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 /* With mergeable buffers we align buffer address and use the low bits to
  52  * encode its true size. Buffer size is up to 1 page so we need to align to
  53  * square root of page size to ensure we reserve enough bits to encode the true
  54  * size.
  55  */
  56 #define MERGEABLE_BUFFER_MIN_ALIGN_SHIFT ((PAGE_SHIFT + 1) / 2)
  57
  58 /* Minimum alignment for mergeable packet buffers. */
  59 #define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, \
  60                                    1 << MERGEABLE_BUFFER_MIN_ALIGN_SHIFT)
  61
  62 #define VIRTNET_DRIVER_VERSION "1.0.0"
  63
  64 struct virtnet_stats {
  65         struct u64_stats_sync tx_syncp;
  66         struct u64_stats_sync rx_syncp;
  67         u64 tx_bytes;
  68         u64 tx_packets;
  69
  70         u64 rx_bytes;
  71         u64 rx_packets;
  72 };
  73
  74 /* Internal representation of a send virtqueue */
  75 struct send_queue {
  76         /* Virtqueue associated with this send _queue */
  77         struct virtqueue *vq;
  78
  79         /* TX: fragments + linear part + virtio header */
  80         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  81
  82         /* Name of the send queue: output.$index */
  83         char name[40];
  84 };
  85
  86 /* Internal representation of a receive virtqueue */
  87 struct receive_queue {
  88         /* Virtqueue associated with this receive_queue */
  89         struct virtqueue *vq;
  90
  91         struct napi_struct napi;
  92
  93         struct bpf_prog __rcu *xdp_prog;
  94
  95         /* Chain pages by the private ptr. */
  96         struct page *pages;
  97
  98         /* Average packet length for mergeable receive buffers. */
  99         struct ewma_pkt_len mrg_avg_pkt_len;
 100
 101         /* Page frag for packet buffer allocation. */
 102         struct page_frag alloc_frag;
 103
 104         /* RX: fragments + linear part + virtio header */
 105         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 106
 107         /* Name of this receive queue: input.$index */
 108         char name[40];
 109 };
 110
 111 struct virtnet_info {
 112         struct virtio_device *vdev;
 113         struct virtqueue *cvq;
 114         struct net_device *dev;
 115         struct send_queue *sq;
 116         struct receive_queue *rq;
 117         unsigned int status;
 118
 119         /* Max # of queue pairs supported by the device */
 120         u16 max_queue_pairs;
 121
 122         /* # of queue pairs currently used by the driver */
 123         u16 curr_queue_pairs;
 124
 125         /* # of XDP queue pairs currently used by the driver */
 126         u16 xdp_queue_pairs;
 127
 128         /* I like... big packets and I cannot lie! */
 129         bool big_packets;
 130
 131         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 132         bool mergeable_rx_bufs;
 133
 134         /* Has control virtqueue */
 135         bool has_cvq;
 136
 137         /* Host can handle any s/g split between our header and packet data */
 138         bool any_header_sg;
 139
 140         /* Packet virtio header size */
 141         u8 hdr_len;
 142
 143         /* Active statistics */
 144         struct virtnet_stats __percpu *stats;
 145
 146         /* Work struct for refilling if we run low on memory. */
 147         struct delayed_work refill;
 148
 149         /* Work struct for config space updates */
 150         struct work_struct config_work;
 151
 152         /* Does the affinity hint is set for virtqueues? */
 153         bool affinity_hint_set;
 154
 155         /* CPU hotplug instances for online & dead */
 156         struct hlist_node node;
 157         struct hlist_node node_dead;
 158
 159         /* Control VQ buffers: protected by the rtnl lock */
 160         struct virtio_net_ctrl_hdr ctrl_hdr;
 161         virtio_net_ctrl_ack ctrl_status;
 162         struct virtio_net_ctrl_mq ctrl_mq;
 163         u8 ctrl_promisc;
 164         u8 ctrl_allmulti;
 165         u16 ctrl_vid;
 166
 167         /* Ethtool settings */
 168         u8 duplex;
 169         u32 speed;
 170 };
 171
 172 struct padded_vnet_hdr {
 173         struct virtio_net_hdr_mrg_rxbuf hdr;
 174         /*
 175          * hdr is in a separate sg buffer, and data sg buffer shares same page
 176          * with this header sg. This padding makes next sg 16 byte aligned
 177          * after the header.
 178          */
 179         char padding[4];
 180 };
 181
 182 /* Converting between virtqueue no. and kernel tx/rx queue no.
 183  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 184  */
 185 static int vq2txq(struct virtqueue *vq)
 186 {
 187         return (vq->index - 1) / 2;
 188 }
 189
 190 static int txq2vq(int txq)
 191 {
 192         return txq * 2 + 1;
 193 }
 194
 195 static int vq2rxq(struct virtqueue *vq)
 196 {
 197         return vq->index / 2;
 198 }
 199
 200 static int rxq2vq(int rxq)
 201 {
 202         return rxq * 2;
 203 }
 204
 205 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 206 {
 207         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 208 }
 209
 210 /*
 211  * private is used to chain pages for big packets, put the whole
 212  * most recent used list in the beginning for reuse
 213  */
 214 static void give_pages(struct receive_queue *rq, struct page *page)
 215 {
 216         struct page *end;
 217
 218         /* Find end of list, sew whole thing into vi->rq.pages. */
 219         for (end = page; end->private; end = (struct page *)end->private);
 220         end->private = (unsigned long)rq->pages;
 221         rq->pages = page;
 222 }
 223
 224 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 225 {
 226         struct page *p = rq->pages;
 227
 228         if (p) {
 229                 rq->pages = (struct page *)p->private;
 230                 /* clear private here, it is used to chain pages */
 231                 p->private = 0;
 232         } else
 233                 p = alloc_page(gfp_mask);
 234         return p;
 235 }
 236
 237 static void skb_xmit_done(struct virtqueue *vq)
 238 {
 239         struct virtnet_info *vi = vq->vdev->priv;
 240
 241         /* Suppress further interrupts. */
 242         virtqueue_disable_cb(vq);
 243
 244         /* We were probably waiting for more output buffers. */
 245         netif_wake_subqueue(vi->dev, vq2txq(vq));
 246 }
 247
 248 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
 249 {
 250         unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
 251         return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
 252 }
 253
 254 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
 255 {
 256         return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
 257
 258 }
 259
 260 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
 261 {
 262         unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
 263         return (unsigned long)buf | (size - 1);
 264 }
 265
 266 /* Called from bottom half context */
 267 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 268                                    struct receive_queue *rq,
 269                                    struct page *page, unsigned int offset,
 270                                    unsigned int len, unsigned int truesize)
 271 {
 272         struct sk_buff *skb;
 273         struct virtio_net_hdr_mrg_rxbuf *hdr;
 274         unsigned int copy, hdr_len, hdr_padded_len;
 275         char *p;
 276
 277         p = page_address(page) + offset;
 278
 279         /* copy small packet so we can reuse these pages for small data */
 280         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 281         if (unlikely(!skb))
 282                 return NULL;
 283
 284         hdr = skb_vnet_hdr(skb);
 285
 286         hdr_len = vi->hdr_len;
 287         if (vi->mergeable_rx_bufs)
 288                 hdr_padded_len = sizeof *hdr;
 289         else
 290                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 291
 292         memcpy(hdr, p, hdr_len);
 293
 294         len -= hdr_len;
 295         offset += hdr_padded_len;
 296         p += hdr_padded_len;
 297
 298         copy = len;
 299         if (copy > skb_tailroom(skb))
 300                 copy = skb_tailroom(skb);
 301         memcpy(skb_put(skb, copy), p, copy);
 302
 303         len -= copy;
 304         offset += copy;
 305
 306         if (vi->mergeable_rx_bufs) {
 307                 if (len)
 308                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 309                 else
 310                         put_page(page);
 311                 return skb;
 312         }
 313
 314         /*
 315          * Verify that we can indeed put this data into a skb.
 316          * This is here to handle cases when the device erroneously
 317          * tries to receive more than is possible. This is usually
 318          * the case of a broken device.
 319          */
 320         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 321                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 322                 dev_kfree_skb(skb);
 323                 return NULL;
 324         }
 325         BUG_ON(offset >= PAGE_SIZE);
 326         while (len) {
 327                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 328                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 329                                 frag_size, truesize);
 330                 len -= frag_size;
 331                 page = (struct page *)page->private;
 332                 offset = 0;
 333         }
 334
 335         if (page)
 336                 give_pages(rq, page);
 337
 338         return skb;
 339 }
 340
 341 static void virtnet_xdp_xmit(struct virtnet_info *vi,
 342                              struct receive_queue *rq,
 343                              struct send_queue *sq,
 344                              struct xdp_buff *xdp,
 345                              void *data)
 346 {
 347         struct virtio_net_hdr_mrg_rxbuf *hdr;
 348         unsigned int num_sg, len;
 349         void *xdp_sent;
 350         int err;
 351
 352         /* Free up any pending old buffers before queueing new ones. */
 353         while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 354                 if (vi->mergeable_rx_bufs) {
 355                         struct page *sent_page = virt_to_head_page(xdp_sent);
 356
 357                         put_page(sent_page);
 358                 } else { /* small buffer */
 359                         struct sk_buff *skb = xdp_sent;
 360
 361                         kfree_skb(skb);
 362                 }
 363         }
 364
 365         if (vi->mergeable_rx_bufs) {
 366                 /* Zero header and leave csum up to XDP layers */
 367                 hdr = xdp->data;
 368                 memset(hdr, 0, vi->hdr_len);
 369
 370                 num_sg = 1;
 371                 sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
 372         } else { /* small buffer */
 373                 struct sk_buff *skb = data;
 374
 375                 /* Zero header and leave csum up to XDP layers */
 376                 hdr = skb_vnet_hdr(skb);
 377                 memset(hdr, 0, vi->hdr_len);
 378
 379                 num_sg = 2;
 380                 sg_init_table(sq->sg, 2);
 381                 sg_set_buf(sq->sg, hdr, vi->hdr_len);
 382                 skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
 383         }
 384         err = virtqueue_add_outbuf(sq->vq, sq->sg, num_sg,
 385                                    data, GFP_ATOMIC);
 386         if (unlikely(err)) {
 387                 if (vi->mergeable_rx_bufs) {
 388                         struct page *page = virt_to_head_page(xdp->data);
 389
 390                         put_page(page);
 391                 } else /* small buffer */
 392                         kfree_skb(data);
 393                 return; // On error abort to avoid unnecessary kick
 394         }
 395
 396         virtqueue_kick(sq->vq);
 397 }
 398
 399 static u32 do_xdp_prog(struct virtnet_info *vi,
 400                        struct receive_queue *rq,
 401                        struct bpf_prog *xdp_prog,
 402                        void *data, int len)
 403 {
 404         int hdr_padded_len;
 405         struct xdp_buff xdp;
 406         void *buf;
 407         unsigned int qp;
 408         u32 act;
 409
 410         if (vi->mergeable_rx_bufs) {
 411                 hdr_padded_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 412                 xdp.data = data + hdr_padded_len;
 413                 xdp.data_end = xdp.data + (len - vi->hdr_len);
 414                 buf = data;
 415         } else { /* small buffers */
 416                 struct sk_buff *skb = data;
 417
 418                 xdp.data = skb->data;
 419                 xdp.data_end = xdp.data + len;
 420                 buf = skb->data;
 421         }
 422
 423         act = bpf_prog_run_xdp(xdp_prog, &xdp);
 424         switch (act) {
 425         case XDP_PASS:
 426                 return XDP_PASS;
 427         case XDP_TX:
 428                 qp = vi->curr_queue_pairs -
 429                         vi->xdp_queue_pairs +
 430                         smp_processor_id();
 431                 xdp.data = buf;
 432                 virtnet_xdp_xmit(vi, rq, &vi->sq[qp], &xdp, data);
 433                 return XDP_TX;
 434         default:
 435                 bpf_warn_invalid_xdp_action(act);
 436         case XDP_ABORTED:
 437         case XDP_DROP:
 438                 return XDP_DROP;
 439         }
 440 }
 441
 442 static struct sk_buff *receive_small(struct net_device *dev,
 443                                      struct virtnet_info *vi,
 444                                      struct receive_queue *rq,
 445                                      void *buf, unsigned int len)
 446 {
 447         struct sk_buff * skb = buf;
 448         struct bpf_prog *xdp_prog;
 449
 450         len -= vi->hdr_len;
 451         skb_trim(skb, len);
 452
 453         rcu_read_lock();
 454         xdp_prog = rcu_dereference(rq->xdp_prog);
 455         if (xdp_prog) {
 456                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 457                 u32 act;
 458
 459                 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
 460                         goto err_xdp;
 461                 act = do_xdp_prog(vi, rq, xdp_prog, skb, len);
 462                 switch (act) {
 463                 case XDP_PASS:
 464                         break;
 465                 case XDP_TX:
 466                         rcu_read_unlock();
 467                         goto xdp_xmit;
 468                 case XDP_DROP:
 469                 default:
 470                         goto err_xdp;
 471                 }
 472         }
 473         rcu_read_unlock();
 474
 475         return skb;
 476
 477 err_xdp:
 478         rcu_read_unlock();
 479         dev->stats.rx_dropped++;
 480         kfree_skb(skb);
 481 xdp_xmit:
 482         return NULL;
 483 }
 484
 485 static struct sk_buff *receive_big(struct net_device *dev,
 486                                    struct virtnet_info *vi,
 487                                    struct receive_queue *rq,
 488                                    void *buf,
 489                                    unsigned int len)
 490 {
 491         struct page *page = buf;
 492         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 493
 494         if (unlikely(!skb))
 495                 goto err;
 496
 497         return skb;
 498
 499 err:
 500         dev->stats.rx_dropped++;
 501         give_pages(rq, page);
 502         return NULL;
 503 }
 504
 505 /* The conditions to enable XDP should preclude the underlying device from
 506  * sending packets across multiple buffers (num_buf > 1). However per spec
 507  * it does not appear to be illegal to do so but rather just against convention.
 508  * So in order to avoid making a system unresponsive the packets are pushed
 509  * into a page and the XDP program is run. This will be extremely slow and we
 510  * push a warning to the user to fix this as soon as possible. Fixing this may
 511  * require resolving the underlying hardware to determine why multiple buffers
 512  * are being received or simply loading the XDP program in the ingress stack
 513  * after the skb is built because there is no advantage to running it here
 514  * anymore.
 515  */
 516 static struct page *xdp_linearize_page(struct receive_queue *rq,
 517                                        u16 *num_buf,
 518                                        struct page *p,
 519                                        int offset,
 520                                        unsigned int *len)
 521 {
 522         struct page *page = alloc_page(GFP_ATOMIC);
 523         unsigned int page_off = 0;
 524
 525         if (!page)
 526                 return NULL;
 527
 528         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 529         page_off += *len;
 530
 531         while (--*num_buf) {
 532                 unsigned int buflen;
 533                 unsigned long ctx;
 534                 void *buf;
 535                 int off;
 536
 537                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &buflen);
 538                 if (unlikely(!ctx))
 539                         goto err_buf;
 540
 541                 buf = mergeable_ctx_to_buf_address(ctx);
 542                 p = virt_to_head_page(buf);
 543                 off = buf - page_address(p);
 544
 545                 /* guard against a misconfigured or uncooperative backend that
 546                  * is sending packet larger than the MTU.
 547                  */
 548                 if ((page_off + buflen) > PAGE_SIZE) {
 549                         put_page(p);
 550                         goto err_buf;
 551                 }
 552
 553                 memcpy(page_address(page) + page_off,
 554                        page_address(p) + off, buflen);
 555                 page_off += buflen;
 556                 put_page(p);
 557         }
 558
 559         *len = page_off;
 560         return page;
 561 err_buf:
 562         __free_pages(page, 0);
 563         return NULL;
 564 }
 565
 566 static struct sk_buff *receive_mergeable(struct net_device *dev,
 567                                          struct virtnet_info *vi,
 568                                          struct receive_queue *rq,
 569                                          unsigned long ctx,
 570                                          unsigned int len)
 571 {
 572         void *buf = mergeable_ctx_to_buf_address(ctx);
 573         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 574         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 575         struct page *page = virt_to_head_page(buf);
 576         int offset = buf - page_address(page);
 577         struct sk_buff *head_skb, *curr_skb;
 578         struct bpf_prog *xdp_prog;
 579         unsigned int truesize;
 580
 581         head_skb = NULL;
 582
 583         rcu_read_lock();
 584         xdp_prog = rcu_dereference(rq->xdp_prog);
 585         if (xdp_prog) {
 586                 struct page *xdp_page;
 587                 u32 act;
 588
 589                 /* This happens when rx buffer size is underestimated */
 590                 if (unlikely(num_buf > 1)) {
 591                         /* linearize data for XDP */
 592                         xdp_page = xdp_linearize_page(rq, &num_buf,
 593                                                       page, offset, &len);
 594                         if (!xdp_page)
 595                                 goto err_xdp;
 596                         offset = 0;
 597                 } else {
 598                         xdp_page = page;
 599                 }
 600
 601                 /* Transient failure which in theory could occur if
 602                  * in-flight packets from before XDP was enabled reach
 603                  * the receive path after XDP is loaded. In practice I
 604                  * was not able to create this condition.
 605                  */
 606                 if (unlikely(hdr->hdr.gso_type))
 607                         goto err_xdp;
 608
 609                 act = do_xdp_prog(vi, rq, xdp_prog,
 610                                   page_address(xdp_page) + offset, len);
 611                 switch (act) {
 612                 case XDP_PASS:
 613                         /* We can only create skb based on xdp_page. */
 614                         if (unlikely(xdp_page != page)) {
 615                                 rcu_read_unlock();
 616                                 put_page(page);
 617                                 head_skb = page_to_skb(vi, rq, xdp_page,
 618                                                        0, len, PAGE_SIZE);
 619                                 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 620                                 return head_skb;
 621                         }
 622                         break;
 623                 case XDP_TX:
 624                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 625                         if (unlikely(xdp_page != page))
 626                                 goto err_xdp;
 627                         rcu_read_unlock();
 628                         goto xdp_xmit;
 629                 case XDP_DROP:
 630                 default:
 631                         if (unlikely(xdp_page != page))
 632                                 __free_pages(xdp_page, 0);
 633                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 634                         goto err_xdp;
 635                 }
 636         }
 637         rcu_read_unlock();
 638
 639         truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 640         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
 641         curr_skb = head_skb;
 642
 643         if (unlikely(!curr_skb))
 644                 goto err_skb;
 645         while (--num_buf) {
 646                 int num_skb_frags;
 647
 648                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 649                 if (unlikely(!ctx)) {
 650                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 651                                  dev->name, num_buf,
 652                                  virtio16_to_cpu(vi->vdev,
 653                                                  hdr->num_buffers));
 654                         dev->stats.rx_length_errors++;
 655                         goto err_buf;
 656                 }
 657
 658                 buf = mergeable_ctx_to_buf_address(ctx);
 659                 page = virt_to_head_page(buf);
 660
 661                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 662                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 663                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 664
 665                         if (unlikely(!nskb))
 666                                 goto err_skb;
 667                         if (curr_skb == head_skb)
 668                                 skb_shinfo(curr_skb)->frag_list = nskb;
 669                         else
 670                                 curr_skb->next = nskb;
 671                         curr_skb = nskb;
 672                         head_skb->truesize += nskb->truesize;
 673                         num_skb_frags = 0;
 674                 }
 675                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 676                 if (curr_skb != head_skb) {
 677                         head_skb->data_len += len;
 678                         head_skb->len += len;
 679                         head_skb->truesize += truesize;
 680                 }
 681                 offset = buf - page_address(page);
 682                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 683                         put_page(page);
 684                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 685                                              len, truesize);
 686                 } else {
 687                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 688                                         offset, len, truesize);
 689                 }
 690         }
 691
 692         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 693         return head_skb;
 694
 695 err_xdp:
 696         rcu_read_unlock();
 697 err_skb:
 698         put_page(page);
 699         while (--num_buf) {
 700                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 701                 if (unlikely(!ctx)) {
 702                         pr_debug("%s: rx error: %d buffers missing\n",
 703                                  dev->name, num_buf);
 704                         dev->stats.rx_length_errors++;
 705                         break;
 706                 }
 707                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 708                 put_page(page);
 709         }
 710 err_buf:
 711         dev->stats.rx_dropped++;
 712         dev_kfree_skb(head_skb);
 713 xdp_xmit:
 714         return NULL;
 715 }
 716
 717 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 718                         void *buf, unsigned int len)
 719 {
 720         struct net_device *dev = vi->dev;
 721         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 722         struct sk_buff *skb;
 723         struct virtio_net_hdr_mrg_rxbuf *hdr;
 724
 725         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 726                 pr_debug("%s: short packet %i\n", dev->name, len);
 727                 dev->stats.rx_length_errors++;
 728                 if (vi->mergeable_rx_bufs) {
 729                         unsigned long ctx = (unsigned long)buf;
 730                         void *base = mergeable_ctx_to_buf_address(ctx);
 731                         put_page(virt_to_head_page(base));
 732                 } else if (vi->big_packets) {
 733                         give_pages(rq, buf);
 734                 } else {
 735                         dev_kfree_skb(buf);
 736                 }
 737                 return;
 738         }
 739
 740         if (vi->mergeable_rx_bufs)
 741                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
 742         else if (vi->big_packets)
 743                 skb = receive_big(dev, vi, rq, buf, len);
 744         else
 745                 skb = receive_small(dev, vi, rq, buf, len);
 746
 747         if (unlikely(!skb))
 748                 return;
 749
 750         hdr = skb_vnet_hdr(skb);
 751
 752         u64_stats_update_begin(&stats->rx_syncp);
 753         stats->rx_bytes += skb->len;
 754         stats->rx_packets++;
 755         u64_stats_update_end(&stats->rx_syncp);
 756
 757         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 758                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 759
 760         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 761                                   virtio_is_little_endian(vi->vdev))) {
 762                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 763                                      dev->name, hdr->hdr.gso_type,
 764                                      hdr->hdr.gso_size);
 765                 goto frame_err;
 766         }
 767
 768         skb->protocol = eth_type_trans(skb, dev);
 769         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 770                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 771
 772         napi_gro_receive(&rq->napi, skb);
 773         return;
 774
 775 frame_err:
 776         dev->stats.rx_frame_errors++;
 777         dev_kfree_skb(skb);
 778 }
 779
 780 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 781                              gfp_t gfp)
 782 {
 783         struct sk_buff *skb;
 784         struct virtio_net_hdr_mrg_rxbuf *hdr;
 785         int err;
 786
 787         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 788         if (unlikely(!skb))
 789                 return -ENOMEM;
 790
 791         skb_put(skb, GOOD_PACKET_LEN);
 792
 793         hdr = skb_vnet_hdr(skb);
 794         sg_init_table(rq->sg, 2);
 795         sg_set_buf(rq->sg, hdr, vi->hdr_len);
 796         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 797
 798         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 799         if (err < 0)
 800                 dev_kfree_skb(skb);
 801
 802         return err;
 803 }
 804
 805 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 806                            gfp_t gfp)
 807 {
 808         struct page *first, *list = NULL;
 809         char *p;
 810         int i, err, offset;
 811
 812         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 813
 814         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 815         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 816                 first = get_a_page(rq, gfp);
 817                 if (!first) {
 818                         if (list)
 819                                 give_pages(rq, list);
 820                         return -ENOMEM;
 821                 }
 822                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 823
 824                 /* chain new page in list head to match sg */
 825                 first->private = (unsigned long)list;
 826                 list = first;
 827         }
 828
 829         first = get_a_page(rq, gfp);
 830         if (!first) {
 831                 give_pages(rq, list);
 832                 return -ENOMEM;
 833         }
 834         p = page_address(first);
 835
 836         /* rq->sg[0], rq->sg[1] share the same page */
 837         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 838         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 839
 840         /* rq->sg[1] for data packet, from offset */
 841         offset = sizeof(struct padded_vnet_hdr);
 842         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 843
 844         /* chain first in list head */
 845         first->private = (unsigned long)list;
 846         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 847                                   first, gfp);
 848         if (err < 0)
 849                 give_pages(rq, first);
 850
 851         return err;
 852 }
 853
 854 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
 855 {
 856         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 857         unsigned int len;
 858
 859         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
 860                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 861         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 862 }
 863
 864 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 865 {
 866         struct page_frag *alloc_frag = &rq->alloc_frag;
 867         char *buf;
 868         unsigned long ctx;
 869         int err;
 870         unsigned int len, hole;
 871
 872         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 873         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 874                 return -ENOMEM;
 875
 876         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 877         ctx = mergeable_buf_to_ctx(buf, len);
 878         get_page(alloc_frag->page);
 879         alloc_frag->offset += len;
 880         hole = alloc_frag->size - alloc_frag->offset;
 881         if (hole < len) {
 882                 /* To avoid internal fragmentation, if there is very likely not
 883                  * enough space for another buffer, add the remaining space to
 884                  * the current buffer. This extra space is not included in
 885                  * the truesize stored in ctx.
 886                  */
 887                 len += hole;
 888                 alloc_frag->offset += hole;
 889         }
 890
 891         sg_init_one(rq->sg, buf, len);
 892         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 893         if (err < 0)
 894                 put_page(virt_to_head_page(buf));
 895
 896         return err;
 897 }
 898
 899 /*
 900  * Returns false if we couldn't fill entirely (OOM).
 901  *
 902  * Normally run in the receive path, but can also be run from ndo_open
 903  * before we're receiving packets, or from refill_work which is
 904  * careful to disable receiving (using napi_disable).
 905  */
 906 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 907                           gfp_t gfp)
 908 {
 909         int err;
 910         bool oom;
 911
 912         gfp |= __GFP_COLD;
 913         do {
 914                 if (vi->mergeable_rx_bufs)
 915                         err = add_recvbuf_mergeable(rq, gfp);
 916                 else if (vi->big_packets)
 917                         err = add_recvbuf_big(vi, rq, gfp);
 918                 else
 919                         err = add_recvbuf_small(vi, rq, gfp);
 920
 921                 oom = err == -ENOMEM;
 922                 if (err)
 923                         break;
 924         } while (rq->vq->num_free);
 925         virtqueue_kick(rq->vq);
 926         return !oom;
 927 }
 928
 929 static void skb_recv_done(struct virtqueue *rvq)
 930 {
 931         struct virtnet_info *vi = rvq->vdev->priv;
 932         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 933
 934         /* Schedule NAPI, Suppress further interrupts if successful. */
 935         if (napi_schedule_prep(&rq->napi)) {
 936                 virtqueue_disable_cb(rvq);
 937                 __napi_schedule(&rq->napi);
 938         }
 939 }
 940
 941 static void virtnet_napi_enable(struct receive_queue *rq)
 942 {
 943         napi_enable(&rq->napi);
 944
 945         /* If all buffers were filled by other side before we napi_enabled, we
 946          * won't get another interrupt, so process any outstanding packets
 947          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 948          * We synchronize against interrupts via NAPI_STATE_SCHED */
 949         if (napi_schedule_prep(&rq->napi)) {
 950                 virtqueue_disable_cb(rq->vq);
 951                 local_bh_disable();
 952                 __napi_schedule(&rq->napi);
 953                 local_bh_enable();
 954         }
 955 }
 956
 957 static void refill_work(struct work_struct *work)
 958 {
 959         struct virtnet_info *vi =
 960                 container_of(work, struct virtnet_info, refill.work);
 961         bool still_empty;
 962         int i;
 963
 964         for (i = 0; i < vi->curr_queue_pairs; i++) {
 965                 struct receive_queue *rq = &vi->rq[i];
 966
 967                 napi_disable(&rq->napi);
 968                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
 969                 virtnet_napi_enable(rq);
 970
 971                 /* In theory, this can happen: if we don't get any buffers in
 972                  * we will *never* try to fill again.
 973                  */
 974                 if (still_empty)
 975                         schedule_delayed_work(&vi->refill, HZ/2);
 976         }
 977 }
 978
 979 static int virtnet_receive(struct receive_queue *rq, int budget)
 980 {
 981         struct virtnet_info *vi = rq->vq->vdev->priv;
 982         unsigned int len, received = 0;
 983         void *buf;
 984
 985         while (received < budget &&
 986                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 987                 receive_buf(vi, rq, buf, len);
 988                 received++;
 989         }
 990
 991         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 992                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 993                         schedule_delayed_work(&vi->refill, 0);
 994         }
 995
 996         return received;
 997 }
 998
 999 static int virtnet_poll(struct napi_struct *napi, int budget)
1000 {
1001         struct receive_queue *rq =
1002                 container_of(napi, struct receive_queue, napi);
1003         unsigned int r, received;
1004
1005         received = virtnet_receive(rq, budget);
1006
1007         /* Out of packets? */
1008         if (received < budget) {
1009                 r = virtqueue_enable_cb_prepare(rq->vq);
1010                 napi_complete_done(napi, received);
1011                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
1012                     napi_schedule_prep(napi)) {
1013                         virtqueue_disable_cb(rq->vq);
1014                         __napi_schedule(napi);
1015                 }
1016         }
1017
1018         return received;
1019 }
1020
1021 #ifdef CONFIG_NET_RX_BUSY_POLL
1022 /* must be called with local_bh_disable()d */
1023 static int virtnet_busy_poll(struct napi_struct *napi)
1024 {
1025         struct receive_queue *rq =
1026                 container_of(napi, struct receive_queue, napi);
1027         struct virtnet_info *vi = rq->vq->vdev->priv;
1028         int r, received = 0, budget = 4;
1029
1030         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
1031                 return LL_FLUSH_FAILED;
1032
1033         if (!napi_schedule_prep(napi))
1034                 return LL_FLUSH_BUSY;
1035
1036         virtqueue_disable_cb(rq->vq);
1037
1038 again:
1039         received += virtnet_receive(rq, budget);
1040
1041         r = virtqueue_enable_cb_prepare(rq->vq);
1042         clear_bit(NAPI_STATE_SCHED, &napi->state);
1043         if (unlikely(virtqueue_poll(rq->vq, r)) &&
1044             napi_schedule_prep(napi)) {
1045                 virtqueue_disable_cb(rq->vq);
1046                 if (received < budget) {
1047                         budget -= received;
1048                         goto again;
1049                 } else {
1050                         __napi_schedule(napi);
1051                 }
1052         }
1053
1054         return received;
1055 }
1056 #endif  /* CONFIG_NET_RX_BUSY_POLL */
1057
1058 static int virtnet_open(struct net_device *dev)
1059 {
1060         struct virtnet_info *vi = netdev_priv(dev);
1061         int i;
1062
1063         for (i = 0; i < vi->max_queue_pairs; i++) {
1064                 if (i < vi->curr_queue_pairs)
1065                         /* Make sure we have some buffers: if oom use wq. */
1066                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1067                                 schedule_delayed_work(&vi->refill, 0);
1068                 virtnet_napi_enable(&vi->rq[i]);
1069         }
1070
1071         return 0;
1072 }
1073
1074 static void free_old_xmit_skbs(struct send_queue *sq)
1075 {
1076         struct sk_buff *skb;
1077         unsigned int len;
1078         struct virtnet_info *vi = sq->vq->vdev->priv;
1079         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1080
1081         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1082                 pr_debug("Sent skb %p\n", skb);
1083
1084                 u64_stats_update_begin(&stats->tx_syncp);
1085                 stats->tx_bytes += skb->len;
1086                 stats->tx_packets++;
1087                 u64_stats_update_end(&stats->tx_syncp);
1088
1089                 dev_kfree_skb_any(skb);
1090         }
1091 }
1092
1093 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1094 {
1095         struct virtio_net_hdr_mrg_rxbuf *hdr;
1096         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1097         struct virtnet_info *vi = sq->vq->vdev->priv;
1098         unsigned num_sg;
1099         unsigned hdr_len = vi->hdr_len;
1100         bool can_push;
1101
1102         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1103
1104         can_push = vi->any_header_sg &&
1105                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1106                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1107         /* Even if we can, don't push here yet as this would skew
1108          * csum_start offset below. */
1109         if (can_push)
1110                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1111         else
1112                 hdr = skb_vnet_hdr(skb);
1113
1114         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1115                                     virtio_is_little_endian(vi->vdev), false))
1116                 BUG();
1117
1118         if (vi->mergeable_rx_bufs)
1119                 hdr->num_buffers = 0;
1120
1121         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1122         if (can_push) {
1123                 __skb_push(skb, hdr_len);
1124                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1125                 /* Pull header back to avoid skew in tx bytes calculations. */
1126                 __skb_pull(skb, hdr_len);
1127         } else {
1128                 sg_set_buf(sq->sg, hdr, hdr_len);
1129                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
1130         }
1131         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1132 }
1133
1134 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1135 {
1136         struct virtnet_info *vi = netdev_priv(dev);
1137         int qnum = skb_get_queue_mapping(skb);
1138         struct send_queue *sq = &vi->sq[qnum];
1139         int err;
1140         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1141         bool kick = !skb->xmit_more;
1142
1143         /* Free up any pending old buffers before queueing new ones. */
1144         free_old_xmit_skbs(sq);
1145
1146         /* timestamp packet in software */
1147         skb_tx_timestamp(skb);
1148
1149         /* Try to transmit */
1150         err = xmit_skb(sq, skb);
1151
1152         /* This should not happen! */
1153         if (unlikely(err)) {
1154                 dev->stats.tx_fifo_errors++;
1155                 if (net_ratelimit())
1156                         dev_warn(&dev->dev,
1157                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1158                 dev->stats.tx_dropped++;
1159                 dev_kfree_skb_any(skb);
1160                 return NETDEV_TX_OK;
1161         }
1162
1163         /* Don't wait up for transmitted skbs to be freed. */
1164         skb_orphan(skb);
1165         nf_reset(skb);
1166
1167         /* If running out of space, stop queue to avoid getting packets that we
1168          * are then unable to transmit.
1169          * An alternative would be to force queuing layer to requeue the skb by
1170          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1171          * returned in a normal path of operation: it means that driver is not
1172          * maintaining the TX queue stop/start state properly, and causes
1173          * the stack to do a non-trivial amount of useless work.
1174          * Since most packets only take 1 or 2 ring slots, stopping the queue
1175          * early means 16 slots are typically wasted.
1176          */
1177         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1178                 netif_stop_subqueue(dev, qnum);
1179                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1180                         /* More just got used, free them then recheck. */
1181                         free_old_xmit_skbs(sq);
1182                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1183                                 netif_start_subqueue(dev, qnum);
1184                                 virtqueue_disable_cb(sq->vq);
1185                         }
1186                 }
1187         }
1188
1189         if (kick || netif_xmit_stopped(txq))
1190                 virtqueue_kick(sq->vq);
1191
1192         return NETDEV_TX_OK;
1193 }
1194
1195 /*
1196  * Send command via the control virtqueue and check status.  Commands
1197  * supported by the hypervisor, as indicated by feature bits, should
1198  * never fail unless improperly formatted.
1199  */
1200 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1201                                  struct scatterlist *out)
1202 {
1203         struct scatterlist *sgs[4], hdr, stat;
1204         unsigned out_num = 0, tmp;
1205
1206         /* Caller should know better */
1207         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1208
1209         vi->ctrl_status = ~0;
1210         vi->ctrl_hdr.class = class;
1211         vi->ctrl_hdr.cmd = cmd;
1212         /* Add header */
1213         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1214         sgs[out_num++] = &hdr;
1215
1216         if (out)
1217                 sgs[out_num++] = out;
1218
1219         /* Add return status. */
1220         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1221         sgs[out_num] = &stat;
1222
1223         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1224         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1225
1226         if (unlikely(!virtqueue_kick(vi->cvq)))
1227                 return vi->ctrl_status == VIRTIO_NET_OK;
1228
1229         /* Spin for a response, the kick causes an ioport write, trapping
1230          * into the hypervisor, so the request should be handled immediately.
1231          */
1232         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1233                !virtqueue_is_broken(vi->cvq))
1234                 cpu_relax();
1235
1236         return vi->ctrl_status == VIRTIO_NET_OK;
1237 }
1238
1239 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1240 {
1241         struct virtnet_info *vi = netdev_priv(dev);
1242         struct virtio_device *vdev = vi->vdev;
1243         int ret;
1244         struct sockaddr *addr;
1245         struct scatterlist sg;
1246
1247         addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1248         if (!addr)
1249                 return -ENOMEM;
1250         memcpy(addr, p, sizeof(*addr));
1251
1252         ret = eth_prepare_mac_addr_change(dev, addr);
1253         if (ret)
1254                 goto out;
1255
1256         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1257                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1258                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1259                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1260                         dev_warn(&vdev->dev,
1261                                  "Failed to set mac address by vq command.\n");
1262                         ret = -EINVAL;
1263                         goto out;
1264                 }
1265         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1266                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1267                 unsigned int i;
1268
1269                 /* Naturally, this has an atomicity problem. */
1270                 for (i = 0; i < dev->addr_len; i++)
1271                         virtio_cwrite8(vdev,
1272                                        offsetof(struct virtio_net_config, mac) +
1273                                        i, addr->sa_data[i]);
1274         }
1275
1276         eth_commit_mac_addr_change(dev, p);
1277         ret = 0;
1278
1279 out:
1280         kfree(addr);
1281         return ret;
1282 }
1283
1284 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1285                                                struct rtnl_link_stats64 *tot)
1286 {
1287         struct virtnet_info *vi = netdev_priv(dev);
1288         int cpu;
1289         unsigned int start;
1290
1291         for_each_possible_cpu(cpu) {
1292                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1293                 u64 tpackets, tbytes, rpackets, rbytes;
1294
1295                 do {
1296                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1297                         tpackets = stats->tx_packets;
1298                         tbytes   = stats->tx_bytes;
1299                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1300
1301                 do {
1302                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1303                         rpackets = stats->rx_packets;
1304                         rbytes   = stats->rx_bytes;
1305                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1306
1307                 tot->rx_packets += rpackets;
1308                 tot->tx_packets += tpackets;
1309                 tot->rx_bytes   += rbytes;
1310                 tot->tx_bytes   += tbytes;
1311         }
1312
1313         tot->tx_dropped = dev->stats.tx_dropped;
1314         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1315         tot->rx_dropped = dev->stats.rx_dropped;
1316         tot->rx_length_errors = dev->stats.rx_length_errors;
1317         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1318
1319         return tot;
1320 }
1321
1322 #ifdef CONFIG_NET_POLL_CONTROLLER
1323 static void virtnet_netpoll(struct net_device *dev)
1324 {
1325         struct virtnet_info *vi = netdev_priv(dev);
1326         int i;
1327
1328         for (i = 0; i < vi->curr_queue_pairs; i++)
1329                 napi_schedule(&vi->rq[i].napi);
1330 }
1331 #endif
1332
1333 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1334 {
1335         rtnl_lock();
1336         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1337                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1338                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1339         rtnl_unlock();
1340 }
1341
1342 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1343 {
1344         struct scatterlist sg;
1345         struct net_device *dev = vi->dev;
1346
1347         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1348                 return 0;
1349
1350         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1351         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1352
1353         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1354                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1355                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1356                          queue_pairs);
1357                 return -EINVAL;
1358         } else {
1359                 vi->curr_queue_pairs = queue_pairs;
1360                 /* virtnet_open() will refill when device is going to up. */
1361                 if (dev->flags & IFF_UP)
1362                         schedule_delayed_work(&vi->refill, 0);
1363         }
1364
1365         return 0;
1366 }
1367
1368 static int virtnet_close(struct net_device *dev)
1369 {
1370         struct virtnet_info *vi = netdev_priv(dev);
1371         int i;
1372
1373         /* Make sure refill_work doesn't re-enable napi! */
1374         cancel_delayed_work_sync(&vi->refill);
1375
1376         for (i = 0; i < vi->max_queue_pairs; i++)
1377                 napi_disable(&vi->rq[i].napi);
1378
1379         return 0;
1380 }
1381
1382 static void virtnet_set_rx_mode(struct net_device *dev)
1383 {
1384         struct virtnet_info *vi = netdev_priv(dev);
1385         struct scatterlist sg[2];
1386         struct virtio_net_ctrl_mac *mac_data;
1387         struct netdev_hw_addr *ha;
1388         int uc_count;
1389         int mc_count;
1390         void *buf;
1391         int i;
1392
1393         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1394         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1395                 return;
1396
1397         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1398         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1399
1400         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1401
1402         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1403                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1404                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1405                          vi->ctrl_promisc ? "en" : "dis");
1406
1407         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1408
1409         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1410                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1411                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1412                          vi->ctrl_allmulti ? "en" : "dis");
1413
1414         uc_count = netdev_uc_count(dev);
1415         mc_count = netdev_mc_count(dev);
1416         /* MAC filter - use one buffer for both lists */
1417         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1418                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1419         mac_data = buf;
1420         if (!buf)
1421                 return;
1422
1423         sg_init_table(sg, 2);
1424
1425         /* Store the unicast list and count in the front of the buffer */
1426         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1427         i = 0;
1428         netdev_for_each_uc_addr(ha, dev)
1429                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1430
1431         sg_set_buf(&sg[0], mac_data,
1432                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1433
1434         /* multicast list and count fill the end */
1435         mac_data = (void *)&mac_data->macs[uc_count][0];
1436
1437         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1438         i = 0;
1439         netdev_for_each_mc_addr(ha, dev)
1440                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1441
1442         sg_set_buf(&sg[1], mac_data,
1443                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1444
1445         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1446                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1447                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1448
1449         kfree(buf);
1450 }
1451
1452 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1453                                    __be16 proto, u16 vid)
1454 {
1455         struct virtnet_info *vi = netdev_priv(dev);
1456         struct scatterlist sg;
1457
1458         vi->ctrl_vid = vid;
1459         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1460
1461         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1462                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1463                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1464         return 0;
1465 }
1466
1467 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1468                                     __be16 proto, u16 vid)
1469 {
1470         struct virtnet_info *vi = netdev_priv(dev);
1471         struct scatterlist sg;
1472
1473         vi->ctrl_vid = vid;
1474         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1475
1476         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1477                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1478                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1479         return 0;
1480 }
1481
1482 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1483 {
1484         int i;
1485
1486         if (vi->affinity_hint_set) {
1487                 for (i = 0; i < vi->max_queue_pairs; i++) {
1488                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1489                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1490                 }
1491
1492                 vi->affinity_hint_set = false;
1493         }
1494 }
1495
1496 static void virtnet_set_affinity(struct virtnet_info *vi)
1497 {
1498         int i;
1499         int cpu;
1500
1501         /* In multiqueue mode, when the number of cpu is equal to the number of
1502          * queue pairs, we let the queue pairs to be private to one cpu by
1503          * setting the affinity hint to eliminate the contention.
1504          */
1505         if (vi->curr_queue_pairs == 1 ||
1506             vi->max_queue_pairs != num_online_cpus()) {
1507                 virtnet_clean_affinity(vi, -1);
1508                 return;
1509         }
1510
1511         i = 0;
1512         for_each_online_cpu(cpu) {
1513                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1514                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1515                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1516                 i++;
1517         }
1518
1519         vi->affinity_hint_set = true;
1520 }
1521
1522 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1523 {
1524         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1525                                                    node);
1526         virtnet_set_affinity(vi);
1527         return 0;
1528 }
1529
1530 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1531 {
1532         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1533                                                    node_dead);
1534         virtnet_set_affinity(vi);
1535         return 0;
1536 }
1537
1538 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1539 {
1540         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1541                                                    node);
1542
1543         virtnet_clean_affinity(vi, cpu);
1544         return 0;
1545 }
1546
1547 static enum cpuhp_state virtionet_online;
1548
1549 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1550 {
1551         int ret;
1552
1553         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1554         if (ret)
1555                 return ret;
1556         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1557                                                &vi->node_dead);
1558         if (!ret)
1559                 return ret;
1560         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1561         return ret;
1562 }
1563
1564 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1565 {
1566         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1567         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1568                                             &vi->node_dead);
1569 }
1570
1571 static void virtnet_get_ringparam(struct net_device *dev,
1572                                 struct ethtool_ringparam *ring)
1573 {
1574         struct virtnet_info *vi = netdev_priv(dev);
1575
1576         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1577         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1578         ring->rx_pending = ring->rx_max_pending;
1579         ring->tx_pending = ring->tx_max_pending;
1580 }
1581
1582
1583 static void virtnet_get_drvinfo(struct net_device *dev,
1584                                 struct ethtool_drvinfo *info)
1585 {
1586         struct virtnet_info *vi = netdev_priv(dev);
1587         struct virtio_device *vdev = vi->vdev;
1588
1589         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1590         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1591         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1592
1593 }
1594
1595 /* TODO: Eliminate OOO packets during switching */
1596 static int virtnet_set_channels(struct net_device *dev,
1597                                 struct ethtool_channels *channels)
1598 {
1599         struct virtnet_info *vi = netdev_priv(dev);
1600         u16 queue_pairs = channels->combined_count;
1601         int err;
1602
1603         /* We don't support separate rx/tx channels.
1604          * We don't allow setting 'other' channels.
1605          */
1606         if (channels->rx_count || channels->tx_count || channels->other_count)
1607                 return -EINVAL;
1608
1609         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1610                 return -EINVAL;
1611
1612         /* For now we don't support modifying channels while XDP is loaded
1613          * also when XDP is loaded all RX queues have XDP programs so we only
1614          * need to check a single RX queue.
1615          */
1616         if (vi->rq[0].xdp_prog)
1617                 return -EINVAL;
1618
1619         get_online_cpus();
1620         err = virtnet_set_queues(vi, queue_pairs);
1621         if (!err) {
1622                 netif_set_real_num_tx_queues(dev, queue_pairs);
1623                 netif_set_real_num_rx_queues(dev, queue_pairs);
1624
1625                 virtnet_set_affinity(vi);
1626         }
1627         put_online_cpus();
1628
1629         return err;
1630 }
1631
1632 static void virtnet_get_channels(struct net_device *dev,
1633                                  struct ethtool_channels *channels)
1634 {
1635         struct virtnet_info *vi = netdev_priv(dev);
1636
1637         channels->combined_count = vi->curr_queue_pairs;
1638         channels->max_combined = vi->max_queue_pairs;
1639         channels->max_other = 0;
1640         channels->rx_count = 0;
1641         channels->tx_count = 0;
1642         channels->other_count = 0;
1643 }
1644
1645 /* Check if the user is trying to change anything besides speed/duplex */
1646 static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1647 {
1648         struct ethtool_cmd diff1 = *cmd;
1649         struct ethtool_cmd diff2 = {};
1650
1651         /* cmd is always set so we need to clear it, validate the port type
1652          * and also without autonegotiation we can ignore advertising
1653          */
1654         ethtool_cmd_speed_set(&diff1, 0);
1655         diff2.port = PORT_OTHER;
1656         diff1.advertising = 0;
1657         diff1.duplex = 0;
1658         diff1.cmd = 0;
1659
1660         return !memcmp(&diff1, &diff2, sizeof(diff1));
1661 }
1662
1663 static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1664 {
1665         struct virtnet_info *vi = netdev_priv(dev);
1666         u32 speed;
1667
1668         speed = ethtool_cmd_speed(cmd);
1669         /* don't allow custom speed and duplex */
1670         if (!ethtool_validate_speed(speed) ||
1671             !ethtool_validate_duplex(cmd->duplex) ||
1672             !virtnet_validate_ethtool_cmd(cmd))
1673                 return -EINVAL;
1674         vi->speed = speed;
1675         vi->duplex = cmd->duplex;
1676
1677         return 0;
1678 }
1679
1680 static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1681 {
1682         struct virtnet_info *vi = netdev_priv(dev);
1683
1684         ethtool_cmd_speed_set(cmd, vi->speed);
1685         cmd->duplex = vi->duplex;
1686         cmd->port = PORT_OTHER;
1687
1688         return 0;
1689 }
1690
1691 static void virtnet_init_settings(struct net_device *dev)
1692 {
1693         struct virtnet_info *vi = netdev_priv(dev);
1694
1695         vi->speed = SPEED_UNKNOWN;
1696         vi->duplex = DUPLEX_UNKNOWN;
1697 }
1698
1699 static const struct ethtool_ops virtnet_ethtool_ops = {
1700         .get_drvinfo = virtnet_get_drvinfo,
1701         .get_link = ethtool_op_get_link,
1702         .get_ringparam = virtnet_get_ringparam,
1703         .set_channels = virtnet_set_channels,
1704         .get_channels = virtnet_get_channels,
1705         .get_ts_info = ethtool_op_get_ts_info,
1706         .get_settings = virtnet_get_settings,
1707         .set_settings = virtnet_set_settings,
1708 };
1709
1710 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog)
1711 {
1712         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1713         struct virtnet_info *vi = netdev_priv(dev);
1714         struct bpf_prog *old_prog;
1715         u16 xdp_qp = 0, curr_qp;
1716         int i, err;
1717
1718         if (prog && prog->xdp_adjust_head) {
1719                 netdev_warn(dev, "Does not support bpf_xdp_adjust_head()\n");
1720                 return -EOPNOTSUPP;
1721         }
1722
1723         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1724             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1725             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
1726             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO)) {
1727                 netdev_warn(dev, "can't set XDP while host is implementing LRO, disable LRO first\n");
1728                 return -EOPNOTSUPP;
1729         }
1730
1731         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1732                 netdev_warn(dev, "XDP expects header/data in single page, any_header_sg required\n");
1733                 return -EINVAL;
1734         }
1735
1736         if (dev->mtu > max_sz) {
1737                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1738                 return -EINVAL;
1739         }
1740
1741         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1742         if (prog)
1743                 xdp_qp = nr_cpu_ids;
1744
1745         /* XDP requires extra queues for XDP_TX */
1746         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1747                 netdev_warn(dev, "request %i queues but max is %i\n",
1748                             curr_qp + xdp_qp, vi->max_queue_pairs);
1749                 return -ENOMEM;
1750         }
1751
1752         err = virtnet_set_queues(vi, curr_qp + xdp_qp);
1753         if (err) {
1754                 dev_warn(&dev->dev, "XDP Device queue allocation failure.\n");
1755                 return err;
1756         }
1757
1758         if (prog) {
1759                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1760                 if (IS_ERR(prog)) {
1761                         virtnet_set_queues(vi, curr_qp);
1762                         return PTR_ERR(prog);
1763                 }
1764         }
1765
1766         vi->xdp_queue_pairs = xdp_qp;
1767         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
1768
1769         for (i = 0; i < vi->max_queue_pairs; i++) {
1770                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1771                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
1772                 if (old_prog)
1773                         bpf_prog_put(old_prog);
1774         }
1775
1776         return 0;
1777 }
1778
1779 static bool virtnet_xdp_query(struct net_device *dev)
1780 {
1781         struct virtnet_info *vi = netdev_priv(dev);
1782         int i;
1783
1784         for (i = 0; i < vi->max_queue_pairs; i++) {
1785                 if (vi->rq[i].xdp_prog)
1786                         return true;
1787         }
1788         return false;
1789 }
1790
1791 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
1792 {
1793         switch (xdp->command) {
1794         case XDP_SETUP_PROG:
1795                 return virtnet_xdp_set(dev, xdp->prog);
1796         case XDP_QUERY_PROG:
1797                 xdp->prog_attached = virtnet_xdp_query(dev);
1798                 return 0;
1799         default:
1800                 return -EINVAL;
1801         }
1802 }
1803
1804 static const struct net_device_ops virtnet_netdev = {
1805         .ndo_open            = virtnet_open,
1806         .ndo_stop            = virtnet_close,
1807         .ndo_start_xmit      = start_xmit,
1808         .ndo_validate_addr   = eth_validate_addr,
1809         .ndo_set_mac_address = virtnet_set_mac_address,
1810         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1811         .ndo_get_stats64     = virtnet_stats,
1812         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1813         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1814 #ifdef CONFIG_NET_POLL_CONTROLLER
1815         .ndo_poll_controller = virtnet_netpoll,
1816 #endif
1817 #ifdef CONFIG_NET_RX_BUSY_POLL
1818         .ndo_busy_poll          = virtnet_busy_poll,
1819 #endif
1820         .ndo_xdp                = virtnet_xdp,
1821 };
1822
1823 static void virtnet_config_changed_work(struct work_struct *work)
1824 {
1825         struct virtnet_info *vi =
1826                 container_of(work, struct virtnet_info, config_work);
1827         u16 v;
1828
1829         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1830                                  struct virtio_net_config, status, &v) < 0)
1831                 return;
1832
1833         if (v & VIRTIO_NET_S_ANNOUNCE) {
1834                 netdev_notify_peers(vi->dev);
1835                 virtnet_ack_link_announce(vi);
1836         }
1837
1838         /* Ignore unknown (future) status bits */
1839         v &= VIRTIO_NET_S_LINK_UP;
1840
1841         if (vi->status == v)
1842                 return;
1843
1844         vi->status = v;
1845
1846         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1847                 netif_carrier_on(vi->dev);
1848                 netif_tx_wake_all_queues(vi->dev);
1849         } else {
1850                 netif_carrier_off(vi->dev);
1851                 netif_tx_stop_all_queues(vi->dev);
1852         }
1853 }
1854
1855 static void virtnet_config_changed(struct virtio_device *vdev)
1856 {
1857         struct virtnet_info *vi = vdev->priv;
1858
1859         schedule_work(&vi->config_work);
1860 }
1861
1862 static void virtnet_free_queues(struct virtnet_info *vi)
1863 {
1864         int i;
1865
1866         for (i = 0; i < vi->max_queue_pairs; i++) {
1867                 napi_hash_del(&vi->rq[i].napi);
1868                 netif_napi_del(&vi->rq[i].napi);
1869         }
1870
1871         /* We called napi_hash_del() before netif_napi_del(),
1872          * we need to respect an RCU grace period before freeing vi->rq
1873          */
1874         synchronize_net();
1875
1876         kfree(vi->rq);
1877         kfree(vi->sq);
1878 }
1879
1880 static void free_receive_bufs(struct virtnet_info *vi)
1881 {
1882         struct bpf_prog *old_prog;
1883         int i;
1884
1885         rtnl_lock();
1886         for (i = 0; i < vi->max_queue_pairs; i++) {
1887                 while (vi->rq[i].pages)
1888                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1889
1890                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1891                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
1892                 if (old_prog)
1893                         bpf_prog_put(old_prog);
1894         }
1895         rtnl_unlock();
1896 }
1897
1898 static void free_receive_page_frags(struct virtnet_info *vi)
1899 {
1900         int i;
1901         for (i = 0; i < vi->max_queue_pairs; i++)
1902                 if (vi->rq[i].alloc_frag.page)
1903                         put_page(vi->rq[i].alloc_frag.page);
1904 }
1905
1906 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1907 {
1908         /* For small receive mode always use kfree_skb variants */
1909         if (!vi->mergeable_rx_bufs)
1910                 return false;
1911
1912         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1913                 return false;
1914         else if (q < vi->curr_queue_pairs)
1915                 return true;
1916         else
1917                 return false;
1918 }
1919
1920 static void free_unused_bufs(struct virtnet_info *vi)
1921 {
1922         void *buf;
1923         int i;
1924
1925         for (i = 0; i < vi->max_queue_pairs; i++) {
1926                 struct virtqueue *vq = vi->sq[i].vq;
1927                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1928                         if (!is_xdp_raw_buffer_queue(vi, i))
1929                                 dev_kfree_skb(buf);
1930                         else
1931                                 put_page(virt_to_head_page(buf));
1932                 }
1933         }
1934
1935         for (i = 0; i < vi->max_queue_pairs; i++) {
1936                 struct virtqueue *vq = vi->rq[i].vq;
1937
1938                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1939                         if (vi->mergeable_rx_bufs) {
1940                                 unsigned long ctx = (unsigned long)buf;
1941                                 void *base = mergeable_ctx_to_buf_address(ctx);
1942                                 put_page(virt_to_head_page(base));
1943                         } else if (vi->big_packets) {
1944                                 give_pages(&vi->rq[i], buf);
1945                         } else {
1946                                 dev_kfree_skb(buf);
1947                         }
1948                 }
1949         }
1950 }
1951
1952 static void virtnet_del_vqs(struct virtnet_info *vi)
1953 {
1954         struct virtio_device *vdev = vi->vdev;
1955
1956         virtnet_clean_affinity(vi, -1);
1957
1958         vdev->config->del_vqs(vdev);
1959
1960         virtnet_free_queues(vi);
1961 }
1962
1963 static int virtnet_find_vqs(struct virtnet_info *vi)
1964 {
1965         vq_callback_t **callbacks;
1966         struct virtqueue **vqs;
1967         int ret = -ENOMEM;
1968         int i, total_vqs;
1969         const char **names;
1970
1971         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1972          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1973          * possible control vq.
1974          */
1975         total_vqs = vi->max_queue_pairs * 2 +
1976                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1977
1978         /* Allocate space for find_vqs parameters */
1979         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1980         if (!vqs)
1981                 goto err_vq;
1982         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1983         if (!callbacks)
1984                 goto err_callback;
1985         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1986         if (!names)
1987                 goto err_names;
1988
1989         /* Parameters for control virtqueue, if any */
1990         if (vi->has_cvq) {
1991                 callbacks[total_vqs - 1] = NULL;
1992                 names[total_vqs - 1] = "control";
1993         }
1994
1995         /* Allocate/initialize parameters for send/receive virtqueues */
1996         for (i = 0; i < vi->max_queue_pairs; i++) {
1997                 callbacks[rxq2vq(i)] = skb_recv_done;
1998                 callbacks[txq2vq(i)] = skb_xmit_done;
1999                 sprintf(vi->rq[i].name, "input.%d", i);
2000                 sprintf(vi->sq[i].name, "output.%d", i);
2001                 names[rxq2vq(i)] = vi->rq[i].name;
2002                 names[txq2vq(i)] = vi->sq[i].name;
2003         }
2004
2005         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2006                                          names);
2007         if (ret)
2008                 goto err_find;
2009
2010         if (vi->has_cvq) {
2011                 vi->cvq = vqs[total_vqs - 1];
2012                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2013                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2014         }
2015
2016         for (i = 0; i < vi->max_queue_pairs; i++) {
2017                 vi->rq[i].vq = vqs[rxq2vq(i)];
2018                 vi->sq[i].vq = vqs[txq2vq(i)];
2019         }
2020
2021         kfree(names);
2022         kfree(callbacks);
2023         kfree(vqs);
2024
2025         return 0;
2026
2027 err_find:
2028         kfree(names);
2029 err_names:
2030         kfree(callbacks);
2031 err_callback:
2032         kfree(vqs);
2033 err_vq:
2034         return ret;
2035 }
2036
2037 static int virtnet_alloc_queues(struct virtnet_info *vi)
2038 {
2039         int i;
2040
2041         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2042         if (!vi->sq)
2043                 goto err_sq;
2044         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2045         if (!vi->rq)
2046                 goto err_rq;
2047
2048         INIT_DELAYED_WORK(&vi->refill, refill_work);
2049         for (i = 0; i < vi->max_queue_pairs; i++) {
2050                 vi->rq[i].pages = NULL;
2051                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2052                                napi_weight);
2053
2054                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2055                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2056                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2057         }
2058
2059         return 0;
2060
2061 err_rq:
2062         kfree(vi->sq);
2063 err_sq:
2064         return -ENOMEM;
2065 }
2066
2067 static int init_vqs(struct virtnet_info *vi)
2068 {
2069         int ret;
2070
2071         /* Allocate send & receive queues */
2072         ret = virtnet_alloc_queues(vi);
2073         if (ret)
2074                 goto err;
2075
2076         ret = virtnet_find_vqs(vi);
2077         if (ret)
2078                 goto err_free;
2079
2080         get_online_cpus();
2081         virtnet_set_affinity(vi);
2082         put_online_cpus();
2083
2084         return 0;
2085
2086 err_free:
2087         virtnet_free_queues(vi);
2088 err:
2089         return ret;
2090 }
2091
2092 #ifdef CONFIG_SYSFS
2093 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2094                 struct rx_queue_attribute *attribute, char *buf)
2095 {
2096         struct virtnet_info *vi = netdev_priv(queue->dev);
2097         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2098         struct ewma_pkt_len *avg;
2099
2100         BUG_ON(queue_index >= vi->max_queue_pairs);
2101         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2102         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
2103 }
2104
2105 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2106         __ATTR_RO(mergeable_rx_buffer_size);
2107
2108 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2109         &mergeable_rx_buffer_size_attribute.attr,
2110         NULL
2111 };
2112
2113 static const struct attribute_group virtio_net_mrg_rx_group = {
2114         .name = "virtio_net",
2115         .attrs = virtio_net_mrg_rx_attrs
2116 };
2117 #endif
2118
2119 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2120                                     unsigned int fbit,
2121                                     const char *fname, const char *dname)
2122 {
2123         if (!virtio_has_feature(vdev, fbit))
2124                 return false;
2125
2126         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2127                 fname, dname);
2128
2129         return true;
2130 }
2131
2132 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2133         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2134
2135 static bool virtnet_validate_features(struct virtio_device *vdev)
2136 {
2137         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2138             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2139                              "VIRTIO_NET_F_CTRL_VQ") ||
2140              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2141                              "VIRTIO_NET_F_CTRL_VQ") ||
2142              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2143                              "VIRTIO_NET_F_CTRL_VQ") ||
2144              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2145              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2146                              "VIRTIO_NET_F_CTRL_VQ"))) {
2147                 return false;
2148         }
2149
2150         return true;
2151 }
2152
2153 #define MIN_MTU ETH_MIN_MTU
2154 #define MAX_MTU ETH_MAX_MTU
2155
2156 static int virtnet_probe(struct virtio_device *vdev)
2157 {
2158         int i, err;
2159         struct net_device *dev;
2160         struct virtnet_info *vi;
2161         u16 max_queue_pairs;
2162         int mtu;
2163
2164         if (!vdev->config->get) {
2165                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2166                         __func__);
2167                 return -EINVAL;
2168         }
2169
2170         if (!virtnet_validate_features(vdev))
2171                 return -EINVAL;
2172
2173         /* Find if host supports multiqueue virtio_net device */
2174         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2175                                    struct virtio_net_config,
2176                                    max_virtqueue_pairs, &max_queue_pairs);
2177
2178         /* We need at least 2 queue's */
2179         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2180             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2181             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2182                 max_queue_pairs = 1;
2183
2184         /* Allocate ourselves a network device with room for our info */
2185         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2186         if (!dev)
2187                 return -ENOMEM;
2188
2189         /* Set up network device as normal. */
2190         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2191         dev->netdev_ops = &virtnet_netdev;
2192         dev->features = NETIF_F_HIGHDMA;
2193
2194         dev->ethtool_ops = &virtnet_ethtool_ops;
2195         SET_NETDEV_DEV(dev, &vdev->dev);
2196
2197         /* Do we support "hardware" checksums? */
2198         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2199                 /* This opens up the world of extra features. */
2200                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2201                 if (csum)
2202                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2203
2204                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2205                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
2206                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2207                 }
2208                 /* Individual feature bits: what can host handle? */
2209                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2210                         dev->hw_features |= NETIF_F_TSO;
2211                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2212                         dev->hw_features |= NETIF_F_TSO6;
2213                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2214                         dev->hw_features |= NETIF_F_TSO_ECN;
2215                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
2216                         dev->hw_features |= NETIF_F_UFO;
2217
2218                 dev->features |= NETIF_F_GSO_ROBUST;
2219
2220                 if (gso)
2221                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
2222                 /* (!csum && gso) case will be fixed by register_netdev() */
2223         }
2224         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2225                 dev->features |= NETIF_F_RXCSUM;
2226
2227         dev->vlan_features = dev->features;
2228
2229         /* MTU range: 68 - 65535 */
2230         dev->min_mtu = MIN_MTU;
2231         dev->max_mtu = MAX_MTU;
2232
2233         /* Configuration may specify what MAC to use.  Otherwise random. */
2234         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2235                 virtio_cread_bytes(vdev,
2236                                    offsetof(struct virtio_net_config, mac),
2237                                    dev->dev_addr, dev->addr_len);
2238         else
2239                 eth_hw_addr_random(dev);
2240
2241         /* Set up our device-specific information */
2242         vi = netdev_priv(dev);
2243         vi->dev = dev;
2244         vi->vdev = vdev;
2245         vdev->priv = vi;
2246         vi->stats = alloc_percpu(struct virtnet_stats);
2247         err = -ENOMEM;
2248         if (vi->stats == NULL)
2249                 goto free;
2250
2251         for_each_possible_cpu(i) {
2252                 struct virtnet_stats *virtnet_stats;
2253                 virtnet_stats = per_cpu_ptr(vi->stats, i);
2254                 u64_stats_init(&virtnet_stats->tx_syncp);
2255                 u64_stats_init(&virtnet_stats->rx_syncp);
2256         }
2257
2258         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2259
2260         /* If we can receive ANY GSO packets, we must allocate large ones. */
2261         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2262             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2263             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2264             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2265                 vi->big_packets = true;
2266
2267         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2268                 vi->mergeable_rx_bufs = true;
2269
2270         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2271             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2272                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2273         else
2274                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2275
2276         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2277             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2278                 vi->any_header_sg = true;
2279
2280         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2281                 vi->has_cvq = true;
2282
2283         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2284                 mtu = virtio_cread16(vdev,
2285                                      offsetof(struct virtio_net_config,
2286                                               mtu));
2287                 if (mtu < dev->min_mtu) {
2288                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2289                 } else {
2290                         dev->mtu = mtu;
2291                         dev->max_mtu = mtu;
2292                 }
2293         }
2294
2295         if (vi->any_header_sg)
2296                 dev->needed_headroom = vi->hdr_len;
2297
2298         /* Enable multiqueue by default */
2299         if (num_online_cpus() >= max_queue_pairs)
2300                 vi->curr_queue_pairs = max_queue_pairs;
2301         else
2302                 vi->curr_queue_pairs = num_online_cpus();
2303         vi->max_queue_pairs = max_queue_pairs;
2304
2305         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2306         err = init_vqs(vi);
2307         if (err)
2308                 goto free_stats;
2309
2310 #ifdef CONFIG_SYSFS
2311         if (vi->mergeable_rx_bufs)
2312                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2313 #endif
2314         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2315         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2316
2317         virtnet_init_settings(dev);
2318
2319         err = register_netdev(dev);
2320         if (err) {
2321                 pr_debug("virtio_net: registering device failed\n");
2322                 goto free_vqs;
2323         }
2324
2325         virtio_device_ready(vdev);
2326
2327         err = virtnet_cpu_notif_add(vi);
2328         if (err) {
2329                 pr_debug("virtio_net: registering cpu notifier failed\n");
2330                 goto free_unregister_netdev;
2331         }
2332
2333         rtnl_lock();
2334         virtnet_set_queues(vi, vi->curr_queue_pairs);
2335         rtnl_unlock();
2336
2337         /* Assume link up if device can't report link status,
2338            otherwise get link status from config. */
2339         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2340                 netif_carrier_off(dev);
2341                 schedule_work(&vi->config_work);
2342         } else {
2343                 vi->status = VIRTIO_NET_S_LINK_UP;
2344                 netif_carrier_on(dev);
2345         }
2346
2347         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2348                  dev->name, max_queue_pairs);
2349
2350         return 0;
2351
2352 free_unregister_netdev:
2353         vi->vdev->config->reset(vdev);
2354
2355         unregister_netdev(dev);
2356 free_vqs:
2357         cancel_delayed_work_sync(&vi->refill);
2358         free_receive_page_frags(vi);
2359         virtnet_del_vqs(vi);
2360 free_stats:
2361         free_percpu(vi->stats);
2362 free:
2363         free_netdev(dev);
2364         return err;
2365 }
2366
2367 static void remove_vq_common(struct virtnet_info *vi)
2368 {
2369         vi->vdev->config->reset(vi->vdev);
2370
2371         /* Free unused buffers in both send and recv, if any. */
2372         free_unused_bufs(vi);
2373
2374         free_receive_bufs(vi);
2375
2376         free_receive_page_frags(vi);
2377
2378         virtnet_del_vqs(vi);
2379 }
2380
2381 static void virtnet_remove(struct virtio_device *vdev)
2382 {
2383         struct virtnet_info *vi = vdev->priv;
2384
2385         virtnet_cpu_notif_remove(vi);
2386
2387         /* Make sure no work handler is accessing the device. */
2388         flush_work(&vi->config_work);
2389
2390         unregister_netdev(vi->dev);
2391
2392         remove_vq_common(vi);
2393
2394         free_percpu(vi->stats);
2395         free_netdev(vi->dev);
2396 }
2397
2398 #ifdef CONFIG_PM_SLEEP
2399 static int virtnet_freeze(struct virtio_device *vdev)
2400 {
2401         struct virtnet_info *vi = vdev->priv;
2402         int i;
2403
2404         virtnet_cpu_notif_remove(vi);
2405
2406         /* Make sure no work handler is accessing the device */
2407         flush_work(&vi->config_work);
2408
2409         netif_device_detach(vi->dev);
2410         cancel_delayed_work_sync(&vi->refill);
2411
2412         if (netif_running(vi->dev)) {
2413                 for (i = 0; i < vi->max_queue_pairs; i++)
2414                         napi_disable(&vi->rq[i].napi);
2415         }
2416
2417         remove_vq_common(vi);
2418
2419         return 0;
2420 }
2421
2422 static int virtnet_restore(struct virtio_device *vdev)
2423 {
2424         struct virtnet_info *vi = vdev->priv;
2425         int err, i;
2426
2427         err = init_vqs(vi);
2428         if (err)
2429                 return err;
2430
2431         virtio_device_ready(vdev);
2432
2433         if (netif_running(vi->dev)) {
2434                 for (i = 0; i < vi->curr_queue_pairs; i++)
2435                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2436                                 schedule_delayed_work(&vi->refill, 0);
2437
2438                 for (i = 0; i < vi->max_queue_pairs; i++)
2439                         virtnet_napi_enable(&vi->rq[i]);
2440         }
2441
2442         netif_device_attach(vi->dev);
2443
2444         rtnl_lock();
2445         virtnet_set_queues(vi, vi->curr_queue_pairs);
2446         rtnl_unlock();
2447
2448         err = virtnet_cpu_notif_add(vi);
2449         if (err)
2450                 return err;
2451
2452         return 0;
2453 }
2454 #endif
2455
2456 static struct virtio_device_id id_table[] = {
2457         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2458         { 0 },
2459 };
2460
2461 #define VIRTNET_FEATURES \
2462         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2463         VIRTIO_NET_F_MAC, \
2464         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2465         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2466         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2467         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2468         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2469         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2470         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2471         VIRTIO_NET_F_MTU
2472
2473 static unsigned int features[] = {
2474         VIRTNET_FEATURES,
2475 };
2476
2477 static unsigned int features_legacy[] = {
2478         VIRTNET_FEATURES,
2479         VIRTIO_NET_F_GSO,
2480         VIRTIO_F_ANY_LAYOUT,
2481 };
2482
2483 static struct virtio_driver virtio_net_driver = {
2484         .feature_table = features,
2485         .feature_table_size = ARRAY_SIZE(features),
2486         .feature_table_legacy = features_legacy,
2487         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2488         .driver.name =  KBUILD_MODNAME,
2489         .driver.owner = THIS_MODULE,
2490         .id_table =     id_table,
2491         .probe =        virtnet_probe,
2492         .remove =       virtnet_remove,
2493         .config_changed = virtnet_config_changed,
2494 #ifdef CONFIG_PM_SLEEP
2495         .freeze =       virtnet_freeze,
2496         .restore =      virtnet_restore,
2497 #endif
2498 };
2499
2500 static __init int virtio_net_driver_init(void)
2501 {
2502         int ret;
2503
2504         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2505                                       virtnet_cpu_online,
2506                                       virtnet_cpu_down_prep);
2507         if (ret < 0)
2508                 goto out;
2509         virtionet_online = ret;
2510         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2511                                       NULL, virtnet_cpu_dead);
2512         if (ret)
2513                 goto err_dead;
2514
2515         ret = register_virtio_driver(&virtio_net_driver);
2516         if (ret)
2517                 goto err_virtio;
2518         return 0;
2519 err_virtio:
2520         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2521 err_dead:
2522         cpuhp_remove_multi_state(virtionet_online);
2523 out:
2524         return ret;
2525 }
2526 module_init(virtio_net_driver_init);
2527
2528 static __exit void virtio_net_driver_exit(void)
2529 {
2530         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2531         cpuhp_remove_multi_state(virtionet_online);
2532         unregister_virtio_driver(&virtio_net_driver);
2533 }
2534 module_exit(virtio_net_driver_exit);
2535
2536 MODULE_DEVICE_TABLE(virtio, id_table);
2537 MODULE_DESCRIPTION("Virtio network driver");
2538 MODULE_LICENSE("GPL");