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