161          * An alternative would be to force queuing layer to requeue the skb by
1162          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1163          * returned in a normal path of operation: it means that driver is not
1164          * maintaining the TX queue stop/start state properly, and causes
1165          * the stack to do a non-trivial amount of useless work.
1166          * Since most packets only take 1 or 2 ring slots, stopping the queue
1167          * early means 16 slots are typically wasted.
1168          */
1169         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1170                 netif_stop_subqueue(dev, qnum);
1171                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1172                         /* More just got used, free them then recheck. */
1173                         free_old_xmit_skbs(sq);
1174                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1175                                 netif_start_subqueue(dev, qnum);
1176                                 virtqueue_disable_cb(sq->vq);
1177                         }
1178                 }
1179         }
1180
1181         if (kick || netif_xmit_stopped(txq))
1182                 virtqueue_kick(sq->vq);
1183
1184         return NETDEV_TX_OK;
1185 }
1186
1187 /*
1188  * Send command via the control virtqueue and check status.  Commands
1189  * supported by the hypervisor, as indicated by feature bits, should
1190  * never fail unless improperly formatted.
1191  */
1192 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1193                                  struct scatterlist *out)
1194 {
1195         struct scatterlist *sgs[4], hdr, stat;
1196         unsigned out_num = 0, tmp;
1197
1198         /* Caller should know better */
1199         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1200
1201         vi->ctrl_status = ~0;
1202         vi->ctrl_hdr.class = class;
1203         vi->ctrl_hdr.cmd = cmd;
1204         /* Add header */
1205         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1206         sgs[out_num++] = &hdr;
1207
1208         if (out)
1209                 sgs[out_num++] = out;
1210
1211         /* Add return status. */
1212         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1213         sgs[out_num] = &stat;
1214
1215         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1216         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1217
1218         if (unlikely(!virtqueue_kick(vi->cvq)))
1219                 return vi->ctrl_status == VIRTIO_NET_OK;
1220
1221         /* Spin for a response, the kick causes an ioport write, trapping
1222          * into the hypervisor, so the request should be handled immediately.
1223          */
1224         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1225                !virtqueue_is_broken(vi->cvq))
1226                 cpu_relax();
1227
1228         return vi->ctrl_status == VIRTIO_NET_OK;
1229 }
1230
1231 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1232 {
1233         struct virtnet_info *vi = netdev_priv(dev);
1234         struct virtio_device *vdev = vi->vdev;
1235         int ret;
1236         struct sockaddr *addr;
1237         struct scatterlist sg;
1238
1239         addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1240         if (!addr)
1241                 return -ENOMEM;
1242         memcpy(addr, p, sizeof(*addr));
1243
1244         ret = eth_prepare_mac_addr_change(dev, addr);
1245         if (ret)
1246                 goto out;
1247
1248         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1249                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1250                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1251                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1252                         dev_warn(&vdev->dev,
1253                                  "Failed to set mac address by vq command.\n");
1254                         ret = -EINVAL;
1255                         goto out;
1256                 }
1257         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1258                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1259                 unsigned int i;
1260
1261                 /* Naturally, this has an atomicity problem. */
1262                 for (i = 0; i < dev->addr_len; i++)
1263                         virtio_cwrite8(vdev,
1264                                        offsetof(struct virtio_net_config, mac) +
1265                                        i, addr->sa_data[i]);
1266         }
1267
1268         eth_commit_mac_addr_change(dev, p);
1269         ret = 0;
1270
1271 out:
1272         kfree(addr);
1273         return ret;
1274 }
1275
1276 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1277                                                struct rtnl_link_stats64 *tot)
1278 {
1279         struct virtnet_info *vi = netdev_priv(dev);
1280         int cpu;
1281         unsigned int start;
1282
1283         for_each_possible_cpu(cpu) {
1284                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1285                 u64 tpackets, tbytes, rpackets, rbytes;
1286
1287                 do {
1288                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1289                         tpackets = stats->tx_packets;
1290                         tbytes   = stats->tx_bytes;
1291                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1292
1293                 do {
1294                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1295                         rpackets = stats->rx_packets;
1296                         rbytes   = stats->rx_bytes;
1297                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1298
1299                 tot->rx_packets += rpackets;
1300                 tot->tx_packets += tpackets;
1301                 tot->rx_bytes   += rbytes;
1302                 tot->tx_bytes   += tbytes;
1303         }
1304
1305         tot->tx_dropped = dev->stats.tx_dropped;
1306         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1307         tot->rx_dropped = dev->stats.rx_dropped;
1308         tot->rx_length_errors = dev->stats.rx_length_errors;
1309         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1310
1311         return tot;
1312 }
1313
1314 #ifdef CONFIG_NET_POLL_CONTROLLER
1315 static void virtnet_netpoll(struct net_device *dev)
1316 {
1317         struct virtnet_info *vi = netdev_priv(dev);
1318         int i;
1319
1320         for (i = 0; i < vi->curr_queue_pairs; i++)
1321                 napi_schedule(&vi->rq[i].napi);
1322 }
1323 #endif
1324
1325 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1326 {
1327         rtnl_lock();
1328         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1329                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1330                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1331         rtnl_unlock();
1332 }
1333
1334 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1335 {
1336         struct scatterlist sg;
1337         struct net_device *dev = vi->dev;
1338
1339         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1340                 return 0;
1341
1342         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1343         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1344
1345         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1346                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1347                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1348                          queue_pairs);
1349                 return -EINVAL;
1350         } else {
1351                 vi->curr_queue_pairs = queue_pairs;
1352                 /* virtnet_open() will refill when device is going to up. */
1353                 if (dev->flags & IFF_UP)
1354                         schedule_delayed_work(&vi->refill, 0);
1355         }
1356
1357         return 0;
1358 }
1359
1360 static int virtnet_close(struct net_device *dev)
1361 {
1362         struct virtnet_info *vi = netdev_priv(dev);
1363         int i;
1364
1365         /* Make sure refill_work doesn't re-enable napi! */
1366         cancel_delayed_work_sync(&vi->refill);
1367
1368         for (i = 0; i < vi->max_queue_pairs; i++)
1369                 napi_disable(&vi->rq[i].napi);
1370
1371         return 0;
1372 }
1373
1374 static void virtnet_set_rx_mode(struct net_device *dev)
1375 {
1376         struct virtnet_info *vi = netdev_priv(dev);
1377         struct scatterlist sg[2];
1378         struct virtio_net_ctrl_mac *mac_data;
1379         struct netdev_hw_addr *ha;
1380         int uc_count;
1381         int mc_count;
1382         void *buf;
1383         int i;
1384
1385         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1386         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1387                 return;
1388
1389         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1390         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1391
1392         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1393
1394         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1395                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1396                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1397                          vi->ctrl_promisc ? "en" : "dis");
1398
1399         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1400
1401         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1402                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1403                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1404                          vi->ctrl_allmulti ? "en" : "dis");
1405
1406         uc_count = netdev_uc_count(dev);
1407         mc_count = netdev_mc_count(dev);
1408         /* MAC filter - use one buffer for both lists */
1409         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1410                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1411         mac_data = buf;
1412         if (!buf)
1413                 return;
1414
1415         sg_init_table(sg, 2);
1416
1417         /* Store the unicast list and count in the front of the buffer */
1418         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1419         i = 0;
1420         netdev_for_each_uc_addr(ha, dev)
1421                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1422
1423         sg_set_buf(&sg[0], mac_data,
1424                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1425
1426         /* multicast list and count fill the end */
1427         mac_data = (void *)&mac_data->macs[uc_count][0];
1428
1429         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1430         i = 0;
1431         netdev_for_each_mc_addr(ha, dev)
1432                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1433
1434         sg_set_buf(&sg[1], mac_data,
1435                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1436
1437         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1438                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1439                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1440
1441         kfree(buf);
1442 }
1443
1444 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1445                                    __be16 proto, u16 vid)
1446 {
1447         struct virtnet_info *vi = netdev_priv(dev);
1448         struct scatterlist sg;
1449
1450         vi->ctrl_vid = vid;
1451         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1452
1453         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1454                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1455                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1456         return 0;
1457 }
1458
1459 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1460                                     __be16 proto, u16 vid)
1461 {
1462         struct virtnet_info *vi = netdev_priv(dev);
1463         struct scatterlist sg;
1464
1465         vi->ctrl_vid = vid;
1466         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1467
1468         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1469                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1470                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1471         return 0;
1472 }
1473
1474 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1475 {
1476         int i;
1477
1478         if (vi->affinity_hint_set) {
1479                 for (i = 0; i < vi->max_queue_pairs; i++) {
1480                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1481                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1482                 }
1483
1484                 vi->affinity_hint_set = false;
1485         }
1486 }
1487
1488 static void virtnet_set_affinity(struct virtnet_info *vi)
1489 {
1490         int i;
1491         int cpu;
1492
1493         /* In multiqueue mode, when the number of cpu is equal to the number of
1494          * queue pairs, we let the queue pairs to be private to one cpu by
1495          * setting the affinity hint to eliminate the contention.
1496          */
1497         if (vi->curr_queue_pairs == 1 ||
1498             vi->max_queue_pairs != num_online_cpus()) {
1499                 virtnet_clean_affinity(vi, -1);
1500                 return;
1501         }
1502
1503         i = 0;
1504         for_each_online_cpu(cpu) {
1505                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1506                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1507                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1508                 i++;
1509         }
1510
1511         vi->affinity_hint_set = true;
1512 }
1513
1514 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1515 {
1516         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1517                                                    node);
1518         virtnet_set_affinity(vi);
1519         return 0;
1520 }
1521
1522 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1523 {
1524         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1525                                                    node_dead);
1526         virtnet_set_affinity(vi);
1527         return 0;
1528 }
1529
1530 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1531 {
1532         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1533                                                    node);
1534
1535         virtnet_clean_affinity(vi, cpu);
1536         return 0;
1537 }
1538
1539 static enum cpuhp_state virtionet_online;
1540
1541 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1542 {
1543         int ret;
1544
1545         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1546         if (ret)
1547                 return ret;
1548         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1549                                                &vi->node_dead);
1550         if (!ret)
1551                 return ret;
1552         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1553         return ret;
1554 }
1555
1556 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1557 {
1558         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1559         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1560                                             &vi->node_dead);
1561 }
1562
1563 static void virtnet_get_ringparam(struct net_device *dev,
1564                                 struct ethtool_ringparam *ring)
1565 {
1566         struct virtnet_info *vi = netdev_priv(dev);
1567
1568         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1569         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1570         ring->rx_pending = ring->rx_max_pending;
1571         ring->tx_pending = ring->tx_max_pending;
1572 }
1573
1574
1575 static void virtnet_get_drvinfo(struct net_device *dev,
1576                                 struct ethtool_drvinfo *info)
1577 {
1578         struct virtnet_info *vi = netdev_priv(dev);
1579         struct virtio_device *vdev = vi->vdev;
1580
1581         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1582         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1583         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1584
1585 }
1586
1587 /* TODO: Eliminate OOO packets during switching */
1588 static int virtnet_set_channels(struct net_device *dev,
1589                                 struct ethtool_channels *channels)
1590 {
1591         struct virtnet_info *vi = netdev_priv(dev);
1592         u16 queue_pairs = channels->combined_count;
1593         int err;
1594
1595         /* We don't support separate rx/tx channels.
1596          * We don't allow setting 'other' channels.
1597          */
1598         if (channels->rx_count || channels->tx_count || channels->other_count)
1599                 return -EINVAL;
1600
1601         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1602                 return -EINVAL;
1603
1604         /* For now we don't support modifying channels while XDP is loaded
1605          * also when XDP is loaded all RX queues have XDP programs so we only
1606          * need to check a single RX queue.
1607          */
1608         if (vi->rq[0].xdp_prog)
1609                 return -EINVAL;
1610
1611         get_online_cpus();
1612         err = virtnet_set_queues(vi, queue_pairs);
1613         if (!err) {
1614                 netif_set_real_num_tx_queues(dev, queue_pairs);
1615                 netif_set_real_num_rx_queues(dev, queue_pairs);
1616
1617                 virtnet_set_affinity(vi);
1618         }
1619         put_online_cpus();
1620
1621         return err;
1622 }
1623
1624 static void virtnet_get_channels(struct net_device *dev,
1625                                  struct ethtool_channels *channels)
1626 {
1627         struct virtnet_info *vi = netdev_priv(dev);
1628
1629         channels->combined_count = vi->curr_queue_pairs;
1630         channels->max_combined = vi->max_queue_pairs;
1631         channels->max_other = 0;
1632         channels->rx_count = 0;
1633         channels->tx_count = 0;
1634         channels->other_count = 0;
1635 }
1636
1637 /* Check if the user is trying to change anything besides speed/duplex */
1638 static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1639 {
1640         struct ethtool_cmd diff1 = *cmd;
1641         struct ethtool_cmd diff2 = {};
1642
1643         /* cmd is always set so we need to clear it, validate the port type
1644          * and also without autonegotiation we can ignore advertising
1645          */
1646         ethtool_cmd_speed_set(&diff1, 0);
1647         diff2.port = PORT_OTHER;
1648         diff1.advertising = 0;
1649         diff1.duplex = 0;
1650         diff1.cmd = 0;
1651
1652         return !memcmp(&diff1, &diff2, sizeof(diff1));
1653 }
1654
1655 static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1656 {
1657         struct virtnet_info *vi = netdev_priv(dev);
1658         u32 speed;
1659
1660         speed = ethtool_cmd_speed(cmd);
1661         /* don't allow custom speed and duplex */
1662         if (!ethtool_validate_speed(speed) ||
1663             !ethtool_validate_duplex(cmd->duplex) ||
1664             !virtnet_validate_ethtool_cmd(cmd))
1665                 return -EINVAL;
1666         vi->speed = speed;
1667         vi->duplex = cmd->duplex;
1668
1669         return 0;
1670 }
1671
1672 static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1673 {
1674         struct virtnet_info *vi = netdev_priv(dev);
1675
1676         ethtool_cmd_speed_set(cmd, vi->speed);
1677         cmd->duplex = vi->duplex;
1678         cmd->port = PORT_OTHER;
1679
1680         return 0;
1681 }
1682
1683 static void virtnet_init_settings(struct net_device *dev)
1684 {
1685         struct virtnet_info *vi = netdev_priv(dev);
1686
1687         vi->speed = SPEED_UNKNOWN;
1688         vi->duplex = DUPLEX_UNKNOWN;
1689 }
1690
1691 static const struct ethtool_ops virtnet_ethtool_ops = {
1692         .get_drvinfo = virtnet_get_drvinfo,
1693         .get_link = ethtool_op_get_link,
1694         .get_ringparam = virtnet_get_ringparam,
1695         .set_channels = virtnet_set_channels,
1696         .get_channels = virtnet_get_channels,
1697         .get_ts_info = ethtool_op_get_ts_info,
1698         .get_settings = virtnet_get_settings,
1699         .set_settings = virtnet_set_settings,
1700 };
1701
1702 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog)
1703 {
1704         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1705         struct virtnet_info *vi = netdev_priv(dev);
1706         struct bpf_prog *old_prog;
1707         u16 xdp_qp = 0, curr_qp;
1708         int i, err;
1709
1710         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1711             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1712             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
1713             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO)) {
1714                 netdev_warn(dev, "can't set XDP while host is implementing LRO, disable LRO first\n");
1715                 return -EOPNOTSUPP;
1716         }
1717
1718         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1719                 netdev_warn(dev, "XDP expects header/data in single page, any_header_sg required\n");
1720                 return -EINVAL;
1721         }
1722
1723         if (dev->mtu > max_sz) {
1724                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1725                 return -EINVAL;
1726         }
1727
1728         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1729         if (prog)
1730                 xdp_qp = nr_cpu_ids;
1731
1732         /* XDP requires extra queues for XDP_TX */
1733         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1734                 netdev_warn(dev, "request %i queues but max is %i\n",
1735                             curr_qp + xdp_qp, vi->max_queue_pairs);
1736                 return -ENOMEM;
1737         }
1738
1739         err = virtnet_set_queues(vi, curr_qp + xdp_qp);
1740         if (err) {
1741                 dev_warn(&dev->dev, "XDP Device queue allocation failure.\n");
1742                 return err;
1743         }
1744
1745         if (prog) {
1746                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1747                 if (IS_ERR(prog)) {
1748                         virtnet_set_queues(vi, curr_qp);
1749                         return PTR_ERR(prog);
1750                 }
1751         }
1752
1753         vi->xdp_queue_pairs = xdp_qp;
1754         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
1755
1756         for (i = 0; i < vi->max_queue_pairs; i++) {
1757                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1758                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
1759                 if (old_prog)
1760                         bpf_prog_put(old_prog);
1761         }
1762
1763         return 0;
1764 }
1765
1766 static bool virtnet_xdp_query(struct net_device *dev)
1767 {
1768         struct virtnet_info *vi = netdev_priv(dev);
1769         int i;
1770
1771         for (i = 0; i < vi->max_queue_pairs; i++) {
1772                 if (vi->rq[i].xdp_prog)
1773                         return true;
1774         }
1775         return false;
1776 }
1777
1778 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
1779 {
1780         switch (xdp->command) {
1781         case XDP_SETUP_PROG:
1782                 return virtnet_xdp_set(dev, xdp->prog);
1783         case XDP_QUERY_PROG:
1784                 xdp->prog_attached = virtnet_xdp_query(dev);
1785                 return 0;
1786         default:
1787                 return -EINVAL;
1788         }
1789 }
1790
1791 static const struct net_device_ops virtnet_netdev = {
1792         .ndo_open            = virtnet_open,
1793         .ndo_stop            = virtnet_close,
1794         .ndo_start_xmit      = start_xmit,
1795         .ndo_validate_addr   = eth_validate_addr,
1796         .ndo_set_mac_address = virtnet_set_mac_address,
1797         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1798         .ndo_get_stats64     = virtnet_stats,
1799         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1800         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1801 #ifdef CONFIG_NET_POLL_CONTROLLER
1802         .ndo_poll_controller = virtnet_netpoll,
1803 #endif
1804 #ifdef CONFIG_NET_RX_BUSY_POLL
1805         .ndo_busy_poll          = virtnet_busy_poll,
1806 #endif
1807         .ndo_xdp                = virtnet_xdp,
1808 };
1809
1810 static void virtnet_config_changed_work(struct work_struct *work)
1811 {
1812         struct virtnet_info *vi =
1813                 container_of(work, struct virtnet_info, config_work);
1814         u16 v;
1815
1816         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1817                                  struct virtio_net_config, status, &v) < 0)
1818                 return;
1819
1820         if (v & VIRTIO_NET_S_ANNOUNCE) {
1821                 netdev_notify_peers(vi->dev);
1822                 virtnet_ack_link_announce(vi);
1823         }
1824
1825         /* Ignore unknown (future) status bits */
1826         v &= VIRTIO_NET_S_LINK_UP;
1827
1828         if (vi->status == v)
1829                 return;
1830
1831         vi->status = v;
1832
1833         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1834                 netif_carrier_on(vi->dev);
1835                 netif_tx_wake_all_queues(vi->dev);
1836         } else {
1837                 netif_carrier_off(vi->dev);
1838                 netif_tx_stop_all_queues(vi->dev);
1839         }
1840 }
1841
1842 static void virtnet_config_changed(struct virtio_device *vdev)
1843 {
1844         struct virtnet_info *vi = vdev->priv;
1845
1846         schedule_work(&vi->config_work);
1847 }
1848
1849 static void virtnet_free_queues(struct virtnet_info *vi)
1850 {
1851         int i;
1852
1853         for (i = 0; i < vi->max_queue_pairs; i++) {
1854                 napi_hash_del(&vi->rq[i].napi);
1855                 netif_napi_del(&vi->rq[i].napi);
1856         }
1857
1858         /* We called napi_hash_del() before netif_napi_del(),
1859          * we need to respect an RCU grace period before freeing vi->rq
1860          */
1861         synchronize_net();
1862
1863         kfree(vi->rq);
1864         kfree(vi->sq);
1865 }
1866
1867 static void free_receive_bufs(struct virtnet_info *vi)
1868 {
1869         struct bpf_prog *old_prog;
1870         int i;
1871
1872         rtnl_lock();
1873         for (i = 0; i < vi->max_queue_pairs; i++) {
1874                 while (vi->rq[i].pages)
1875                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1876
1877                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1878                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
1879                 if (old_prog)
1880                         bpf_prog_put(old_prog);
1881         }
1882         rtnl_unlock();
1883 }
1884
1885 static void free_receive_page_frags(struct virtnet_info *vi)
1886 {
1887         int i;
1888         for (i = 0; i < vi->max_queue_pairs; i++)
1889                 if (vi->rq[i].alloc_frag.page)
1890                         put_page(vi->rq[i].alloc_frag.page);
1891 }
1892
1893 static bool is_xdp_queue(struct virtnet_info *vi, int q)
1894 {
1895         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1896                 return false;
1897         else if (q < vi->curr_queue_pairs)
1898                 return true;
1899         else
1900                 return false;
1901 }
1902
1903 static void free_unused_bufs(struct virtnet_info *vi)
1904 {
1905         void *buf;
1906         int i;
1907
1908         for (i = 0; i < vi->max_queue_pairs; i++) {
1909                 struct virtqueue *vq = vi->sq[i].vq;
1910                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1911                         if (!is_xdp_queue(vi, i))
1912                                 dev_kfree_skb(buf);
1913                         else
1914                                 put_page(virt_to_head_page(buf));
1915                 }
1916         }
1917
1918         for (i = 0; i < vi->max_queue_pairs; i++) {
1919                 struct virtqueue *vq = vi->rq[i].vq;
1920
1921                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1922                         if (vi->mergeable_rx_bufs) {
1923                                 unsigned long ctx = (unsigned long)buf;
1924                                 void *base = mergeable_ctx_to_buf_address(ctx);
1925                                 put_page(virt_to_head_page(base));
1926                         } else if (vi->big_packets) {
1927                                 give_pages(&vi->rq[i], buf);
1928                         } else {
1929                                 dev_kfree_skb(buf);
1930                         }
1931                 }
1932         }
1933 }
1934
1935 static void virtnet_del_vqs(struct virtnet_info *vi)
1936 {
1937         struct virtio_device *vdev = vi->vdev;
1938
1939         virtnet_clean_affinity(vi, -1);
1940
1941         vdev->config->del_vqs(vdev);
1942
1943         virtnet_free_queues(vi);
1944 }
1945
1946 static int virtnet_find_vqs(struct virtnet_info *vi)
1947 {
1948         vq_callback_t **callbacks;
1949         struct virtqueue **vqs;
1950         int ret = -ENOMEM;
1951         int i, total_vqs;
1952         const char **names;
1953
1954         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1955          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1956          * possible control vq.
1957          */
1958         total_vqs = vi->max_queue_pairs * 2 +
1959                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1960
1961         /* Allocate space for find_vqs parameters */
1962         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1963         if (!vqs)
1964                 goto err_vq;
1965         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1966         if (!callbacks)
1967                 goto err_callback;
1968         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1969         if (!names)
1970                 goto err_names;
1971
1972         /* Parameters for control virtqueue, if any */
1973         if (vi->has_cvq) {
1974                 callbacks[total_vqs - 1] = NULL;
1975                 names[total_vqs - 1] = "control";
1976         }
1977
1978         /* Allocate/initialize parameters for send/receive virtqueues */
1979         for (i = 0; i < vi->max_queue_pairs; i++) {
1980                 callbacks[rxq2vq(i)] = skb_recv_done;
1981                 callbacks[txq2vq(i)] = skb_xmit_done;
1982                 sprintf(vi->rq[i].name, "input.%d", i);
1983                 sprintf(vi->sq[i].name, "output.%d", i);
1984                 names[rxq2vq(i)] = vi->rq[i].name;
1985                 names[txq2vq(i)] = vi->sq[i].name;
1986         }
1987
1988         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1989                                          names);
1990         if (ret)
1991                 goto err_find;
1992
1993         if (vi->has_cvq) {
1994                 vi->cvq = vqs[total_vqs - 1];
1995                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1996                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1997         }
1998
1999         for (i = 0; i < vi->max_queue_pairs; i++) {
2000                 vi->rq[i].vq = vqs[rxq2vq(i)];
2001                 vi->sq[i].vq = vqs[txq2vq(i)];
2002         }
2003
2004         kfree(names);
2005         kfree(callbacks);
2006         kfree(vqs);
2007
2008         return 0;
2009
2010 err_find:
2011         kfree(names);
2012 err_names:
2013         kfree(callbacks);
2014 err_callback:
2015         kfree(vqs);
2016 err_vq:
2017         return ret;
2018 }
2019
2020 static int virtnet_alloc_queues(struct virtnet_info *vi)
2021 {
2022         int i;
2023
2024         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2025         if (!vi->sq)
2026                 goto err_sq;
2027         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2028         if (!vi->rq)
2029                 goto err_rq;
2030
2031         INIT_DELAYED_WORK(&vi->refill, refill_work);
2032         for (i = 0; i < vi->max_queue_pairs; i++) {
2033                 vi->rq[i].pages = NULL;
2034                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2035                                napi_weight);
2036
2037                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2038                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2039                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2040         }
2041
2042         return 0;
2043
2044 err_rq:
2045         kfree(vi->sq);
2046 err_sq:
2047         return -ENOMEM;
2048 }
2049
2050 static int init_vqs(struct virtnet_info *vi)
2051 {
2052         int ret;
2053
2054         /* Allocate send & receive queues */
2055         ret = virtnet_alloc_queues(vi);
2056         if (ret)
2057                 goto err;
2058
2059         ret = virtnet_find_vqs(vi);
2060         if (ret)
2061                 goto err_free;
2062
2063         get_online_cpus();
2064         virtnet_set_affinity(vi);
2065         put_online_cpus();
2066
2067         return 0;
2068
2069 err_free:
2070         virtnet_free_queues(vi);
2071 err:
2072         return ret;
2073 }
2074
2075 #ifdef CONFIG_SYSFS
2076 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2077                 struct rx_queue_attribute *attribute, char *buf)
2078 {
2079         struct virtnet_info *vi = netdev_priv(queue->dev);
2080         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2081         struct ewma_pkt_len *avg;
2082
2083         BUG_ON(queue_index >= vi->max_queue_pairs);
2084         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2085         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
2086 }
2087
2088 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2089         __ATTR_RO(mergeable_rx_buffer_size);
2090
2091 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2092         &mergeable_rx_buffer_size_attribute.attr,
2093         NULL
2094 };
2095
2096 static const struct attribute_group virtio_net_mrg_rx_group = {
2097         .name = "virtio_net",
2098         .attrs = virtio_net_mrg_rx_attrs
2099 };
2100 #endif
2101
2102 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2103                                     unsigned int fbit,
2104                                     const char *fname, const char *dname)
2105 {
2106         if (!virtio_has_feature(vdev, fbit))
2107                 return false;
2108
2109         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2110                 fname, dname);
2111
2112         return true;
2113 }
2114
2115 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2116         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2117
2118 static bool virtnet_validate_features(struct virtio_device *vdev)
2119 {
2120         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2121             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2122                              "VIRTIO_NET_F_CTRL_VQ") ||
2123              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2124                              "VIRTIO_NET_F_CTRL_VQ") ||
2125              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2126                              "VIRTIO_NET_F_CTRL_VQ") ||
2127              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2128              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2129                              "VIRTIO_NET_F_CTRL_VQ"))) {
2130                 return false;
2131         }
2132
2133         return true;
2134 }
2135
2136 #define MIN_MTU ETH_MIN_MTU
2137 #define MAX_MTU ETH_MAX_MTU
2138
2139 static int virtnet_probe(struct virtio_device *vdev)
2140 {
2141         int i, err;
2142         struct net_device *dev;
2143         struct virtnet_info *vi;
2144         u16 max_queue_pairs;
2145         int mtu;
2146
2147         if (!vdev->config->get) {
2148                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2149                         __func__);
2150                 return -EINVAL;
2151         }
2152
2153         if (!virtnet_validate_features(vdev))
2154                 return -EINVAL;
2155
2156         /* Find if host supports multiqueue virtio_net device */
2157         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2158                                    struct virtio_net_config,
2159                                    max_virtqueue_pairs, &max_queue_pairs);
2160
2161         /* We need at least 2 queue's */
2162         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2163             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2164             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2165                 max_queue_pairs = 1;
2166
2167         /* Allocate ourselves a network device with room for our info */
2168         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2169         if (!dev)
2170                 return -ENOMEM;
2171
2172         /* Set up network device as normal. */
2173         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2174         dev->netdev_ops = &virtnet_netdev;
2175         dev->features = NETIF_F_HIGHDMA;
2176
2177         dev->ethtool_ops = &virtnet_ethtool_ops;
2178         SET_NETDEV_DEV(dev, &vdev->dev);
2179
2180         /* Do we support "hardware" checksums? */
2181         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2182                 /* This opens up the world of extra features. */
2183                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2184                 if (csum)
2185                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2186
2187                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2188                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
2189                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2190                 }
2191                 /* Individual feature bits: what can host handle? */
2192                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2193                         dev->hw_features |= NETIF_F_TSO;
2194                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2195                         dev->hw_features |= NETIF_F_TSO6;
2196                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2197                         dev->hw_features |= NETIF_F_TSO_ECN;
2198                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
2199                         dev->hw_features |= NETIF_F_UFO;
2200
2201                 dev->features |= NETIF_F_GSO_ROBUST;
2202
2203                 if (gso)
2204                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
2205                 /* (!csum && gso) case will be fixed by register_netdev() */
2206         }
2207         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2208                 dev->features |= NETIF_F_RXCSUM;
2209
2210         dev->vlan_features = dev->features;
2211
2212         /* MTU range: 68 - 65535 */
2213         dev->min_mtu = MIN_MTU;
2214         dev->max_mtu = MAX_MTU;
2215
2216         /* Configuration may specify what MAC to use.  Otherwise random. */
2217         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2218                 virtio_cread_bytes(vdev,
2219                                    offsetof(struct virtio_net_config, mac),
2220                                    dev->dev_addr, dev->addr_len);
2221         else
2222                 eth_hw_addr_random(dev);
2223
2224         /* Set up our device-specific information */
2225         vi = netdev_priv(dev);
2226         vi->dev = dev;
2227         vi->vdev = vdev;
2228         vdev->priv = vi;
2229         vi->stats = alloc_percpu(struct virtnet_stats);
2230         err = -ENOMEM;
2231         if (vi->stats == NULL)
2232                 goto free;
2233
2234         for_each_possible_cpu(i) {
2235                 struct virtnet_stats *virtnet_stats;
2236                 virtnet_stats = per_cpu_ptr(vi->stats, i);
2237                 u64_stats_init(&virtnet_stats->tx_syncp);
2238                 u64_stats_init(&virtnet_stats->rx_syncp);
2239         }
2240
2241         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2242
2243         /* If we can receive ANY GSO packets, we must allocate large ones. */
2244         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2245             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2246             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2247             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2248                 vi->big_packets = true;
2249
2250         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2251                 vi->mergeable_rx_bufs = true;
2252
2253         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2254             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2255                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2256         else
2257                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2258
2259         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2260             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2261                 vi->any_header_sg = true;
2262
2263         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2264                 vi->has_cvq = true;
2265
2266         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2267                 mtu = virtio_cread16(vdev,
2268                                      offsetof(struct virtio_net_config,
2269                                               mtu));
2270                 if (mtu < dev->min_mtu) {
2271                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2272                 } else {
2273                         dev->mtu = mtu;
2274                         dev->max_mtu = mtu;
2275                 }
2276         }
2277
2278         if (vi->any_header_sg)
2279                 dev->needed_headroom = vi->hdr_len;
2280
2281         /* Enable multiqueue by default */
2282         if (num_online_cpus() >= max_queue_pairs)
2283                 vi->curr_queue_pairs = max_queue_pairs;
2284         else
2285                 vi->curr_queue_pairs = num_online_cpus();
2286         vi->max_queue_pairs = max_queue_pairs;
2287
2288         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2289         err = init_vqs(vi);
2290         if (err)
2291                 goto free_stats;
2292
2293 #ifdef CONFIG_SYSFS
2294         if (vi->mergeable_rx_bufs)
2295                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2296 #endif
2297         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2298         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2299
2300         virtnet_init_settings(dev);
2301
2302         err = register_netdev(dev);
2303         if (err) {
2304                 pr_debug("virtio_net: registering device failed\n");
2305                 goto free_vqs;
2306         }
2307
2308         virtio_device_ready(vdev);
2309
2310         err = virtnet_cpu_notif_add(vi);
2311         if (err) {
2312                 pr_debug("virtio_net: registering cpu notifier failed\n");
2313                 goto free_unregister_netdev;
2314         }
2315
2316         rtnl_lock();
2317         virtnet_set_queues(vi, vi->curr_queue_pairs);
2318         rtnl_unlock();
2319
2320         /* Assume link up if device can't report link status,
2321            otherwise get link status from config. */
2322         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2323                 netif_carrier_off(dev);
2324                 schedule_work(&vi->config_work);
2325         } else {
2326                 vi->status = VIRTIO_NET_S_LINK_UP;
2327                 netif_carrier_on(dev);
2328         }
2329
2330         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2331                  dev->name, max_queue_pairs);
2332
2333         return 0;
2334
2335 free_unregister_netdev:
2336         vi->vdev->config->reset(vdev);
2337
2338         unregister_netdev(dev);
2339 free_vqs:
2340         cancel_delayed_work_sync(&vi->refill);
2341         free_receive_page_frags(vi);
2342         virtnet_del_vqs(vi);
2343 free_stats:
2344         free_percpu(vi->stats);
2345 free:
2346         free_netdev(dev);
2347         return err;
2348 }
2349
2350 static void remove_vq_common(struct virtnet_info *vi)
2351 {
2352         vi->vdev->config->reset(vi->vdev);
2353
2354         /* Free unused buffers in both send and recv, if any. */
2355         free_unused_bufs(vi);
2356
2357         free_receive_bufs(vi);
2358
2359         free_receive_page_frags(vi);
2360
2361         virtnet_del_vqs(vi);
2362 }
2363
2364 static void virtnet_remove(struct virtio_device *vdev)
2365 {
2366         struct virtnet_info *vi = vdev->priv;
2367
2368         virtnet_cpu_notif_remove(vi);
2369
2370         /* Make sure no work handler is accessing the device. */
2371         flush_work(&vi->config_work);
2372
2373         unregister_netdev(vi->dev);
2374
2375         remove_vq_common(vi);
2376
2377         free_percpu(vi->stats);
2378         free_netdev(vi->dev);
2379 }
2380
2381 #ifdef CONFIG_PM_SLEEP
2382 static int virtnet_freeze(struct virtio_device *vdev)
2383 {
2384         struct virtnet_info *vi = vdev->priv;
2385         int i;
2386
2387         virtnet_cpu_notif_remove(vi);
2388
2389         /* Make sure no work handler is accessing the device */
2390         flush_work(&vi->config_work);
2391
2392         netif_device_detach(vi->dev);
2393         cancel_delayed_work_sync(&vi->refill);
2394
2395         if (netif_running(vi->dev)) {
2396                 for (i = 0; i < vi->max_queue_pairs; i++)
2397                         napi_disable(&vi->rq[i].napi);
2398         }
2399
2400         remove_vq_common(vi);
2401
2402         return 0;
2403 }
2404
2405 static int virtnet_restore(struct virtio_device *vdev)
2406 {
2407         struct virtnet_info *vi = vdev->priv;
2408         int err, i;
2409
2410         err = init_vqs(vi);
2411         if (err)
2412                 return err;
2413
2414         virtio_device_ready(vdev);
2415
2416         if (netif_running(vi->dev)) {
2417                 for (i = 0; i < vi->curr_queue_pairs; i++)
2418                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2419                                 schedule_delayed_work(&vi->refill, 0);
2420
2421                 for (i = 0; i < vi->max_queue_pairs; i++)
2422                         virtnet_napi_enable(&vi->rq[i]);
2423         }
2424
2425         netif_device_attach(vi->dev);
2426
2427         rtnl_lock();
2428         virtnet_set_queues(vi, vi->curr_queue_pairs);
2429         rtnl_unlock();
2430
2431         err = virtnet_cpu_notif_add(vi);
2432         if (err)
2433                 return err;
2434
2435         return 0;
2436 }
2437 #endif
2438
2439 static struct virtio_device_id id_table[] = {
2440         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2441         { 0 },
2442 };
2443
2444 #define VIRTNET_FEATURES \
2445         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2446         VIRTIO_NET_F_MAC, \
2447         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2448         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2449         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2450         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2451         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2452         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2453         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2454         VIRTIO_NET_F_MTU
2455
2456 static unsigned int features[] = {
2457         VIRTNET_FEATURES,
2458 };
2459
2460 static unsigned int features_legacy[] = {
2461         VIRTNET_FEATURES,
2462         VIRTIO_NET_F_GSO,
2463         VIRTIO_F_ANY_LAYOUT,
2464 };
2465
2466 static struct virtio_driver virtio_net_driver = {
2467         .feature_table = features,
2468         .feature_table_size = ARRAY_SIZE(features),
2469         .feature_table_legacy = features_legacy,
2470         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2471         .driver.name =  KBUILD_MODNAME,
2472         .driver.owner = THIS_MODULE,
2473         .id_table =     id_table,
2474         .probe =        virtnet_probe,
2475         .remove =       virtnet_remove,
2476         .config_changed = virtnet_config_changed,
2477 #ifdef CONFIG_PM_SLEEP
2478         .freeze =       virtnet_freeze,
2479         .restore =      virtnet_restore,
2480 #endif
2481 };
2482
2483 static __init int virtio_net_driver_init(void)
2484 {
2485         int ret;
2486
2487         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2488                                       virtnet_cpu_online,
2489                                       virtnet_cpu_down_prep);
2490         if (ret < 0)
2491                 goto out;
2492         virtionet_online = ret;
2493         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2494                                       NULL, virtnet_cpu_dead);
2495         if (ret)
2496                 goto err_dead;
2497
2498         ret = register_virtio_driver(&virtio_net_driver);
2499         if (ret)
2500                 goto err_virtio;
2501         return 0;
2502 err_virtio:
2503         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2504 err_dead:
2505         cpuhp_remove_multi_state(virtionet_online);
2506 out:
2507         return ret;
2508 }
2509 module_init(virtio_net_driver_init);
2510
2511 static __exit void virtio_net_driver_exit(void)
2512 {
2513         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2514         cpuhp_remove_multi_state(virtionet_online);
2515         unregister_virtio_driver(&virtio_net_driver);
2516 }
2517 module_exit(virtio_net_driver_exit);
2518
2519 MODULE_DEVICE_TABLE(virtio, id_table);
2520 MODULE_DESCRIPTION("Virtio network driver");
2521 MODULE_LICENSE("GPL");