drivers/net/virtio_net.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* A network driver using virtio.
   3  *
   4  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   5  */
   6 //#define DEBUG
   7 #include <linux/netdevice.h>
   8 #include <linux/etherdevice.h>
   9 #include <linux/ethtool.h>
  10 #include <linux/module.h>
  11 #include <linux/virtio.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/bpf.h>
  14 #include <linux/bpf_trace.h>
  15 #include <linux/scatterlist.h>
  16 #include <linux/if_vlan.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/average.h>
  20 #include <linux/filter.h>
  21 #include <linux/kernel.h>
  22 #include <net/route.h>
  23 #include <net/xdp.h>
  24 #include <net/net_failover.h>
  25
  26 static int napi_weight = NAPI_POLL_WEIGHT;
  27 module_param(napi_weight, int, 0444);
  28
  29 static bool csum = true, gso = true, napi_tx = true;
  30 module_param(csum, bool, 0444);
  31 module_param(gso, bool, 0444);
  32 module_param(napi_tx, bool, 0644);
  33
  34 /* FIXME: MTU in config. */
  35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  36 #define GOOD_COPY_LEN   128
  37
  38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  39
  40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  41 #define VIRTIO_XDP_HEADROOM 256
  42
  43 /* Separating two types of XDP xmit */
  44 #define VIRTIO_XDP_TX           BIT(0)
  45 #define VIRTIO_XDP_REDIR        BIT(1)
  46
  47 #define VIRTIO_XDP_FLAG BIT(0)
  48
  49 /* RX packet size EWMA. The average packet size is used to determine the packet
  50  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  51  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  52  * term, transient changes in packet size.
  53  */
  54 DECLARE_EWMA(pkt_len, 0, 64)
  55
  56 #define VIRTNET_DRIVER_VERSION "1.0.0"
  57
  58 static const unsigned long guest_offloads[] = {
  59         VIRTIO_NET_F_GUEST_TSO4,
  60         VIRTIO_NET_F_GUEST_TSO6,
  61         VIRTIO_NET_F_GUEST_ECN,
  62         VIRTIO_NET_F_GUEST_UFO,
  63         VIRTIO_NET_F_GUEST_CSUM
  64 };
  65
  66 struct virtnet_stat_desc {
  67         char desc[ETH_GSTRING_LEN];
  68         size_t offset;
  69 };
  70
  71 struct virtnet_sq_stats {
  72         struct u64_stats_sync syncp;
  73         u64 packets;
  74         u64 bytes;
  75         u64 xdp_tx;
  76         u64 xdp_tx_drops;
  77         u64 kicks;
  78 };
  79
  80 struct virtnet_rq_stats {
  81         struct u64_stats_sync syncp;
  82         u64 packets;
  83         u64 bytes;
  84         u64 drops;
  85         u64 xdp_packets;
  86         u64 xdp_tx;
  87         u64 xdp_redirects;
  88         u64 xdp_drops;
  89         u64 kicks;
  90 };
  91
  92 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
  93 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
  94
  95 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
  96         { "packets",            VIRTNET_SQ_STAT(packets) },
  97         { "bytes",              VIRTNET_SQ_STAT(bytes) },
  98         { "xdp_tx",             VIRTNET_SQ_STAT(xdp_tx) },
  99         { "xdp_tx_drops",       VIRTNET_SQ_STAT(xdp_tx_drops) },
 100         { "kicks",              VIRTNET_SQ_STAT(kicks) },
 101 };
 102
 103 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
 104         { "packets",            VIRTNET_RQ_STAT(packets) },
 105         { "bytes",              VIRTNET_RQ_STAT(bytes) },
 106         { "drops",              VIRTNET_RQ_STAT(drops) },
 107         { "xdp_packets",        VIRTNET_RQ_STAT(xdp_packets) },
 108         { "xdp_tx",             VIRTNET_RQ_STAT(xdp_tx) },
 109         { "xdp_redirects",      VIRTNET_RQ_STAT(xdp_redirects) },
 110         { "xdp_drops",          VIRTNET_RQ_STAT(xdp_drops) },
 111         { "kicks",              VIRTNET_RQ_STAT(kicks) },
 112 };
 113
 114 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 115 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 116
 117 /* Internal representation of a send virtqueue */
 118 struct send_queue {
 119         /* Virtqueue associated with this send _queue */
 120         struct virtqueue *vq;
 121
 122         /* TX: fragments + linear part + virtio header */
 123         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 124
 125         /* Name of the send queue: output.$index */
 126         char name[40];
 127
 128         struct virtnet_sq_stats stats;
 129
 130         struct napi_struct napi;
 131 };
 132
 133 /* Internal representation of a receive virtqueue */
 134 struct receive_queue {
 135         /* Virtqueue associated with this receive_queue */
 136         struct virtqueue *vq;
 137
 138         struct napi_struct napi;
 139
 140         struct bpf_prog __rcu *xdp_prog;
 141
 142         struct virtnet_rq_stats stats;
 143
 144         /* Chain pages by the private ptr. */
 145         struct page *pages;
 146
 147         /* Average packet length for mergeable receive buffers. */
 148         struct ewma_pkt_len mrg_avg_pkt_len;
 149
 150         /* Page frag for packet buffer allocation. */
 151         struct page_frag alloc_frag;
 152
 153         /* RX: fragments + linear part + virtio header */
 154         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 155
 156         /* Min single buffer size for mergeable buffers case. */
 157         unsigned int min_buf_len;
 158
 159         /* Name of this receive queue: input.$index */
 160         char name[40];
 161
 162         struct xdp_rxq_info xdp_rxq;
 163 };
 164
 165 /* Control VQ buffers: protected by the rtnl lock */
 166 struct control_buf {
 167         struct virtio_net_ctrl_hdr hdr;
 168         virtio_net_ctrl_ack status;
 169         struct virtio_net_ctrl_mq mq;
 170         u8 promisc;
 171         u8 allmulti;
 172         __virtio16 vid;
 173         __virtio64 offloads;
 174 };
 175
 176 struct virtnet_info {
 177         struct virtio_device *vdev;
 178         struct virtqueue *cvq;
 179         struct net_device *dev;
 180         struct send_queue *sq;
 181         struct receive_queue *rq;
 182         unsigned int status;
 183
 184         /* Max # of queue pairs supported by the device */
 185         u16 max_queue_pairs;
 186
 187         /* # of queue pairs currently used by the driver */
 188         u16 curr_queue_pairs;
 189
 190         /* # of XDP queue pairs currently used by the driver */
 191         u16 xdp_queue_pairs;
 192
 193         /* I like... big packets and I cannot lie! */
 194         bool big_packets;
 195
 196         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 197         bool mergeable_rx_bufs;
 198
 199         /* Has control virtqueue */
 200         bool has_cvq;
 201
 202         /* Host can handle any s/g split between our header and packet data */
 203         bool any_header_sg;
 204
 205         /* Packet virtio header size */
 206         u8 hdr_len;
 207
 208         /* Work struct for refilling if we run low on memory. */
 209         struct delayed_work refill;
 210
 211         /* Work struct for config space updates */
 212         struct work_struct config_work;
 213
 214         /* Does the affinity hint is set for virtqueues? */
 215         bool affinity_hint_set;
 216
 217         /* CPU hotplug instances for online & dead */
 218         struct hlist_node node;
 219         struct hlist_node node_dead;
 220
 221         struct control_buf *ctrl;
 222
 223         /* Ethtool settings */
 224         u8 duplex;
 225         u32 speed;
 226
 227         unsigned long guest_offloads;
 228         unsigned long guest_offloads_capable;
 229
 230         /* failover when STANDBY feature enabled */
 231         struct failover *failover;
 232 };
 233
 234 struct padded_vnet_hdr {
 235         struct virtio_net_hdr_mrg_rxbuf hdr;
 236         /*
 237          * hdr is in a separate sg buffer, and data sg buffer shares same page
 238          * with this header sg. This padding makes next sg 16 byte aligned
 239          * after the header.
 240          */
 241         char padding[4];
 242 };
 243
 244 static bool is_xdp_frame(void *ptr)
 245 {
 246         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 247 }
 248
 249 static void *xdp_to_ptr(struct xdp_frame *ptr)
 250 {
 251         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 252 }
 253
 254 static struct xdp_frame *ptr_to_xdp(void *ptr)
 255 {
 256         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 257 }
 258
 259 /* Converting between virtqueue no. and kernel tx/rx queue no.
 260  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 261  */
 262 static int vq2txq(struct virtqueue *vq)
 263 {
 264         return (vq->index - 1) / 2;
 265 }
 266
 267 static int txq2vq(int txq)
 268 {
 269         return txq * 2 + 1;
 270 }
 271
 272 static int vq2rxq(struct virtqueue *vq)
 273 {
 274         return vq->index / 2;
 275 }
 276
 277 static int rxq2vq(int rxq)
 278 {
 279         return rxq * 2;
 280 }
 281
 282 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 283 {
 284         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 285 }
 286
 287 /*
 288  * private is used to chain pages for big packets, put the whole
 289  * most recent used list in the beginning for reuse
 290  */
 291 static void give_pages(struct receive_queue *rq, struct page *page)
 292 {
 293         struct page *end;
 294
 295         /* Find end of list, sew whole thing into vi->rq.pages. */
 296         for (end = page; end->private; end = (struct page *)end->private);
 297         end->private = (unsigned long)rq->pages;
 298         rq->pages = page;
 299 }
 300
 301 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 302 {
 303         struct page *p = rq->pages;
 304
 305         if (p) {
 306                 rq->pages = (struct page *)p->private;
 307                 /* clear private here, it is used to chain pages */
 308                 p->private = 0;
 309         } else
 310                 p = alloc_page(gfp_mask);
 311         return p;
 312 }
 313
 314 static void virtqueue_napi_schedule(struct napi_struct *napi,
 315                                     struct virtqueue *vq)
 316 {
 317         if (napi_schedule_prep(napi)) {
 318                 virtqueue_disable_cb(vq);
 319                 __napi_schedule(napi);
 320         }
 321 }
 322
 323 static void virtqueue_napi_complete(struct napi_struct *napi,
 324                                     struct virtqueue *vq, int processed)
 325 {
 326         int opaque;
 327
 328         opaque = virtqueue_enable_cb_prepare(vq);
 329         if (napi_complete_done(napi, processed)) {
 330                 if (unlikely(virtqueue_poll(vq, opaque)))
 331                         virtqueue_napi_schedule(napi, vq);
 332         } else {
 333                 virtqueue_disable_cb(vq);
 334         }
 335 }
 336
 337 static void skb_xmit_done(struct virtqueue *vq)
 338 {
 339         struct virtnet_info *vi = vq->vdev->priv;
 340         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 341
 342         /* Suppress further interrupts. */
 343         virtqueue_disable_cb(vq);
 344
 345         if (napi->weight)
 346                 virtqueue_napi_schedule(napi, vq);
 347         else
 348                 /* We were probably waiting for more output buffers. */
 349                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 350 }
 351
 352 #define MRG_CTX_HEADER_SHIFT 22
 353 static void *mergeable_len_to_ctx(unsigned int truesize,
 354                                   unsigned int headroom)
 355 {
 356         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 357 }
 358
 359 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 360 {
 361         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 362 }
 363
 364 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 365 {
 366         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 367 }
 368
 369 /* Called from bottom half context */
 370 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 371                                    struct receive_queue *rq,
 372                                    struct page *page, unsigned int offset,
 373                                    unsigned int len, unsigned int truesize,
 374                                    bool hdr_valid, unsigned int metasize)
 375 {
 376         struct sk_buff *skb;
 377         struct virtio_net_hdr_mrg_rxbuf *hdr;
 378         unsigned int copy, hdr_len, hdr_padded_len;
 379         char *p;
 380
 381         p = page_address(page) + offset;
 382
 383         /* copy small packet so we can reuse these pages for small data */
 384         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 385         if (unlikely(!skb))
 386                 return NULL;
 387
 388         hdr = skb_vnet_hdr(skb);
 389
 390         hdr_len = vi->hdr_len;
 391         if (vi->mergeable_rx_bufs)
 392                 hdr_padded_len = sizeof(*hdr);
 393         else
 394                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 395
 396         /* hdr_valid means no XDP, so we can copy the vnet header */
 397         if (hdr_valid)
 398                 memcpy(hdr, p, hdr_len);
 399
 400         len -= hdr_len;
 401         offset += hdr_padded_len;
 402         p += hdr_padded_len;
 403
 404         copy = len;
 405         if (copy > skb_tailroom(skb))
 406                 copy = skb_tailroom(skb);
 407         skb_put_data(skb, p, copy);
 408
 409         if (metasize) {
 410                 __skb_pull(skb, metasize);
 411                 skb_metadata_set(skb, metasize);
 412         }
 413
 414         len -= copy;
 415         offset += copy;
 416
 417         if (vi->mergeable_rx_bufs) {
 418                 if (len)
 419                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 420                 else
 421                         put_page(page);
 422                 return skb;
 423         }
 424
 425         /*
 426          * Verify that we can indeed put this data into a skb.
 427          * This is here to handle cases when the device erroneously
 428          * tries to receive more than is possible. This is usually
 429          * the case of a broken device.
 430          */
 431         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 432                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 433                 dev_kfree_skb(skb);
 434                 return NULL;
 435         }
 436         BUG_ON(offset >= PAGE_SIZE);
 437         while (len) {
 438                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 439                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 440                                 frag_size, truesize);
 441                 len -= frag_size;
 442                 page = (struct page *)page->private;
 443                 offset = 0;
 444         }
 445
 446         if (page)
 447                 give_pages(rq, page);
 448
 449         return skb;
 450 }
 451
 452 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 453                                    struct send_queue *sq,
 454                                    struct xdp_frame *xdpf)
 455 {
 456         struct virtio_net_hdr_mrg_rxbuf *hdr;
 457         int err;
 458
 459         if (unlikely(xdpf->headroom < vi->hdr_len))
 460                 return -EOVERFLOW;
 461
 462         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 463         xdpf->data -= vi->hdr_len;
 464         /* Zero header and leave csum up to XDP layers */
 465         hdr = xdpf->data;
 466         memset(hdr, 0, vi->hdr_len);
 467         xdpf->len   += vi->hdr_len;
 468
 469         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 470
 471         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
 472                                    GFP_ATOMIC);
 473         if (unlikely(err))
 474                 return -ENOSPC; /* Caller handle free/refcnt */
 475
 476         return 0;
 477 }
 478
 479 static struct send_queue *virtnet_xdp_sq(struct virtnet_info *vi)
 480 {
 481         unsigned int qp;
 482
 483         qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
 484         return &vi->sq[qp];
 485 }
 486
 487 static int virtnet_xdp_xmit(struct net_device *dev,
 488                             int n, struct xdp_frame **frames, u32 flags)
 489 {
 490         struct virtnet_info *vi = netdev_priv(dev);
 491         struct receive_queue *rq = vi->rq;
 492         struct bpf_prog *xdp_prog;
 493         struct send_queue *sq;
 494         unsigned int len;
 495         int packets = 0;
 496         int bytes = 0;
 497         int drops = 0;
 498         int kicks = 0;
 499         int ret, err;
 500         void *ptr;
 501         int i;
 502
 503         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 504          * indicate XDP resources have been successfully allocated.
 505          */
 506         xdp_prog = rcu_access_pointer(rq->xdp_prog);
 507         if (!xdp_prog)
 508                 return -ENXIO;
 509
 510         sq = virtnet_xdp_sq(vi);
 511
 512         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 513                 ret = -EINVAL;
 514                 drops = n;
 515                 goto out;
 516         }
 517
 518         /* Free up any pending old buffers before queueing new ones. */
 519         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 520                 if (likely(is_xdp_frame(ptr))) {
 521                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 522
 523                         bytes += frame->len;
 524                         xdp_return_frame(frame);
 525                 } else {
 526                         struct sk_buff *skb = ptr;
 527
 528                         bytes += skb->len;
 529                         napi_consume_skb(skb, false);
 530                 }
 531                 packets++;
 532         }
 533
 534         for (i = 0; i < n; i++) {
 535                 struct xdp_frame *xdpf = frames[i];
 536
 537                 err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
 538                 if (err) {
 539                         xdp_return_frame_rx_napi(xdpf);
 540                         drops++;
 541                 }
 542         }
 543         ret = n - drops;
 544
 545         if (flags & XDP_XMIT_FLUSH) {
 546                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
 547                         kicks = 1;
 548         }
 549 out:
 550         u64_stats_update_begin(&sq->stats.syncp);
 551         sq->stats.bytes += bytes;
 552         sq->stats.packets += packets;
 553         sq->stats.xdp_tx += n;
 554         sq->stats.xdp_tx_drops += drops;
 555         sq->stats.kicks += kicks;
 556         u64_stats_update_end(&sq->stats.syncp);
 557
 558         return ret;
 559 }
 560
 561 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 562 {
 563         return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
 564 }
 565
 566 /* We copy the packet for XDP in the following cases:
 567  *
 568  * 1) Packet is scattered across multiple rx buffers.
 569  * 2) Headroom space is insufficient.
 570  *
 571  * This is inefficient but it's a temporary condition that
 572  * we hit right after XDP is enabled and until queue is refilled
 573  * with large buffers with sufficient headroom - so it should affect
 574  * at most queue size packets.
 575  * Afterwards, the conditions to enable
 576  * XDP should preclude the underlying device from sending packets
 577  * across multiple buffers (num_buf > 1), and we make sure buffers
 578  * have enough headroom.
 579  */
 580 static struct page *xdp_linearize_page(struct receive_queue *rq,
 581                                        u16 *num_buf,
 582                                        struct page *p,
 583                                        int offset,
 584                                        int page_off,
 585                                        unsigned int *len)
 586 {
 587         struct page *page = alloc_page(GFP_ATOMIC);
 588
 589         if (!page)
 590                 return NULL;
 591
 592         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 593         page_off += *len;
 594
 595         while (--*num_buf) {
 596                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 597                 unsigned int buflen;
 598                 void *buf;
 599                 int off;
 600
 601                 buf = virtqueue_get_buf(rq->vq, &buflen);
 602                 if (unlikely(!buf))
 603                         goto err_buf;
 604
 605                 p = virt_to_head_page(buf);
 606                 off = buf - page_address(p);
 607
 608                 /* guard against a misconfigured or uncooperative backend that
 609                  * is sending packet larger than the MTU.
 610                  */
 611                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 612                         put_page(p);
 613                         goto err_buf;
 614                 }
 615
 616                 memcpy(page_address(page) + page_off,
 617                        page_address(p) + off, buflen);
 618                 page_off += buflen;
 619                 put_page(p);
 620         }
 621
 622         /* Headroom does not contribute to packet length */
 623         *len = page_off - VIRTIO_XDP_HEADROOM;
 624         return page;
 625 err_buf:
 626         __free_pages(page, 0);
 627         return NULL;
 628 }
 629
 630 static struct sk_buff *receive_small(struct net_device *dev,
 631                                      struct virtnet_info *vi,
 632                                      struct receive_queue *rq,
 633                                      void *buf, void *ctx,
 634                                      unsigned int len,
 635                                      unsigned int *xdp_xmit,
 636                                      struct virtnet_rq_stats *stats)
 637 {
 638         struct sk_buff *skb;
 639         struct bpf_prog *xdp_prog;
 640         unsigned int xdp_headroom = (unsigned long)ctx;
 641         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 642         unsigned int headroom = vi->hdr_len + header_offset;
 643         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 644                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 645         struct page *page = virt_to_head_page(buf);
 646         unsigned int delta = 0;
 647         struct page *xdp_page;
 648         int err;
 649         unsigned int metasize = 0;
 650
 651         len -= vi->hdr_len;
 652         stats->bytes += len;
 653
 654         rcu_read_lock();
 655         xdp_prog = rcu_dereference(rq->xdp_prog);
 656         if (xdp_prog) {
 657                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 658                 struct xdp_frame *xdpf;
 659                 struct xdp_buff xdp;
 660                 void *orig_data;
 661                 u32 act;
 662
 663                 if (unlikely(hdr->hdr.gso_type))
 664                         goto err_xdp;
 665
 666                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 667                         int offset = buf - page_address(page) + header_offset;
 668                         unsigned int tlen = len + vi->hdr_len;
 669                         u16 num_buf = 1;
 670
 671                         xdp_headroom = virtnet_get_headroom(vi);
 672                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 673                         headroom = vi->hdr_len + header_offset;
 674                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 675                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 676                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 677                                                       offset, header_offset,
 678                                                       &tlen);
 679                         if (!xdp_page)
 680                                 goto err_xdp;
 681
 682                         buf = page_address(xdp_page);
 683                         put_page(page);
 684                         page = xdp_page;
 685                 }
 686
 687                 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
 688                 xdp.data = xdp.data_hard_start + xdp_headroom;
 689                 xdp.data_end = xdp.data + len;
 690                 xdp.data_meta = xdp.data;
 691                 xdp.rxq = &rq->xdp_rxq;
 692                 orig_data = xdp.data;
 693                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 694                 stats->xdp_packets++;
 695
 696                 switch (act) {
 697                 case XDP_PASS:
 698                         /* Recalculate length in case bpf program changed it */
 699                         delta = orig_data - xdp.data;
 700                         len = xdp.data_end - xdp.data;
 701                         metasize = xdp.data - xdp.data_meta;
 702                         break;
 703                 case XDP_TX:
 704                         stats->xdp_tx++;
 705                         xdpf = convert_to_xdp_frame(&xdp);
 706                         if (unlikely(!xdpf))
 707                                 goto err_xdp;
 708                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 709                         if (unlikely(err < 0)) {
 710                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 711                                 goto err_xdp;
 712                         }
 713                         *xdp_xmit |= VIRTIO_XDP_TX;
 714                         rcu_read_unlock();
 715                         goto xdp_xmit;
 716                 case XDP_REDIRECT:
 717                         stats->xdp_redirects++;
 718                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 719                         if (err)
 720                                 goto err_xdp;
 721                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 722                         rcu_read_unlock();
 723                         goto xdp_xmit;
 724                 default:
 725                         bpf_warn_invalid_xdp_action(act);
 726                         /* fall through */
 727                 case XDP_ABORTED:
 728                         trace_xdp_exception(vi->dev, xdp_prog, act);
 729                 case XDP_DROP:
 730                         goto err_xdp;
 731                 }
 732         }
 733         rcu_read_unlock();
 734
 735         skb = build_skb(buf, buflen);
 736         if (!skb) {
 737                 put_page(page);
 738                 goto err;
 739         }
 740         skb_reserve(skb, headroom - delta);
 741         skb_put(skb, len);
 742         if (!xdp_prog) {
 743                 buf += header_offset;
 744                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 745         } /* keep zeroed vnet hdr since XDP is loaded */
 746
 747         if (metasize)
 748                 skb_metadata_set(skb, metasize);
 749
 750 err:
 751         return skb;
 752
 753 err_xdp:
 754         rcu_read_unlock();
 755         stats->xdp_drops++;
 756         stats->drops++;
 757         put_page(page);
 758 xdp_xmit:
 759         return NULL;
 760 }
 761
 762 static struct sk_buff *receive_big(struct net_device *dev,
 763                                    struct virtnet_info *vi,
 764                                    struct receive_queue *rq,
 765                                    void *buf,
 766                                    unsigned int len,
 767                                    struct virtnet_rq_stats *stats)
 768 {
 769         struct page *page = buf;
 770         struct sk_buff *skb =
 771                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 772
 773         stats->bytes += len - vi->hdr_len;
 774         if (unlikely(!skb))
 775                 goto err;
 776
 777         return skb;
 778
 779 err:
 780         stats->drops++;
 781         give_pages(rq, page);
 782         return NULL;
 783 }
 784
 785 static struct sk_buff *receive_mergeable(struct net_device *dev,
 786                                          struct virtnet_info *vi,
 787                                          struct receive_queue *rq,
 788                                          void *buf,
 789                                          void *ctx,
 790                                          unsigned int len,
 791                                          unsigned int *xdp_xmit,
 792                                          struct virtnet_rq_stats *stats)
 793 {
 794         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 795         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 796         struct page *page = virt_to_head_page(buf);
 797         int offset = buf - page_address(page);
 798         struct sk_buff *head_skb, *curr_skb;
 799         struct bpf_prog *xdp_prog;
 800         unsigned int truesize;
 801         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 802         int err;
 803         unsigned int metasize = 0;
 804
 805         head_skb = NULL;
 806         stats->bytes += len - vi->hdr_len;
 807
 808         rcu_read_lock();
 809         xdp_prog = rcu_dereference(rq->xdp_prog);
 810         if (xdp_prog) {
 811                 struct xdp_frame *xdpf;
 812                 struct page *xdp_page;
 813                 struct xdp_buff xdp;
 814                 void *data;
 815                 u32 act;
 816
 817                 /* Transient failure which in theory could occur if
 818                  * in-flight packets from before XDP was enabled reach
 819                  * the receive path after XDP is loaded.
 820                  */
 821                 if (unlikely(hdr->hdr.gso_type))
 822                         goto err_xdp;
 823
 824                 /* This happens when rx buffer size is underestimated
 825                  * or headroom is not enough because of the buffer
 826                  * was refilled before XDP is set. This should only
 827                  * happen for the first several packets, so we don't
 828                  * care much about its performance.
 829                  */
 830                 if (unlikely(num_buf > 1 ||
 831                              headroom < virtnet_get_headroom(vi))) {
 832                         /* linearize data for XDP */
 833                         xdp_page = xdp_linearize_page(rq, &num_buf,
 834                                                       page, offset,
 835                                                       VIRTIO_XDP_HEADROOM,
 836                                                       &len);
 837                         if (!xdp_page)
 838                                 goto err_xdp;
 839                         offset = VIRTIO_XDP_HEADROOM;
 840                 } else {
 841                         xdp_page = page;
 842                 }
 843
 844                 /* Allow consuming headroom but reserve enough space to push
 845                  * the descriptor on if we get an XDP_TX return code.
 846                  */
 847                 data = page_address(xdp_page) + offset;
 848                 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
 849                 xdp.data = data + vi->hdr_len;
 850                 xdp.data_end = xdp.data + (len - vi->hdr_len);
 851                 xdp.data_meta = xdp.data;
 852                 xdp.rxq = &rq->xdp_rxq;
 853
 854                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 855                 stats->xdp_packets++;
 856
 857                 switch (act) {
 858                 case XDP_PASS:
 859                         metasize = xdp.data - xdp.data_meta;
 860
 861                         /* recalculate offset to account for any header
 862                          * adjustments and minus the metasize to copy the
 863                          * metadata in page_to_skb(). Note other cases do not
 864                          * build an skb and avoid using offset
 865                          */
 866                         offset = xdp.data - page_address(xdp_page) -
 867                                  vi->hdr_len - metasize;
 868
 869                         /* recalculate len if xdp.data, xdp.data_end or
 870                          * xdp.data_meta were adjusted
 871                          */
 872                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 873                         /* We can only create skb based on xdp_page. */
 874                         if (unlikely(xdp_page != page)) {
 875                                 rcu_read_unlock();
 876                                 put_page(page);
 877                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
 878                                                        len, PAGE_SIZE, false,
 879                                                        metasize);
 880                                 return head_skb;
 881                         }
 882                         break;
 883                 case XDP_TX:
 884                         stats->xdp_tx++;
 885                         xdpf = convert_to_xdp_frame(&xdp);
 886                         if (unlikely(!xdpf))
 887                                 goto err_xdp;
 888                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 889                         if (unlikely(err < 0)) {
 890                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 891                                 if (unlikely(xdp_page != page))
 892                                         put_page(xdp_page);
 893                                 goto err_xdp;
 894                         }
 895                         *xdp_xmit |= VIRTIO_XDP_TX;
 896                         if (unlikely(xdp_page != page))
 897                                 put_page(page);
 898                         rcu_read_unlock();
 899                         goto xdp_xmit;
 900                 case XDP_REDIRECT:
 901                         stats->xdp_redirects++;
 902                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 903                         if (err) {
 904                                 if (unlikely(xdp_page != page))
 905                                         put_page(xdp_page);
 906                                 goto err_xdp;
 907                         }
 908                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 909                         if (unlikely(xdp_page != page))
 910                                 put_page(page);
 911                         rcu_read_unlock();
 912                         goto xdp_xmit;
 913                 default:
 914                         bpf_warn_invalid_xdp_action(act);
 915                         /* fall through */
 916                 case XDP_ABORTED:
 917                         trace_xdp_exception(vi->dev, xdp_prog, act);
 918                         /* fall through */
 919                 case XDP_DROP:
 920                         if (unlikely(xdp_page != page))
 921                                 __free_pages(xdp_page, 0);
 922                         goto err_xdp;
 923                 }
 924         }
 925         rcu_read_unlock();
 926
 927         truesize = mergeable_ctx_to_truesize(ctx);
 928         if (unlikely(len > truesize)) {
 929                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 930                          dev->name, len, (unsigned long)ctx);
 931                 dev->stats.rx_length_errors++;
 932                 goto err_skb;
 933         }
 934
 935         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
 936                                metasize);
 937         curr_skb = head_skb;
 938
 939         if (unlikely(!curr_skb))
 940                 goto err_skb;
 941         while (--num_buf) {
 942                 int num_skb_frags;
 943
 944                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
 945                 if (unlikely(!buf)) {
 946                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 947                                  dev->name, num_buf,
 948                                  virtio16_to_cpu(vi->vdev,
 949                                                  hdr->num_buffers));
 950                         dev->stats.rx_length_errors++;
 951                         goto err_buf;
 952                 }
 953
 954                 stats->bytes += len;
 955                 page = virt_to_head_page(buf);
 956
 957                 truesize = mergeable_ctx_to_truesize(ctx);
 958                 if (unlikely(len > truesize)) {
 959                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 960                                  dev->name, len, (unsigned long)ctx);
 961                         dev->stats.rx_length_errors++;
 962                         goto err_skb;
 963                 }
 964
 965                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 966                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 967                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 968
 969                         if (unlikely(!nskb))
 970                                 goto err_skb;
 971                         if (curr_skb == head_skb)
 972                                 skb_shinfo(curr_skb)->frag_list = nskb;
 973                         else
 974                                 curr_skb->next = nskb;
 975                         curr_skb = nskb;
 976                         head_skb->truesize += nskb->truesize;
 977                         num_skb_frags = 0;
 978                 }
 979                 if (curr_skb != head_skb) {
 980                         head_skb->data_len += len;
 981                         head_skb->len += len;
 982                         head_skb->truesize += truesize;
 983                 }
 984                 offset = buf - page_address(page);
 985                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 986                         put_page(page);
 987                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 988                                              len, truesize);
 989                 } else {
 990                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 991                                         offset, len, truesize);
 992                 }
 993         }
 994
 995         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 996         return head_skb;
 997
 998 err_xdp:
 999         rcu_read_unlock();
1000         stats->xdp_drops++;
1001 err_skb:
1002         put_page(page);
1003         while (num_buf-- > 1) {
1004                 buf = virtqueue_get_buf(rq->vq, &len);
1005                 if (unlikely(!buf)) {
1006                         pr_debug("%s: rx error: %d buffers missing\n",
1007                                  dev->name, num_buf);
1008                         dev->stats.rx_length_errors++;
1009                         break;
1010                 }
1011                 stats->bytes += len;
1012                 page = virt_to_head_page(buf);
1013                 put_page(page);
1014         }
1015 err_buf:
1016         stats->drops++;
1017         dev_kfree_skb(head_skb);
1018 xdp_xmit:
1019         return NULL;
1020 }
1021
1022 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1023                         void *buf, unsigned int len, void **ctx,
1024                         unsigned int *xdp_xmit,
1025                         struct virtnet_rq_stats *stats)
1026 {
1027         struct net_device *dev = vi->dev;
1028         struct sk_buff *skb;
1029         struct virtio_net_hdr_mrg_rxbuf *hdr;
1030
1031         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1032                 pr_debug("%s: short packet %i\n", dev->name, len);
1033                 dev->stats.rx_length_errors++;
1034                 if (vi->mergeable_rx_bufs) {
1035                         put_page(virt_to_head_page(buf));
1036                 } else if (vi->big_packets) {
1037                         give_pages(rq, buf);
1038                 } else {
1039                         put_page(virt_to_head_page(buf));
1040                 }
1041                 return;
1042         }
1043
1044         if (vi->mergeable_rx_bufs)
1045                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1046                                         stats);
1047         else if (vi->big_packets)
1048                 skb = receive_big(dev, vi, rq, buf, len, stats);
1049         else
1050                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1051
1052         if (unlikely(!skb))
1053                 return;
1054
1055         hdr = skb_vnet_hdr(skb);
1056
1057         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1058                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1059
1060         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1061                                   virtio_is_little_endian(vi->vdev))) {
1062                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1063                                      dev->name, hdr->hdr.gso_type,
1064                                      hdr->hdr.gso_size);
1065                 goto frame_err;
1066         }
1067
1068         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1069         skb->protocol = eth_type_trans(skb, dev);
1070         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1071                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1072
1073         napi_gro_receive(&rq->napi, skb);
1074         return;
1075
1076 frame_err:
1077         dev->stats.rx_frame_errors++;
1078         dev_kfree_skb(skb);
1079 }
1080
1081 /* Unlike mergeable buffers, all buffers are allocated to the
1082  * same size, except for the headroom. For this reason we do
1083  * not need to use  mergeable_len_to_ctx here - it is enough
1084  * to store the headroom as the context ignoring the truesize.
1085  */
1086 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1087                              gfp_t gfp)
1088 {
1089         struct page_frag *alloc_frag = &rq->alloc_frag;
1090         char *buf;
1091         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1092         void *ctx = (void *)(unsigned long)xdp_headroom;
1093         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1094         int err;
1095
1096         len = SKB_DATA_ALIGN(len) +
1097               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1098         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1099                 return -ENOMEM;
1100
1101         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1102         get_page(alloc_frag->page);
1103         alloc_frag->offset += len;
1104         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1105                     vi->hdr_len + GOOD_PACKET_LEN);
1106         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1107         if (err < 0)
1108                 put_page(virt_to_head_page(buf));
1109         return err;
1110 }
1111
1112 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1113                            gfp_t gfp)
1114 {
1115         struct page *first, *list = NULL;
1116         char *p;
1117         int i, err, offset;
1118
1119         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1120
1121         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1122         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1123                 first = get_a_page(rq, gfp);
1124                 if (!first) {
1125                         if (list)
1126                                 give_pages(rq, list);
1127                         return -ENOMEM;
1128                 }
1129                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1130
1131                 /* chain new page in list head to match sg */
1132                 first->private = (unsigned long)list;
1133                 list = first;
1134         }
1135
1136         first = get_a_page(rq, gfp);
1137         if (!first) {
1138                 give_pages(rq, list);
1139                 return -ENOMEM;
1140         }
1141         p = page_address(first);
1142
1143         /* rq->sg[0], rq->sg[1] share the same page */
1144         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1145         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1146
1147         /* rq->sg[1] for data packet, from offset */
1148         offset = sizeof(struct padded_vnet_hdr);
1149         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1150
1151         /* chain first in list head */
1152         first->private = (unsigned long)list;
1153         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1154                                   first, gfp);
1155         if (err < 0)
1156                 give_pages(rq, first);
1157
1158         return err;
1159 }
1160
1161 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1162                                           struct ewma_pkt_len *avg_pkt_len,
1163                                           unsigned int room)
1164 {
1165         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1166         unsigned int len;
1167
1168         if (room)
1169                 return PAGE_SIZE - room;
1170
1171         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1172                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1173
1174         return ALIGN(len, L1_CACHE_BYTES);
1175 }
1176
1177 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1178                                  struct receive_queue *rq, gfp_t gfp)
1179 {
1180         struct page_frag *alloc_frag = &rq->alloc_frag;
1181         unsigned int headroom = virtnet_get_headroom(vi);
1182         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1183         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1184         char *buf;
1185         void *ctx;
1186         int err;
1187         unsigned int len, hole;
1188
1189         /* Extra tailroom is needed to satisfy XDP's assumption. This
1190          * means rx frags coalescing won't work, but consider we've
1191          * disabled GSO for XDP, it won't be a big issue.
1192          */
1193         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1194         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1195                 return -ENOMEM;
1196
1197         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1198         buf += headroom; /* advance address leaving hole at front of pkt */
1199         get_page(alloc_frag->page);
1200         alloc_frag->offset += len + room;
1201         hole = alloc_frag->size - alloc_frag->offset;
1202         if (hole < len + room) {
1203                 /* To avoid internal fragmentation, if there is very likely not
1204                  * enough space for another buffer, add the remaining space to
1205                  * the current buffer.
1206                  */
1207                 len += hole;
1208                 alloc_frag->offset += hole;
1209         }
1210
1211         sg_init_one(rq->sg, buf, len);
1212         ctx = mergeable_len_to_ctx(len, headroom);
1213         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1214         if (err < 0)
1215                 put_page(virt_to_head_page(buf));
1216
1217         return err;
1218 }
1219
1220 /*
1221  * Returns false if we couldn't fill entirely (OOM).
1222  *
1223  * Normally run in the receive path, but can also be run from ndo_open
1224  * before we're receiving packets, or from refill_work which is
1225  * careful to disable receiving (using napi_disable).
1226  */
1227 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1228                           gfp_t gfp)
1229 {
1230         int err;
1231         bool oom;
1232
1233         do {
1234                 if (vi->mergeable_rx_bufs)
1235                         err = add_recvbuf_mergeable(vi, rq, gfp);
1236                 else if (vi->big_packets)
1237                         err = add_recvbuf_big(vi, rq, gfp);
1238                 else
1239                         err = add_recvbuf_small(vi, rq, gfp);
1240
1241                 oom = err == -ENOMEM;
1242                 if (err)
1243                         break;
1244         } while (rq->vq->num_free);
1245         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1246                 unsigned long flags;
1247
1248                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1249                 rq->stats.kicks++;
1250                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1251         }
1252
1253         return !oom;
1254 }
1255
1256 static void skb_recv_done(struct virtqueue *rvq)
1257 {
1258         struct virtnet_info *vi = rvq->vdev->priv;
1259         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1260
1261         virtqueue_napi_schedule(&rq->napi, rvq);
1262 }
1263
1264 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1265 {
1266         napi_enable(napi);
1267
1268         /* If all buffers were filled by other side before we napi_enabled, we
1269          * won't get another interrupt, so process any outstanding packets now.
1270          * Call local_bh_enable after to trigger softIRQ processing.
1271          */
1272         local_bh_disable();
1273         virtqueue_napi_schedule(napi, vq);
1274         local_bh_enable();
1275 }
1276
1277 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1278                                    struct virtqueue *vq,
1279                                    struct napi_struct *napi)
1280 {
1281         if (!napi->weight)
1282                 return;
1283
1284         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1285          * enable the feature if this is likely affine with the transmit path.
1286          */
1287         if (!vi->affinity_hint_set) {
1288                 napi->weight = 0;
1289                 return;
1290         }
1291
1292         return virtnet_napi_enable(vq, napi);
1293 }
1294
1295 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1296 {
1297         if (napi->weight)
1298                 napi_disable(napi);
1299 }
1300
1301 static void refill_work(struct work_struct *work)
1302 {
1303         struct virtnet_info *vi =
1304                 container_of(work, struct virtnet_info, refill.work);
1305         bool still_empty;
1306         int i;
1307
1308         for (i = 0; i < vi->curr_queue_pairs; i++) {
1309                 struct receive_queue *rq = &vi->rq[i];
1310
1311                 napi_disable(&rq->napi);
1312                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1313                 virtnet_napi_enable(rq->vq, &rq->napi);
1314
1315                 /* In theory, this can happen: if we don't get any buffers in
1316                  * we will *never* try to fill again.
1317                  */
1318                 if (still_empty)
1319                         schedule_delayed_work(&vi->refill, HZ/2);
1320         }
1321 }
1322
1323 static int virtnet_receive(struct receive_queue *rq, int budget,
1324                            unsigned int *xdp_xmit)
1325 {
1326         struct virtnet_info *vi = rq->vq->vdev->priv;
1327         struct virtnet_rq_stats stats = {};
1328         unsigned int len;
1329         void *buf;
1330         int i;
1331
1332         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1333                 void *ctx;
1334
1335                 while (stats.packets < budget &&
1336                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1337                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1338                         stats.packets++;
1339                 }
1340         } else {
1341                 while (stats.packets < budget &&
1342                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1343                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1344                         stats.packets++;
1345                 }
1346         }
1347
1348         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1349                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1350                         schedule_delayed_work(&vi->refill, 0);
1351         }
1352
1353         u64_stats_update_begin(&rq->stats.syncp);
1354         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1355                 size_t offset = virtnet_rq_stats_desc[i].offset;
1356                 u64 *item;
1357
1358                 item = (u64 *)((u8 *)&rq->stats + offset);
1359                 *item += *(u64 *)((u8 *)&stats + offset);
1360         }
1361         u64_stats_update_end(&rq->stats.syncp);
1362
1363         return stats.packets;
1364 }
1365
1366 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1367 {
1368         unsigned int len;
1369         unsigned int packets = 0;
1370         unsigned int bytes = 0;
1371         void *ptr;
1372
1373         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1374                 if (likely(!is_xdp_frame(ptr))) {
1375                         struct sk_buff *skb = ptr;
1376
1377                         pr_debug("Sent skb %p\n", skb);
1378
1379                         bytes += skb->len;
1380                         napi_consume_skb(skb, in_napi);
1381                 } else {
1382                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1383
1384                         bytes += frame->len;
1385                         xdp_return_frame(frame);
1386                 }
1387                 packets++;
1388         }
1389
1390         /* Avoid overhead when no packets have been processed
1391          * happens when called speculatively from start_xmit.
1392          */
1393         if (!packets)
1394                 return;
1395
1396         u64_stats_update_begin(&sq->stats.syncp);
1397         sq->stats.bytes += bytes;
1398         sq->stats.packets += packets;
1399         u64_stats_update_end(&sq->stats.syncp);
1400 }
1401
1402 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1403 {
1404         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1405                 return false;
1406         else if (q < vi->curr_queue_pairs)
1407                 return true;
1408         else
1409                 return false;
1410 }
1411
1412 static void virtnet_poll_cleantx(struct receive_queue *rq)
1413 {
1414         struct virtnet_info *vi = rq->vq->vdev->priv;
1415         unsigned int index = vq2rxq(rq->vq);
1416         struct send_queue *sq = &vi->sq[index];
1417         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1418
1419         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1420                 return;
1421
1422         if (__netif_tx_trylock(txq)) {
1423                 free_old_xmit_skbs(sq, true);
1424                 __netif_tx_unlock(txq);
1425         }
1426
1427         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1428                 netif_tx_wake_queue(txq);
1429 }
1430
1431 static int virtnet_poll(struct napi_struct *napi, int budget)
1432 {
1433         struct receive_queue *rq =
1434                 container_of(napi, struct receive_queue, napi);
1435         struct virtnet_info *vi = rq->vq->vdev->priv;
1436         struct send_queue *sq;
1437         unsigned int received;
1438         unsigned int xdp_xmit = 0;
1439
1440         virtnet_poll_cleantx(rq);
1441
1442         received = virtnet_receive(rq, budget, &xdp_xmit);
1443
1444         /* Out of packets? */
1445         if (received < budget)
1446                 virtqueue_napi_complete(napi, rq->vq, received);
1447
1448         if (xdp_xmit & VIRTIO_XDP_REDIR)
1449                 xdp_do_flush();
1450
1451         if (xdp_xmit & VIRTIO_XDP_TX) {
1452                 sq = virtnet_xdp_sq(vi);
1453                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1454                         u64_stats_update_begin(&sq->stats.syncp);
1455                         sq->stats.kicks++;
1456                         u64_stats_update_end(&sq->stats.syncp);
1457                 }
1458         }
1459
1460         return received;
1461 }
1462
1463 static int virtnet_open(struct net_device *dev)
1464 {
1465         struct virtnet_info *vi = netdev_priv(dev);
1466         int i, err;
1467
1468         for (i = 0; i < vi->max_queue_pairs; i++) {
1469                 if (i < vi->curr_queue_pairs)
1470                         /* Make sure we have some buffers: if oom use wq. */
1471                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1472                                 schedule_delayed_work(&vi->refill, 0);
1473
1474                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1475                 if (err < 0)
1476                         return err;
1477
1478                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1479                                                  MEM_TYPE_PAGE_SHARED, NULL);
1480                 if (err < 0) {
1481                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1482                         return err;
1483                 }
1484
1485                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1486                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1487         }
1488
1489         return 0;
1490 }
1491
1492 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1493 {
1494         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1495         struct virtnet_info *vi = sq->vq->vdev->priv;
1496         unsigned int index = vq2txq(sq->vq);
1497         struct netdev_queue *txq;
1498
1499         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1500                 /* We don't need to enable cb for XDP */
1501                 napi_complete_done(napi, 0);
1502                 return 0;
1503         }
1504
1505         txq = netdev_get_tx_queue(vi->dev, index);
1506         __netif_tx_lock(txq, raw_smp_processor_id());
1507         free_old_xmit_skbs(sq, true);
1508         __netif_tx_unlock(txq);
1509
1510         virtqueue_napi_complete(napi, sq->vq, 0);
1511
1512         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1513                 netif_tx_wake_queue(txq);
1514
1515         return 0;
1516 }
1517
1518 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1519 {
1520         struct virtio_net_hdr_mrg_rxbuf *hdr;
1521         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1522         struct virtnet_info *vi = sq->vq->vdev->priv;
1523         int num_sg;
1524         unsigned hdr_len = vi->hdr_len;
1525         bool can_push;
1526
1527         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1528
1529         can_push = vi->any_header_sg &&
1530                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1531                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1532         /* Even if we can, don't push here yet as this would skew
1533          * csum_start offset below. */
1534         if (can_push)
1535                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1536         else
1537                 hdr = skb_vnet_hdr(skb);
1538
1539         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1540                                     virtio_is_little_endian(vi->vdev), false,
1541                                     0))
1542                 BUG();
1543
1544         if (vi->mergeable_rx_bufs)
1545                 hdr->num_buffers = 0;
1546
1547         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1548         if (can_push) {
1549                 __skb_push(skb, hdr_len);
1550                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1551                 if (unlikely(num_sg < 0))
1552                         return num_sg;
1553                 /* Pull header back to avoid skew in tx bytes calculations. */
1554                 __skb_pull(skb, hdr_len);
1555         } else {
1556                 sg_set_buf(sq->sg, hdr, hdr_len);
1557                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1558                 if (unlikely(num_sg < 0))
1559                         return num_sg;
1560                 num_sg++;
1561         }
1562         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1563 }
1564
1565 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1566 {
1567         struct virtnet_info *vi = netdev_priv(dev);
1568         int qnum = skb_get_queue_mapping(skb);
1569         struct send_queue *sq = &vi->sq[qnum];
1570         int err;
1571         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1572         bool kick = !netdev_xmit_more();
1573         bool use_napi = sq->napi.weight;
1574
1575         /* Free up any pending old buffers before queueing new ones. */
1576         free_old_xmit_skbs(sq, false);
1577
1578         if (use_napi && kick)
1579                 virtqueue_enable_cb_delayed(sq->vq);
1580
1581         /* timestamp packet in software */
1582         skb_tx_timestamp(skb);
1583
1584         /* Try to transmit */
1585         err = xmit_skb(sq, skb);
1586
1587         /* This should not happen! */
1588         if (unlikely(err)) {
1589                 dev->stats.tx_fifo_errors++;
1590                 if (net_ratelimit())
1591                         dev_warn(&dev->dev,
1592                                  "Unexpected TXQ (%d) queue failure: %d\n",
1593                                  qnum, err);
1594                 dev->stats.tx_dropped++;
1595                 dev_kfree_skb_any(skb);
1596                 return NETDEV_TX_OK;
1597         }
1598
1599         /* Don't wait up for transmitted skbs to be freed. */
1600         if (!use_napi) {
1601                 skb_orphan(skb);
1602                 nf_reset_ct(skb);
1603         }
1604
1605         /* If running out of space, stop queue to avoid getting packets that we
1606          * are then unable to transmit.
1607          * An alternative would be to force queuing layer to requeue the skb by
1608          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1609          * returned in a normal path of operation: it means that driver is not
1610          * maintaining the TX queue stop/start state properly, and causes
1611          * the stack to do a non-trivial amount of useless work.
1612          * Since most packets only take 1 or 2 ring slots, stopping the queue
1613          * early means 16 slots are typically wasted.
1614          */
1615         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1616                 netif_stop_subqueue(dev, qnum);
1617                 if (!use_napi &&
1618                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1619                         /* More just got used, free them then recheck. */
1620                         free_old_xmit_skbs(sq, false);
1621                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1622                                 netif_start_subqueue(dev, qnum);
1623                                 virtqueue_disable_cb(sq->vq);
1624                         }
1625                 }
1626         }
1627
1628         if (kick || netif_xmit_stopped(txq)) {
1629                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1630                         u64_stats_update_begin(&sq->stats.syncp);
1631                         sq->stats.kicks++;
1632                         u64_stats_update_end(&sq->stats.syncp);
1633                 }
1634         }
1635
1636         return NETDEV_TX_OK;
1637 }
1638
1639 /*
1640  * Send command via the control virtqueue and check status.  Commands
1641  * supported by the hypervisor, as indicated by feature bits, should
1642  * never fail unless improperly formatted.
1643  */
1644 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1645                                  struct scatterlist *out)
1646 {
1647         struct scatterlist *sgs[4], hdr, stat;
1648         unsigned out_num = 0, tmp;
1649
1650         /* Caller should know better */
1651         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1652
1653         vi->ctrl->status = ~0;
1654         vi->ctrl->hdr.class = class;
1655         vi->ctrl->hdr.cmd = cmd;
1656         /* Add header */
1657         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1658         sgs[out_num++] = &hdr;
1659
1660         if (out)
1661                 sgs[out_num++] = out;
1662
1663         /* Add return status. */
1664         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1665         sgs[out_num] = &stat;
1666
1667         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1668         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1669
1670         if (unlikely(!virtqueue_kick(vi->cvq)))
1671                 return vi->ctrl->status == VIRTIO_NET_OK;
1672
1673         /* Spin for a response, the kick causes an ioport write, trapping
1674          * into the hypervisor, so the request should be handled immediately.
1675          */
1676         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1677                !virtqueue_is_broken(vi->cvq))
1678                 cpu_relax();
1679
1680         return vi->ctrl->status == VIRTIO_NET_OK;
1681 }
1682
1683 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1684 {
1685         struct virtnet_info *vi = netdev_priv(dev);
1686         struct virtio_device *vdev = vi->vdev;
1687         int ret;
1688         struct sockaddr *addr;
1689         struct scatterlist sg;
1690
1691         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1692                 return -EOPNOTSUPP;
1693
1694         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1695         if (!addr)
1696                 return -ENOMEM;
1697
1698         ret = eth_prepare_mac_addr_change(dev, addr);
1699         if (ret)
1700                 goto out;
1701
1702         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1703                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1704                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1705                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1706                         dev_warn(&vdev->dev,
1707                                  "Failed to set mac address by vq command.\n");
1708                         ret = -EINVAL;
1709                         goto out;
1710                 }
1711         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1712                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1713                 unsigned int i;
1714
1715                 /* Naturally, this has an atomicity problem. */
1716                 for (i = 0; i < dev->addr_len; i++)
1717                         virtio_cwrite8(vdev,
1718                                        offsetof(struct virtio_net_config, mac) +
1719                                        i, addr->sa_data[i]);
1720         }
1721
1722         eth_commit_mac_addr_change(dev, p);
1723         ret = 0;
1724
1725 out:
1726         kfree(addr);
1727         return ret;
1728 }
1729
1730 static void virtnet_stats(struct net_device *dev,
1731                           struct rtnl_link_stats64 *tot)
1732 {
1733         struct virtnet_info *vi = netdev_priv(dev);
1734         unsigned int start;
1735         int i;
1736
1737         for (i = 0; i < vi->max_queue_pairs; i++) {
1738                 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1739                 struct receive_queue *rq = &vi->rq[i];
1740                 struct send_queue *sq = &vi->sq[i];
1741
1742                 do {
1743                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1744                         tpackets = sq->stats.packets;
1745                         tbytes   = sq->stats.bytes;
1746                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1747
1748                 do {
1749                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1750                         rpackets = rq->stats.packets;
1751                         rbytes   = rq->stats.bytes;
1752                         rdrops   = rq->stats.drops;
1753                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1754
1755                 tot->rx_packets += rpackets;
1756                 tot->tx_packets += tpackets;
1757                 tot->rx_bytes   += rbytes;
1758                 tot->tx_bytes   += tbytes;
1759                 tot->rx_dropped += rdrops;
1760         }
1761
1762         tot->tx_dropped = dev->stats.tx_dropped;
1763         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1764         tot->rx_length_errors = dev->stats.rx_length_errors;
1765         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1766 }
1767
1768 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1769 {
1770         rtnl_lock();
1771         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1772                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1773                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1774         rtnl_unlock();
1775 }
1776
1777 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1778 {
1779         struct scatterlist sg;
1780         struct net_device *dev = vi->dev;
1781
1782         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1783                 return 0;
1784
1785         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1786         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1787
1788         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1789                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1790                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1791                          queue_pairs);
1792                 return -EINVAL;
1793         } else {
1794                 vi->curr_queue_pairs = queue_pairs;
1795                 /* virtnet_open() will refill when device is going to up. */
1796                 if (dev->flags & IFF_UP)
1797                         schedule_delayed_work(&vi->refill, 0);
1798         }
1799
1800         return 0;
1801 }
1802
1803 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1804 {
1805         int err;
1806
1807         rtnl_lock();
1808         err = _virtnet_set_queues(vi, queue_pairs);
1809         rtnl_unlock();
1810         return err;
1811 }
1812
1813 static int virtnet_close(struct net_device *dev)
1814 {
1815         struct virtnet_info *vi = netdev_priv(dev);
1816         int i;
1817
1818         /* Make sure refill_work doesn't re-enable napi! */
1819         cancel_delayed_work_sync(&vi->refill);
1820
1821         for (i = 0; i < vi->max_queue_pairs; i++) {
1822                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1823                 napi_disable(&vi->rq[i].napi);
1824                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1825         }
1826
1827         return 0;
1828 }
1829
1830 static void virtnet_set_rx_mode(struct net_device *dev)
1831 {
1832         struct virtnet_info *vi = netdev_priv(dev);
1833         struct scatterlist sg[2];
1834         struct virtio_net_ctrl_mac *mac_data;
1835         struct netdev_hw_addr *ha;
1836         int uc_count;
1837         int mc_count;
1838         void *buf;
1839         int i;
1840
1841         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1842         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1843                 return;
1844
1845         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1846         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1847
1848         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1849
1850         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1851                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1852                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1853                          vi->ctrl->promisc ? "en" : "dis");
1854
1855         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1856
1857         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1858                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1859                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1860                          vi->ctrl->allmulti ? "en" : "dis");
1861
1862         uc_count = netdev_uc_count(dev);
1863         mc_count = netdev_mc_count(dev);
1864         /* MAC filter - use one buffer for both lists */
1865         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1866                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1867         mac_data = buf;
1868         if (!buf)
1869                 return;
1870
1871         sg_init_table(sg, 2);
1872
1873         /* Store the unicast list and count in the front of the buffer */
1874         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1875         i = 0;
1876         netdev_for_each_uc_addr(ha, dev)
1877                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1878
1879         sg_set_buf(&sg[0], mac_data,
1880                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1881
1882         /* multicast list and count fill the end */
1883         mac_data = (void *)&mac_data->macs[uc_count][0];
1884
1885         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1886         i = 0;
1887         netdev_for_each_mc_addr(ha, dev)
1888                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1889
1890         sg_set_buf(&sg[1], mac_data,
1891                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1892
1893         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1894                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1895                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1896
1897         kfree(buf);
1898 }
1899
1900 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1901                                    __be16 proto, u16 vid)
1902 {
1903         struct virtnet_info *vi = netdev_priv(dev);
1904         struct scatterlist sg;
1905
1906         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1907         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1908
1909         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1910                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1911                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1912         return 0;
1913 }
1914
1915 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1916                                     __be16 proto, u16 vid)
1917 {
1918         struct virtnet_info *vi = netdev_priv(dev);
1919         struct scatterlist sg;
1920
1921         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1922         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1923
1924         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1925                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1926                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1927         return 0;
1928 }
1929
1930 static void virtnet_clean_affinity(struct virtnet_info *vi)
1931 {
1932         int i;
1933
1934         if (vi->affinity_hint_set) {
1935                 for (i = 0; i < vi->max_queue_pairs; i++) {
1936                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
1937                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
1938                 }
1939
1940                 vi->affinity_hint_set = false;
1941         }
1942 }
1943
1944 static void virtnet_set_affinity(struct virtnet_info *vi)
1945 {
1946         cpumask_var_t mask;
1947         int stragglers;
1948         int group_size;
1949         int i, j, cpu;
1950         int num_cpu;
1951         int stride;
1952
1953         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
1954                 virtnet_clean_affinity(vi);
1955                 return;
1956         }
1957
1958         num_cpu = num_online_cpus();
1959         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
1960         stragglers = num_cpu >= vi->curr_queue_pairs ?
1961                         num_cpu % vi->curr_queue_pairs :
1962                         0;
1963         cpu = cpumask_next(-1, cpu_online_mask);
1964
1965         for (i = 0; i < vi->curr_queue_pairs; i++) {
1966                 group_size = stride + (i < stragglers ? 1 : 0);
1967
1968                 for (j = 0; j < group_size; j++) {
1969                         cpumask_set_cpu(cpu, mask);
1970                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
1971                                                 nr_cpu_ids, false);
1972                 }
1973                 virtqueue_set_affinity(vi->rq[i].vq, mask);
1974                 virtqueue_set_affinity(vi->sq[i].vq, mask);
1975                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, false);
1976                 cpumask_clear(mask);
1977         }
1978
1979         vi->affinity_hint_set = true;
1980         free_cpumask_var(mask);
1981 }
1982
1983 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1984 {
1985         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1986                                                    node);
1987         virtnet_set_affinity(vi);
1988         return 0;
1989 }
1990
1991 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1992 {
1993         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1994                                                    node_dead);
1995         virtnet_set_affinity(vi);
1996         return 0;
1997 }
1998
1999 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2000 {
2001         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2002                                                    node);
2003
2004         virtnet_clean_affinity(vi);
2005         return 0;
2006 }
2007
2008 static enum cpuhp_state virtionet_online;
2009
2010 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2011 {
2012         int ret;
2013
2014         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2015         if (ret)
2016                 return ret;
2017         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2018                                                &vi->node_dead);
2019         if (!ret)
2020                 return ret;
2021         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2022         return ret;
2023 }
2024
2025 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2026 {
2027         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2028         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2029                                             &vi->node_dead);
2030 }
2031
2032 static void virtnet_get_ringparam(struct net_device *dev,
2033                                 struct ethtool_ringparam *ring)
2034 {
2035         struct virtnet_info *vi = netdev_priv(dev);
2036
2037         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2038         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2039         ring->rx_pending = ring->rx_max_pending;
2040         ring->tx_pending = ring->tx_max_pending;
2041 }
2042
2043
2044 static void virtnet_get_drvinfo(struct net_device *dev,
2045                                 struct ethtool_drvinfo *info)
2046 {
2047         struct virtnet_info *vi = netdev_priv(dev);
2048         struct virtio_device *vdev = vi->vdev;
2049
2050         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2051         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2052         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2053
2054 }
2055
2056 /* TODO: Eliminate OOO packets during switching */
2057 static int virtnet_set_channels(struct net_device *dev,
2058                                 struct ethtool_channels *channels)
2059 {
2060         struct virtnet_info *vi = netdev_priv(dev);
2061         u16 queue_pairs = channels->combined_count;
2062         int err;
2063
2064         /* We don't support separate rx/tx channels.
2065          * We don't allow setting 'other' channels.
2066          */
2067         if (channels->rx_count || channels->tx_count || channels->other_count)
2068                 return -EINVAL;
2069
2070         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2071                 return -EINVAL;
2072
2073         /* For now we don't support modifying channels while XDP is loaded
2074          * also when XDP is loaded all RX queues have XDP programs so we only
2075          * need to check a single RX queue.
2076          */
2077         if (vi->rq[0].xdp_prog)
2078                 return -EINVAL;
2079
2080         get_online_cpus();
2081         err = _virtnet_set_queues(vi, queue_pairs);
2082         if (!err) {
2083                 netif_set_real_num_tx_queues(dev, queue_pairs);
2084                 netif_set_real_num_rx_queues(dev, queue_pairs);
2085
2086                 virtnet_set_affinity(vi);
2087         }
2088         put_online_cpus();
2089
2090         return err;
2091 }
2092
2093 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2094 {
2095         struct virtnet_info *vi = netdev_priv(dev);
2096         char *p = (char *)data;
2097         unsigned int i, j;
2098
2099         switch (stringset) {
2100         case ETH_SS_STATS:
2101                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2102                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2103                                 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
2104                                          i, virtnet_rq_stats_desc[j].desc);
2105                                 p += ETH_GSTRING_LEN;
2106                         }
2107                 }
2108
2109                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2110                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2111                                 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
2112                                          i, virtnet_sq_stats_desc[j].desc);
2113                                 p += ETH_GSTRING_LEN;
2114                         }
2115                 }
2116                 break;
2117         }
2118 }
2119
2120 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2121 {
2122         struct virtnet_info *vi = netdev_priv(dev);
2123
2124         switch (sset) {
2125         case ETH_SS_STATS:
2126                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2127                                                VIRTNET_SQ_STATS_LEN);
2128         default:
2129                 return -EOPNOTSUPP;
2130         }
2131 }
2132
2133 static void virtnet_get_ethtool_stats(struct net_device *dev,
2134                                       struct ethtool_stats *stats, u64 *data)
2135 {
2136         struct virtnet_info *vi = netdev_priv(dev);
2137         unsigned int idx = 0, start, i, j;
2138         const u8 *stats_base;
2139         size_t offset;
2140
2141         for (i = 0; i < vi->curr_queue_pairs; i++) {
2142                 struct receive_queue *rq = &vi->rq[i];
2143
2144                 stats_base = (u8 *)&rq->stats;
2145                 do {
2146                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2147                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2148                                 offset = virtnet_rq_stats_desc[j].offset;
2149                                 data[idx + j] = *(u64 *)(stats_base + offset);
2150                         }
2151                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2152                 idx += VIRTNET_RQ_STATS_LEN;
2153         }
2154
2155         for (i = 0; i < vi->curr_queue_pairs; i++) {
2156                 struct send_queue *sq = &vi->sq[i];
2157
2158                 stats_base = (u8 *)&sq->stats;
2159                 do {
2160                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2161                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2162                                 offset = virtnet_sq_stats_desc[j].offset;
2163                                 data[idx + j] = *(u64 *)(stats_base + offset);
2164                         }
2165                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2166                 idx += VIRTNET_SQ_STATS_LEN;
2167         }
2168 }
2169
2170 static void virtnet_get_channels(struct net_device *dev,
2171                                  struct ethtool_channels *channels)
2172 {
2173         struct virtnet_info *vi = netdev_priv(dev);
2174
2175         channels->combined_count = vi->curr_queue_pairs;
2176         channels->max_combined = vi->max_queue_pairs;
2177         channels->max_other = 0;
2178         channels->rx_count = 0;
2179         channels->tx_count = 0;
2180         channels->other_count = 0;
2181 }
2182
2183 static int virtnet_set_link_ksettings(struct net_device *dev,
2184                                       const struct ethtool_link_ksettings *cmd)
2185 {
2186         struct virtnet_info *vi = netdev_priv(dev);
2187
2188         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2189                                                   &vi->speed, &vi->duplex);
2190 }
2191
2192 static int virtnet_get_link_ksettings(struct net_device *dev,
2193                                       struct ethtool_link_ksettings *cmd)
2194 {
2195         struct virtnet_info *vi = netdev_priv(dev);
2196
2197         cmd->base.speed = vi->speed;
2198         cmd->base.duplex = vi->duplex;
2199         cmd->base.port = PORT_OTHER;
2200
2201         return 0;
2202 }
2203
2204 static int virtnet_set_coalesce(struct net_device *dev,
2205                                 struct ethtool_coalesce *ec)
2206 {
2207         struct virtnet_info *vi = netdev_priv(dev);
2208         int i, napi_weight;
2209
2210         if (ec->tx_max_coalesced_frames > 1 ||
2211             ec->rx_max_coalesced_frames != 1)
2212                 return -EINVAL;
2213
2214         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2215         if (napi_weight ^ vi->sq[0].napi.weight) {
2216                 if (dev->flags & IFF_UP)
2217                         return -EBUSY;
2218                 for (i = 0; i < vi->max_queue_pairs; i++)
2219                         vi->sq[i].napi.weight = napi_weight;
2220         }
2221
2222         return 0;
2223 }
2224
2225 static int virtnet_get_coalesce(struct net_device *dev,
2226                                 struct ethtool_coalesce *ec)
2227 {
2228         struct ethtool_coalesce ec_default = {
2229                 .cmd = ETHTOOL_GCOALESCE,
2230                 .rx_max_coalesced_frames = 1,
2231         };
2232         struct virtnet_info *vi = netdev_priv(dev);
2233
2234         memcpy(ec, &ec_default, sizeof(ec_default));
2235
2236         if (vi->sq[0].napi.weight)
2237                 ec->tx_max_coalesced_frames = 1;
2238
2239         return 0;
2240 }
2241
2242 static void virtnet_init_settings(struct net_device *dev)
2243 {
2244         struct virtnet_info *vi = netdev_priv(dev);
2245
2246         vi->speed = SPEED_UNKNOWN;
2247         vi->duplex = DUPLEX_UNKNOWN;
2248 }
2249
2250 static void virtnet_update_settings(struct virtnet_info *vi)
2251 {
2252         u32 speed;
2253         u8 duplex;
2254
2255         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2256                 return;
2257
2258         speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2259                                                   speed));
2260         if (ethtool_validate_speed(speed))
2261                 vi->speed = speed;
2262         duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2263                                                   duplex));
2264         if (ethtool_validate_duplex(duplex))
2265                 vi->duplex = duplex;
2266 }
2267
2268 static const struct ethtool_ops virtnet_ethtool_ops = {
2269         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2270         .get_drvinfo = virtnet_get_drvinfo,
2271         .get_link = ethtool_op_get_link,
2272         .get_ringparam = virtnet_get_ringparam,
2273         .get_strings = virtnet_get_strings,
2274         .get_sset_count = virtnet_get_sset_count,
2275         .get_ethtool_stats = virtnet_get_ethtool_stats,
2276         .set_channels = virtnet_set_channels,
2277         .get_channels = virtnet_get_channels,
2278         .get_ts_info = ethtool_op_get_ts_info,
2279         .get_link_ksettings = virtnet_get_link_ksettings,
2280         .set_link_ksettings = virtnet_set_link_ksettings,
2281         .set_coalesce = virtnet_set_coalesce,
2282         .get_coalesce = virtnet_get_coalesce,
2283 };
2284
2285 static void virtnet_freeze_down(struct virtio_device *vdev)
2286 {
2287         struct virtnet_info *vi = vdev->priv;
2288         int i;
2289
2290         /* Make sure no work handler is accessing the device */
2291         flush_work(&vi->config_work);
2292
2293         netif_tx_lock_bh(vi->dev);
2294         netif_device_detach(vi->dev);
2295         netif_tx_unlock_bh(vi->dev);
2296         cancel_delayed_work_sync(&vi->refill);
2297
2298         if (netif_running(vi->dev)) {
2299                 for (i = 0; i < vi->max_queue_pairs; i++) {
2300                         napi_disable(&vi->rq[i].napi);
2301                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2302                 }
2303         }
2304 }
2305
2306 static int init_vqs(struct virtnet_info *vi);
2307
2308 static int virtnet_restore_up(struct virtio_device *vdev)
2309 {
2310         struct virtnet_info *vi = vdev->priv;
2311         int err, i;
2312
2313         err = init_vqs(vi);
2314         if (err)
2315                 return err;
2316
2317         virtio_device_ready(vdev);
2318
2319         if (netif_running(vi->dev)) {
2320                 for (i = 0; i < vi->curr_queue_pairs; i++)
2321                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2322                                 schedule_delayed_work(&vi->refill, 0);
2323
2324                 for (i = 0; i < vi->max_queue_pairs; i++) {
2325                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2326                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2327                                                &vi->sq[i].napi);
2328                 }
2329         }
2330
2331         netif_tx_lock_bh(vi->dev);
2332         netif_device_attach(vi->dev);
2333         netif_tx_unlock_bh(vi->dev);
2334         return err;
2335 }
2336
2337 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2338 {
2339         struct scatterlist sg;
2340         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2341
2342         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2343
2344         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2345                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2346                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2347                 return -EINVAL;
2348         }
2349
2350         return 0;
2351 }
2352
2353 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2354 {
2355         u64 offloads = 0;
2356
2357         if (!vi->guest_offloads)
2358                 return 0;
2359
2360         return virtnet_set_guest_offloads(vi, offloads);
2361 }
2362
2363 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2364 {
2365         u64 offloads = vi->guest_offloads;
2366
2367         if (!vi->guest_offloads)
2368                 return 0;
2369
2370         return virtnet_set_guest_offloads(vi, offloads);
2371 }
2372
2373 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2374                            struct netlink_ext_ack *extack)
2375 {
2376         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2377         struct virtnet_info *vi = netdev_priv(dev);
2378         struct bpf_prog *old_prog;
2379         u16 xdp_qp = 0, curr_qp;
2380         int i, err;
2381
2382         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2383             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2384                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2385                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2386                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2387                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2388                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2389                 return -EOPNOTSUPP;
2390         }
2391
2392         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2393                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2394                 return -EINVAL;
2395         }
2396
2397         if (dev->mtu > max_sz) {
2398                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2399                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2400                 return -EINVAL;
2401         }
2402
2403         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2404         if (prog)
2405                 xdp_qp = nr_cpu_ids;
2406
2407         /* XDP requires extra queues for XDP_TX */
2408         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2409                 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2410                 netdev_warn(dev, "request %i queues but max is %i\n",
2411                             curr_qp + xdp_qp, vi->max_queue_pairs);
2412                 return -ENOMEM;
2413         }
2414
2415         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2416         if (!prog && !old_prog)
2417                 return 0;
2418
2419         if (prog)
2420                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2421
2422         /* Make sure NAPI is not using any XDP TX queues for RX. */
2423         if (netif_running(dev)) {
2424                 for (i = 0; i < vi->max_queue_pairs; i++) {
2425                         napi_disable(&vi->rq[i].napi);
2426                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2427                 }
2428         }
2429
2430         if (!prog) {
2431                 for (i = 0; i < vi->max_queue_pairs; i++) {
2432                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2433                         if (i == 0)
2434                                 virtnet_restore_guest_offloads(vi);
2435                 }
2436                 synchronize_net();
2437         }
2438
2439         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2440         if (err)
2441                 goto err;
2442         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2443         vi->xdp_queue_pairs = xdp_qp;
2444
2445         if (prog) {
2446                 for (i = 0; i < vi->max_queue_pairs; i++) {
2447                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2448                         if (i == 0 && !old_prog)
2449                                 virtnet_clear_guest_offloads(vi);
2450                 }
2451         }
2452
2453         for (i = 0; i < vi->max_queue_pairs; i++) {
2454                 if (old_prog)
2455                         bpf_prog_put(old_prog);
2456                 if (netif_running(dev)) {
2457                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2458                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2459                                                &vi->sq[i].napi);
2460                 }
2461         }
2462
2463         return 0;
2464
2465 err:
2466         if (!prog) {
2467                 virtnet_clear_guest_offloads(vi);
2468                 for (i = 0; i < vi->max_queue_pairs; i++)
2469                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2470         }
2471
2472         if (netif_running(dev)) {
2473                 for (i = 0; i < vi->max_queue_pairs; i++) {
2474                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2475                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2476                                                &vi->sq[i].napi);
2477                 }
2478         }
2479         if (prog)
2480                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2481         return err;
2482 }
2483
2484 static u32 virtnet_xdp_query(struct net_device *dev)
2485 {
2486         struct virtnet_info *vi = netdev_priv(dev);
2487         const struct bpf_prog *xdp_prog;
2488         int i;
2489
2490         for (i = 0; i < vi->max_queue_pairs; i++) {
2491                 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2492                 if (xdp_prog)
2493                         return xdp_prog->aux->id;
2494         }
2495         return 0;
2496 }
2497
2498 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2499 {
2500         switch (xdp->command) {
2501         case XDP_SETUP_PROG:
2502                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2503         case XDP_QUERY_PROG:
2504                 xdp->prog_id = virtnet_xdp_query(dev);
2505                 return 0;
2506         default:
2507                 return -EINVAL;
2508         }
2509 }
2510
2511 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2512                                       size_t len)
2513 {
2514         struct virtnet_info *vi = netdev_priv(dev);
2515         int ret;
2516
2517         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2518                 return -EOPNOTSUPP;
2519
2520         ret = snprintf(buf, len, "sby");
2521         if (ret >= len)
2522                 return -EOPNOTSUPP;
2523
2524         return 0;
2525 }
2526
2527 static int virtnet_set_features(struct net_device *dev,
2528                                 netdev_features_t features)
2529 {
2530         struct virtnet_info *vi = netdev_priv(dev);
2531         u64 offloads;
2532         int err;
2533
2534         if ((dev->features ^ features) & NETIF_F_LRO) {
2535                 if (vi->xdp_queue_pairs)
2536                         return -EBUSY;
2537
2538                 if (features & NETIF_F_LRO)
2539                         offloads = vi->guest_offloads_capable;
2540                 else
2541                         offloads = 0;
2542
2543                 err = virtnet_set_guest_offloads(vi, offloads);
2544                 if (err)
2545                         return err;
2546                 vi->guest_offloads = offloads;
2547         }
2548
2549         return 0;
2550 }
2551
2552 static const struct net_device_ops virtnet_netdev = {
2553         .ndo_open            = virtnet_open,
2554         .ndo_stop            = virtnet_close,
2555         .ndo_start_xmit      = start_xmit,
2556         .ndo_validate_addr   = eth_validate_addr,
2557         .ndo_set_mac_address = virtnet_set_mac_address,
2558         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2559         .ndo_get_stats64     = virtnet_stats,
2560         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2561         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2562         .ndo_bpf                = virtnet_xdp,
2563         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2564         .ndo_features_check     = passthru_features_check,
2565         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2566         .ndo_set_features       = virtnet_set_features,
2567 };
2568
2569 static void virtnet_config_changed_work(struct work_struct *work)
2570 {
2571         struct virtnet_info *vi =
2572                 container_of(work, struct virtnet_info, config_work);
2573         u16 v;
2574
2575         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2576                                  struct virtio_net_config, status, &v) < 0)
2577                 return;
2578
2579         if (v & VIRTIO_NET_S_ANNOUNCE) {
2580                 netdev_notify_peers(vi->dev);
2581                 virtnet_ack_link_announce(vi);
2582         }
2583
2584         /* Ignore unknown (future) status bits */
2585         v &= VIRTIO_NET_S_LINK_UP;
2586
2587         if (vi->status == v)
2588                 return;
2589
2590         vi->status = v;
2591
2592         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2593                 virtnet_update_settings(vi);
2594                 netif_carrier_on(vi->dev);
2595                 netif_tx_wake_all_queues(vi->dev);
2596         } else {
2597                 netif_carrier_off(vi->dev);
2598                 netif_tx_stop_all_queues(vi->dev);
2599         }
2600 }
2601
2602 static void virtnet_config_changed(struct virtio_device *vdev)
2603 {
2604         struct virtnet_info *vi = vdev->priv;
2605
2606         schedule_work(&vi->config_work);
2607 }
2608
2609 static void virtnet_free_queues(struct virtnet_info *vi)
2610 {
2611         int i;
2612
2613         for (i = 0; i < vi->max_queue_pairs; i++) {
2614                 napi_hash_del(&vi->rq[i].napi);
2615                 netif_napi_del(&vi->rq[i].napi);
2616                 netif_napi_del(&vi->sq[i].napi);
2617         }
2618
2619         /* We called napi_hash_del() before netif_napi_del(),
2620          * we need to respect an RCU grace period before freeing vi->rq
2621          */
2622         synchronize_net();
2623
2624         kfree(vi->rq);
2625         kfree(vi->sq);
2626         kfree(vi->ctrl);
2627 }
2628
2629 static void _free_receive_bufs(struct virtnet_info *vi)
2630 {
2631         struct bpf_prog *old_prog;
2632         int i;
2633
2634         for (i = 0; i < vi->max_queue_pairs; i++) {
2635                 while (vi->rq[i].pages)
2636                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2637
2638                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2639                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2640                 if (old_prog)
2641                         bpf_prog_put(old_prog);
2642         }
2643 }
2644
2645 static void free_receive_bufs(struct virtnet_info *vi)
2646 {
2647         rtnl_lock();
2648         _free_receive_bufs(vi);
2649         rtnl_unlock();
2650 }
2651
2652 static void free_receive_page_frags(struct virtnet_info *vi)
2653 {
2654         int i;
2655         for (i = 0; i < vi->max_queue_pairs; i++)
2656                 if (vi->rq[i].alloc_frag.page)
2657                         put_page(vi->rq[i].alloc_frag.page);
2658 }
2659
2660 static void free_unused_bufs(struct virtnet_info *vi)
2661 {
2662         void *buf;
2663         int i;
2664
2665         for (i = 0; i < vi->max_queue_pairs; i++) {
2666                 struct virtqueue *vq = vi->sq[i].vq;
2667                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2668                         if (!is_xdp_frame(buf))
2669                                 dev_kfree_skb(buf);
2670                         else
2671                                 xdp_return_frame(ptr_to_xdp(buf));
2672                 }
2673         }
2674
2675         for (i = 0; i < vi->max_queue_pairs; i++) {
2676                 struct virtqueue *vq = vi->rq[i].vq;
2677
2678                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2679                         if (vi->mergeable_rx_bufs) {
2680                                 put_page(virt_to_head_page(buf));
2681                         } else if (vi->big_packets) {
2682                                 give_pages(&vi->rq[i], buf);
2683                         } else {
2684                                 put_page(virt_to_head_page(buf));
2685                         }
2686                 }
2687         }
2688 }
2689
2690 static void virtnet_del_vqs(struct virtnet_info *vi)
2691 {
2692         struct virtio_device *vdev = vi->vdev;
2693
2694         virtnet_clean_affinity(vi);
2695
2696         vdev->config->del_vqs(vdev);
2697
2698         virtnet_free_queues(vi);
2699 }
2700
2701 /* How large should a single buffer be so a queue full of these can fit at
2702  * least one full packet?
2703  * Logic below assumes the mergeable buffer header is used.
2704  */
2705 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2706 {
2707         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2708         unsigned int rq_size = virtqueue_get_vring_size(vq);
2709         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2710         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2711         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2712
2713         return max(max(min_buf_len, hdr_len) - hdr_len,
2714                    (unsigned int)GOOD_PACKET_LEN);
2715 }
2716
2717 static int virtnet_find_vqs(struct virtnet_info *vi)
2718 {
2719         vq_callback_t **callbacks;
2720         struct virtqueue **vqs;
2721         int ret = -ENOMEM;
2722         int i, total_vqs;
2723         const char **names;
2724         bool *ctx;
2725
2726         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2727          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2728          * possible control vq.
2729          */
2730         total_vqs = vi->max_queue_pairs * 2 +
2731                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2732
2733         /* Allocate space for find_vqs parameters */
2734         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2735         if (!vqs)
2736                 goto err_vq;
2737         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2738         if (!callbacks)
2739                 goto err_callback;
2740         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2741         if (!names)
2742                 goto err_names;
2743         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2744                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2745                 if (!ctx)
2746                         goto err_ctx;
2747         } else {
2748                 ctx = NULL;
2749         }
2750
2751         /* Parameters for control virtqueue, if any */
2752         if (vi->has_cvq) {
2753                 callbacks[total_vqs - 1] = NULL;
2754                 names[total_vqs - 1] = "control";
2755         }
2756
2757         /* Allocate/initialize parameters for send/receive virtqueues */
2758         for (i = 0; i < vi->max_queue_pairs; i++) {
2759                 callbacks[rxq2vq(i)] = skb_recv_done;
2760                 callbacks[txq2vq(i)] = skb_xmit_done;
2761                 sprintf(vi->rq[i].name, "input.%d", i);
2762                 sprintf(vi->sq[i].name, "output.%d", i);
2763                 names[rxq2vq(i)] = vi->rq[i].name;
2764                 names[txq2vq(i)] = vi->sq[i].name;
2765                 if (ctx)
2766                         ctx[rxq2vq(i)] = true;
2767         }
2768
2769         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2770                                          names, ctx, NULL);
2771         if (ret)
2772                 goto err_find;
2773
2774         if (vi->has_cvq) {
2775                 vi->cvq = vqs[total_vqs - 1];
2776                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2777                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2778         }
2779
2780         for (i = 0; i < vi->max_queue_pairs; i++) {
2781                 vi->rq[i].vq = vqs[rxq2vq(i)];
2782                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2783                 vi->sq[i].vq = vqs[txq2vq(i)];
2784         }
2785
2786         /* run here: ret == 0. */
2787
2788
2789 err_find:
2790         kfree(ctx);
2791 err_ctx:
2792         kfree(names);
2793 err_names:
2794         kfree(callbacks);
2795 err_callback:
2796         kfree(vqs);
2797 err_vq:
2798         return ret;
2799 }
2800
2801 static int virtnet_alloc_queues(struct virtnet_info *vi)
2802 {
2803         int i;
2804
2805         vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2806         if (!vi->ctrl)
2807                 goto err_ctrl;
2808         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2809         if (!vi->sq)
2810                 goto err_sq;
2811         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2812         if (!vi->rq)
2813                 goto err_rq;
2814
2815         INIT_DELAYED_WORK(&vi->refill, refill_work);
2816         for (i = 0; i < vi->max_queue_pairs; i++) {
2817                 vi->rq[i].pages = NULL;
2818                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2819                                napi_weight);
2820                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2821                                   napi_tx ? napi_weight : 0);
2822
2823                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2824                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2825                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2826
2827                 u64_stats_init(&vi->rq[i].stats.syncp);
2828                 u64_stats_init(&vi->sq[i].stats.syncp);
2829         }
2830
2831         return 0;
2832
2833 err_rq:
2834         kfree(vi->sq);
2835 err_sq:
2836         kfree(vi->ctrl);
2837 err_ctrl:
2838         return -ENOMEM;
2839 }
2840
2841 static int init_vqs(struct virtnet_info *vi)
2842 {
2843         int ret;
2844
2845         /* Allocate send & receive queues */
2846         ret = virtnet_alloc_queues(vi);
2847         if (ret)
2848                 goto err;
2849
2850         ret = virtnet_find_vqs(vi);
2851         if (ret)
2852                 goto err_free;
2853
2854         get_online_cpus();
2855         virtnet_set_affinity(vi);
2856         put_online_cpus();
2857
2858         return 0;
2859
2860 err_free:
2861         virtnet_free_queues(vi);
2862 err:
2863         return ret;
2864 }
2865
2866 #ifdef CONFIG_SYSFS
2867 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2868                 char *buf)
2869 {
2870         struct virtnet_info *vi = netdev_priv(queue->dev);
2871         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2872         unsigned int headroom = virtnet_get_headroom(vi);
2873         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2874         struct ewma_pkt_len *avg;
2875
2876         BUG_ON(queue_index >= vi->max_queue_pairs);
2877         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2878         return sprintf(buf, "%u\n",
2879                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
2880                                        SKB_DATA_ALIGN(headroom + tailroom)));
2881 }
2882
2883 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2884         __ATTR_RO(mergeable_rx_buffer_size);
2885
2886 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2887         &mergeable_rx_buffer_size_attribute.attr,
2888         NULL
2889 };
2890
2891 static const struct attribute_group virtio_net_mrg_rx_group = {
2892         .name = "virtio_net",
2893         .attrs = virtio_net_mrg_rx_attrs
2894 };
2895 #endif
2896
2897 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2898                                     unsigned int fbit,
2899                                     const char *fname, const char *dname)
2900 {
2901         if (!virtio_has_feature(vdev, fbit))
2902                 return false;
2903
2904         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2905                 fname, dname);
2906
2907         return true;
2908 }
2909
2910 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2911         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2912
2913 static bool virtnet_validate_features(struct virtio_device *vdev)
2914 {
2915         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2916             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2917                              "VIRTIO_NET_F_CTRL_VQ") ||
2918              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2919                              "VIRTIO_NET_F_CTRL_VQ") ||
2920              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2921                              "VIRTIO_NET_F_CTRL_VQ") ||
2922              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2923              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2924                              "VIRTIO_NET_F_CTRL_VQ"))) {
2925                 return false;
2926         }
2927
2928         return true;
2929 }
2930
2931 #define MIN_MTU ETH_MIN_MTU
2932 #define MAX_MTU ETH_MAX_MTU
2933
2934 static int virtnet_validate(struct virtio_device *vdev)
2935 {
2936         if (!vdev->config->get) {
2937                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2938                         __func__);
2939                 return -EINVAL;
2940         }
2941
2942         if (!virtnet_validate_features(vdev))
2943                 return -EINVAL;
2944
2945         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2946                 int mtu = virtio_cread16(vdev,
2947                                          offsetof(struct virtio_net_config,
2948                                                   mtu));
2949                 if (mtu < MIN_MTU)
2950                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2951         }
2952
2953         return 0;
2954 }
2955
2956 static int virtnet_probe(struct virtio_device *vdev)
2957 {
2958         int i, err = -ENOMEM;
2959         struct net_device *dev;
2960         struct virtnet_info *vi;
2961         u16 max_queue_pairs;
2962         int mtu;
2963
2964         /* Find if host supports multiqueue virtio_net device */
2965         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2966                                    struct virtio_net_config,
2967                                    max_virtqueue_pairs, &max_queue_pairs);
2968
2969         /* We need at least 2 queue's */
2970         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2971             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2972             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2973                 max_queue_pairs = 1;
2974
2975         /* Allocate ourselves a network device with room for our info */
2976         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2977         if (!dev)
2978                 return -ENOMEM;
2979
2980         /* Set up network device as normal. */
2981         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2982         dev->netdev_ops = &virtnet_netdev;
2983         dev->features = NETIF_F_HIGHDMA;
2984
2985         dev->ethtool_ops = &virtnet_ethtool_ops;
2986         SET_NETDEV_DEV(dev, &vdev->dev);
2987
2988         /* Do we support "hardware" checksums? */
2989         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2990                 /* This opens up the world of extra features. */
2991                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2992                 if (csum)
2993                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2994
2995                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2996                         dev->hw_features |= NETIF_F_TSO
2997                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2998                 }
2999                 /* Individual feature bits: what can host handle? */
3000                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3001                         dev->hw_features |= NETIF_F_TSO;
3002                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3003                         dev->hw_features |= NETIF_F_TSO6;
3004                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3005                         dev->hw_features |= NETIF_F_TSO_ECN;
3006
3007                 dev->features |= NETIF_F_GSO_ROBUST;
3008
3009                 if (gso)
3010                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3011                 /* (!csum && gso) case will be fixed by register_netdev() */
3012         }
3013         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3014                 dev->features |= NETIF_F_RXCSUM;
3015         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3016             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3017                 dev->features |= NETIF_F_LRO;
3018         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3019                 dev->hw_features |= NETIF_F_LRO;
3020
3021         dev->vlan_features = dev->features;
3022
3023         /* MTU range: 68 - 65535 */
3024         dev->min_mtu = MIN_MTU;
3025         dev->max_mtu = MAX_MTU;
3026
3027         /* Configuration may specify what MAC to use.  Otherwise random. */
3028         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3029                 virtio_cread_bytes(vdev,
3030                                    offsetof(struct virtio_net_config, mac),
3031                                    dev->dev_addr, dev->addr_len);
3032         else
3033                 eth_hw_addr_random(dev);
3034
3035         /* Set up our device-specific information */
3036         vi = netdev_priv(dev);
3037         vi->dev = dev;
3038         vi->vdev = vdev;
3039         vdev->priv = vi;
3040
3041         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3042
3043         /* If we can receive ANY GSO packets, we must allocate large ones. */
3044         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3045             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3046             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3047             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3048                 vi->big_packets = true;
3049
3050         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3051                 vi->mergeable_rx_bufs = true;
3052
3053         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3054             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3055                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3056         else
3057                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3058
3059         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3060             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3061                 vi->any_header_sg = true;
3062
3063         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3064                 vi->has_cvq = true;
3065
3066         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3067                 mtu = virtio_cread16(vdev,
3068                                      offsetof(struct virtio_net_config,
3069                                               mtu));
3070                 if (mtu < dev->min_mtu) {
3071                         /* Should never trigger: MTU was previously validated
3072                          * in virtnet_validate.
3073                          */
3074                         dev_err(&vdev->dev,
3075                                 "device MTU appears to have changed it is now %d < %d",
3076                                 mtu, dev->min_mtu);
3077                         goto free;
3078                 }
3079
3080                 dev->mtu = mtu;
3081                 dev->max_mtu = mtu;
3082
3083                 /* TODO: size buffers correctly in this case. */
3084                 if (dev->mtu > ETH_DATA_LEN)
3085                         vi->big_packets = true;
3086         }
3087
3088         if (vi->any_header_sg)
3089                 dev->needed_headroom = vi->hdr_len;
3090
3091         /* Enable multiqueue by default */
3092         if (num_online_cpus() >= max_queue_pairs)
3093                 vi->curr_queue_pairs = max_queue_pairs;
3094         else
3095                 vi->curr_queue_pairs = num_online_cpus();
3096         vi->max_queue_pairs = max_queue_pairs;
3097
3098         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3099         err = init_vqs(vi);
3100         if (err)
3101                 goto free;
3102
3103 #ifdef CONFIG_SYSFS
3104         if (vi->mergeable_rx_bufs)
3105                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3106 #endif
3107         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3108         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3109
3110         virtnet_init_settings(dev);
3111
3112         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3113                 vi->failover = net_failover_create(vi->dev);
3114                 if (IS_ERR(vi->failover)) {
3115                         err = PTR_ERR(vi->failover);
3116                         goto free_vqs;
3117                 }
3118         }
3119
3120         err = register_netdev(dev);
3121         if (err) {
3122                 pr_debug("virtio_net: registering device failed\n");
3123                 goto free_failover;
3124         }
3125
3126         virtio_device_ready(vdev);
3127
3128         err = virtnet_cpu_notif_add(vi);
3129         if (err) {
3130                 pr_debug("virtio_net: registering cpu notifier failed\n");
3131                 goto free_unregister_netdev;
3132         }
3133
3134         virtnet_set_queues(vi, vi->curr_queue_pairs);
3135
3136         /* Assume link up if device can't report link status,
3137            otherwise get link status from config. */
3138         netif_carrier_off(dev);
3139         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3140                 schedule_work(&vi->config_work);
3141         } else {
3142                 vi->status = VIRTIO_NET_S_LINK_UP;
3143                 virtnet_update_settings(vi);
3144                 netif_carrier_on(dev);
3145         }
3146
3147         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3148                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3149                         set_bit(guest_offloads[i], &vi->guest_offloads);
3150         vi->guest_offloads_capable = vi->guest_offloads;
3151
3152         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3153                  dev->name, max_queue_pairs);
3154
3155         return 0;
3156
3157 free_unregister_netdev:
3158         vi->vdev->config->reset(vdev);
3159
3160         unregister_netdev(dev);
3161 free_failover:
3162         net_failover_destroy(vi->failover);
3163 free_vqs:
3164         cancel_delayed_work_sync(&vi->refill);
3165         free_receive_page_frags(vi);
3166         virtnet_del_vqs(vi);
3167 free:
3168         free_netdev(dev);
3169         return err;
3170 }
3171
3172 static void remove_vq_common(struct virtnet_info *vi)
3173 {
3174         vi->vdev->config->reset(vi->vdev);
3175
3176         /* Free unused buffers in both send and recv, if any. */
3177         free_unused_bufs(vi);
3178
3179         free_receive_bufs(vi);
3180
3181         free_receive_page_frags(vi);
3182
3183         virtnet_del_vqs(vi);
3184 }
3185
3186 static void virtnet_remove(struct virtio_device *vdev)
3187 {
3188         struct virtnet_info *vi = vdev->priv;
3189
3190         virtnet_cpu_notif_remove(vi);
3191
3192         /* Make sure no work handler is accessing the device. */
3193         flush_work(&vi->config_work);
3194
3195         unregister_netdev(vi->dev);
3196
3197         net_failover_destroy(vi->failover);
3198
3199         remove_vq_common(vi);
3200
3201         free_netdev(vi->dev);
3202 }
3203
3204 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3205 {
3206         struct virtnet_info *vi = vdev->priv;
3207
3208         virtnet_cpu_notif_remove(vi);
3209         virtnet_freeze_down(vdev);
3210         remove_vq_common(vi);
3211
3212         return 0;
3213 }
3214
3215 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3216 {
3217         struct virtnet_info *vi = vdev->priv;
3218         int err;
3219
3220         err = virtnet_restore_up(vdev);
3221         if (err)
3222                 return err;
3223         virtnet_set_queues(vi, vi->curr_queue_pairs);
3224
3225         err = virtnet_cpu_notif_add(vi);
3226         if (err)
3227                 return err;
3228
3229         return 0;
3230 }
3231
3232 static struct virtio_device_id id_table[] = {
3233         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3234         { 0 },
3235 };
3236
3237 #define VIRTNET_FEATURES \
3238         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3239         VIRTIO_NET_F_MAC, \
3240         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3241         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3242         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3243         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3244         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3245         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3246         VIRTIO_NET_F_CTRL_MAC_ADDR, \
3247         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3248         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3249
3250 static unsigned int features[] = {
3251         VIRTNET_FEATURES,
3252 };
3253
3254 static unsigned int features_legacy[] = {
3255         VIRTNET_FEATURES,
3256         VIRTIO_NET_F_GSO,
3257         VIRTIO_F_ANY_LAYOUT,
3258 };
3259
3260 static struct virtio_driver virtio_net_driver = {
3261         .feature_table = features,
3262         .feature_table_size = ARRAY_SIZE(features),
3263         .feature_table_legacy = features_legacy,
3264         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3265         .driver.name =  KBUILD_MODNAME,
3266         .driver.owner = THIS_MODULE,
3267         .id_table =     id_table,
3268         .validate =     virtnet_validate,
3269         .probe =        virtnet_probe,
3270         .remove =       virtnet_remove,
3271         .config_changed = virtnet_config_changed,
3272 #ifdef CONFIG_PM_SLEEP
3273         .freeze =       virtnet_freeze,
3274         .restore =      virtnet_restore,
3275 #endif
3276 };
3277
3278 static __init int virtio_net_driver_init(void)
3279 {
3280         int ret;
3281
3282         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3283                                       virtnet_cpu_online,
3284                                       virtnet_cpu_down_prep);
3285         if (ret < 0)
3286                 goto out;
3287         virtionet_online = ret;
3288         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3289                                       NULL, virtnet_cpu_dead);
3290         if (ret)
3291                 goto err_dead;
3292
3293         ret = register_virtio_driver(&virtio_net_driver);
3294         if (ret)
3295                 goto err_virtio;
3296         return 0;
3297 err_virtio:
3298         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3299 err_dead:
3300         cpuhp_remove_multi_state(virtionet_online);
3301 out:
3302         return ret;
3303 }
3304 module_init(virtio_net_driver_init);
3305
3306 static __exit void virtio_net_driver_exit(void)
3307 {
3308         unregister_virtio_driver(&virtio_net_driver);
3309         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3310         cpuhp_remove_multi_state(virtionet_online);
3311 }
3312 module_exit(virtio_net_driver_exit);
3313
3314 MODULE_DEVICE_TABLE(virtio, id_table);
3315 MODULE_DESCRIPTION("Virtio network driver");
3316 MODULE_LICENSE("GPL");