drivers/net/virtio_net.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* A network driver using virtio.
   3  *
   4  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   5  */
   6 //#define DEBUG
   7 #include <linux/netdevice.h>
   8 #include <linux/etherdevice.h>
   9 #include <linux/ethtool.h>
  10 #include <linux/module.h>
  11 #include <linux/virtio.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/bpf.h>
  14 #include <linux/bpf_trace.h>
  15 #include <linux/scatterlist.h>
  16 #include <linux/if_vlan.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/average.h>
  20 #include <linux/filter.h>
  21 #include <linux/kernel.h>
  22 #include <net/route.h>
  23 #include <net/xdp.h>
  24 #include <net/net_failover.h>
  25
  26 static int napi_weight = NAPI_POLL_WEIGHT;
  27 module_param(napi_weight, int, 0444);
  28
  29 static bool csum = true, gso = true, napi_tx = true;
  30 module_param(csum, bool, 0444);
  31 module_param(gso, bool, 0444);
  32 module_param(napi_tx, bool, 0644);
  33
  34 /* FIXME: MTU in config. */
  35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  36 #define GOOD_COPY_LEN   128
  37
  38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  39
  40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  41 #define VIRTIO_XDP_HEADROOM 256
  42
  43 /* Separating two types of XDP xmit */
  44 #define VIRTIO_XDP_TX           BIT(0)
  45 #define VIRTIO_XDP_REDIR        BIT(1)
  46
  47 #define VIRTIO_XDP_FLAG BIT(0)
  48
  49 /* RX packet size EWMA. The average packet size is used to determine the packet
  50  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  51  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  52  * term, transient changes in packet size.
  53  */
  54 DECLARE_EWMA(pkt_len, 0, 64)
  55
  56 #define VIRTNET_DRIVER_VERSION "1.0.0"
  57
  58 static const unsigned long guest_offloads[] = {
  59         VIRTIO_NET_F_GUEST_TSO4,
  60         VIRTIO_NET_F_GUEST_TSO6,
  61         VIRTIO_NET_F_GUEST_ECN,
  62         VIRTIO_NET_F_GUEST_UFO,
  63         VIRTIO_NET_F_GUEST_CSUM
  64 };
  65
  66 #define GUEST_OFFLOAD_LRO_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
  67                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
  68                                 (1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
  69                                 (1ULL << VIRTIO_NET_F_GUEST_UFO))
  70
  71 #define GUEST_OFFLOAD_CSUM_MASK (1ULL << VIRTIO_NET_F_GUEST_CSUM)
  72
  73 struct virtnet_stat_desc {
  74         char desc[ETH_GSTRING_LEN];
  75         size_t offset;
  76 };
  77
  78 struct virtnet_sq_stats {
  79         struct u64_stats_sync syncp;
  80         u64 packets;
  81         u64 bytes;
  82         u64 xdp_tx;
  83         u64 xdp_tx_drops;
  84         u64 kicks;
  85 };
  86
  87 struct virtnet_rq_stats {
  88         struct u64_stats_sync syncp;
  89         u64 packets;
  90         u64 bytes;
  91         u64 drops;
  92         u64 xdp_packets;
  93         u64 xdp_tx;
  94         u64 xdp_redirects;
  95         u64 xdp_drops;
  96         u64 kicks;
  97 };
  98
  99 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
 100 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
 101
 102 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
 103         { "packets",            VIRTNET_SQ_STAT(packets) },
 104         { "bytes",              VIRTNET_SQ_STAT(bytes) },
 105         { "xdp_tx",             VIRTNET_SQ_STAT(xdp_tx) },
 106         { "xdp_tx_drops",       VIRTNET_SQ_STAT(xdp_tx_drops) },
 107         { "kicks",              VIRTNET_SQ_STAT(kicks) },
 108 };
 109
 110 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
 111         { "packets",            VIRTNET_RQ_STAT(packets) },
 112         { "bytes",              VIRTNET_RQ_STAT(bytes) },
 113         { "drops",              VIRTNET_RQ_STAT(drops) },
 114         { "xdp_packets",        VIRTNET_RQ_STAT(xdp_packets) },
 115         { "xdp_tx",             VIRTNET_RQ_STAT(xdp_tx) },
 116         { "xdp_redirects",      VIRTNET_RQ_STAT(xdp_redirects) },
 117         { "xdp_drops",          VIRTNET_RQ_STAT(xdp_drops) },
 118         { "kicks",              VIRTNET_RQ_STAT(kicks) },
 119 };
 120
 121 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 122 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 123
 124 /* Internal representation of a send virtqueue */
 125 struct send_queue {
 126         /* Virtqueue associated with this send _queue */
 127         struct virtqueue *vq;
 128
 129         /* TX: fragments + linear part + virtio header */
 130         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 131
 132         /* Name of the send queue: output.$index */
 133         char name[40];
 134
 135         struct virtnet_sq_stats stats;
 136
 137         struct napi_struct napi;
 138 };
 139
 140 /* Internal representation of a receive virtqueue */
 141 struct receive_queue {
 142         /* Virtqueue associated with this receive_queue */
 143         struct virtqueue *vq;
 144
 145         struct napi_struct napi;
 146
 147         struct bpf_prog __rcu *xdp_prog;
 148
 149         struct virtnet_rq_stats stats;
 150
 151         /* Chain pages by the private ptr. */
 152         struct page *pages;
 153
 154         /* Average packet length for mergeable receive buffers. */
 155         struct ewma_pkt_len mrg_avg_pkt_len;
 156
 157         /* Page frag for packet buffer allocation. */
 158         struct page_frag alloc_frag;
 159
 160         /* RX: fragments + linear part + virtio header */
 161         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 162
 163         /* Min single buffer size for mergeable buffers case. */
 164         unsigned int min_buf_len;
 165
 166         /* Name of this receive queue: input.$index */
 167         char name[40];
 168
 169         struct xdp_rxq_info xdp_rxq;
 170 };
 171
 172 /* Control VQ buffers: protected by the rtnl lock */
 173 struct control_buf {
 174         struct virtio_net_ctrl_hdr hdr;
 175         virtio_net_ctrl_ack status;
 176         struct virtio_net_ctrl_mq mq;
 177         u8 promisc;
 178         u8 allmulti;
 179         __virtio16 vid;
 180         __virtio64 offloads;
 181 };
 182
 183 struct virtnet_info {
 184         struct virtio_device *vdev;
 185         struct virtqueue *cvq;
 186         struct net_device *dev;
 187         struct send_queue *sq;
 188         struct receive_queue *rq;
 189         unsigned int status;
 190
 191         /* Max # of queue pairs supported by the device */
 192         u16 max_queue_pairs;
 193
 194         /* # of queue pairs currently used by the driver */
 195         u16 curr_queue_pairs;
 196
 197         /* # of XDP queue pairs currently used by the driver */
 198         u16 xdp_queue_pairs;
 199
 200         /* I like... big packets and I cannot lie! */
 201         bool big_packets;
 202
 203         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 204         bool mergeable_rx_bufs;
 205
 206         /* Has control virtqueue */
 207         bool has_cvq;
 208
 209         /* Host can handle any s/g split between our header and packet data */
 210         bool any_header_sg;
 211
 212         /* Packet virtio header size */
 213         u8 hdr_len;
 214
 215         /* Work struct for refilling if we run low on memory. */
 216         struct delayed_work refill;
 217
 218         /* Work struct for config space updates */
 219         struct work_struct config_work;
 220
 221         /* Does the affinity hint is set for virtqueues? */
 222         bool affinity_hint_set;
 223
 224         /* CPU hotplug instances for online & dead */
 225         struct hlist_node node;
 226         struct hlist_node node_dead;
 227
 228         struct control_buf *ctrl;
 229
 230         /* Ethtool settings */
 231         u8 duplex;
 232         u32 speed;
 233
 234         unsigned long guest_offloads;
 235         unsigned long guest_offloads_capable;
 236
 237         /* failover when STANDBY feature enabled */
 238         struct failover *failover;
 239 };
 240
 241 struct padded_vnet_hdr {
 242         struct virtio_net_hdr_mrg_rxbuf hdr;
 243         /*
 244          * hdr is in a separate sg buffer, and data sg buffer shares same page
 245          * with this header sg. This padding makes next sg 16 byte aligned
 246          * after the header.
 247          */
 248         char padding[4];
 249 };
 250
 251 static bool is_xdp_frame(void *ptr)
 252 {
 253         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 254 }
 255
 256 static void *xdp_to_ptr(struct xdp_frame *ptr)
 257 {
 258         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 259 }
 260
 261 static struct xdp_frame *ptr_to_xdp(void *ptr)
 262 {
 263         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 264 }
 265
 266 /* Converting between virtqueue no. and kernel tx/rx queue no.
 267  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 268  */
 269 static int vq2txq(struct virtqueue *vq)
 270 {
 271         return (vq->index - 1) / 2;
 272 }
 273
 274 static int txq2vq(int txq)
 275 {
 276         return txq * 2 + 1;
 277 }
 278
 279 static int vq2rxq(struct virtqueue *vq)
 280 {
 281         return vq->index / 2;
 282 }
 283
 284 static int rxq2vq(int rxq)
 285 {
 286         return rxq * 2;
 287 }
 288
 289 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 290 {
 291         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 292 }
 293
 294 /*
 295  * private is used to chain pages for big packets, put the whole
 296  * most recent used list in the beginning for reuse
 297  */
 298 static void give_pages(struct receive_queue *rq, struct page *page)
 299 {
 300         struct page *end;
 301
 302         /* Find end of list, sew whole thing into vi->rq.pages. */
 303         for (end = page; end->private; end = (struct page *)end->private);
 304         end->private = (unsigned long)rq->pages;
 305         rq->pages = page;
 306 }
 307
 308 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 309 {
 310         struct page *p = rq->pages;
 311
 312         if (p) {
 313                 rq->pages = (struct page *)p->private;
 314                 /* clear private here, it is used to chain pages */
 315                 p->private = 0;
 316         } else
 317                 p = alloc_page(gfp_mask);
 318         return p;
 319 }
 320
 321 static void virtqueue_napi_schedule(struct napi_struct *napi,
 322                                     struct virtqueue *vq)
 323 {
 324         if (napi_schedule_prep(napi)) {
 325                 virtqueue_disable_cb(vq);
 326                 __napi_schedule(napi);
 327         }
 328 }
 329
 330 static void virtqueue_napi_complete(struct napi_struct *napi,
 331                                     struct virtqueue *vq, int processed)
 332 {
 333         int opaque;
 334
 335         opaque = virtqueue_enable_cb_prepare(vq);
 336         if (napi_complete_done(napi, processed)) {
 337                 if (unlikely(virtqueue_poll(vq, opaque)))
 338                         virtqueue_napi_schedule(napi, vq);
 339         } else {
 340                 virtqueue_disable_cb(vq);
 341         }
 342 }
 343
 344 static void skb_xmit_done(struct virtqueue *vq)
 345 {
 346         struct virtnet_info *vi = vq->vdev->priv;
 347         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 348
 349         /* Suppress further interrupts. */
 350         virtqueue_disable_cb(vq);
 351
 352         if (napi->weight)
 353                 virtqueue_napi_schedule(napi, vq);
 354         else
 355                 /* We were probably waiting for more output buffers. */
 356                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 357 }
 358
 359 #define MRG_CTX_HEADER_SHIFT 22
 360 static void *mergeable_len_to_ctx(unsigned int truesize,
 361                                   unsigned int headroom)
 362 {
 363         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 364 }
 365
 366 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 367 {
 368         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 369 }
 370
 371 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 372 {
 373         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 374 }
 375
 376 /* Called from bottom half context */
 377 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 378                                    struct receive_queue *rq,
 379                                    struct page *page, unsigned int offset,
 380                                    unsigned int len, unsigned int truesize,
 381                                    bool hdr_valid, unsigned int metasize)
 382 {
 383         struct sk_buff *skb;
 384         struct virtio_net_hdr_mrg_rxbuf *hdr;
 385         unsigned int copy, hdr_len, hdr_padded_len;
 386         char *p;
 387
 388         p = page_address(page) + offset;
 389
 390         /* copy small packet so we can reuse these pages for small data */
 391         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 392         if (unlikely(!skb))
 393                 return NULL;
 394
 395         hdr = skb_vnet_hdr(skb);
 396
 397         hdr_len = vi->hdr_len;
 398         if (vi->mergeable_rx_bufs)
 399                 hdr_padded_len = sizeof(*hdr);
 400         else
 401                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 402
 403         /* hdr_valid means no XDP, so we can copy the vnet header */
 404         if (hdr_valid)
 405                 memcpy(hdr, p, hdr_len);
 406
 407         len -= hdr_len;
 408         offset += hdr_padded_len;
 409         p += hdr_padded_len;
 410
 411         copy = len;
 412         if (copy > skb_tailroom(skb))
 413                 copy = skb_tailroom(skb);
 414         skb_put_data(skb, p, copy);
 415
 416         if (metasize) {
 417                 __skb_pull(skb, metasize);
 418                 skb_metadata_set(skb, metasize);
 419         }
 420
 421         len -= copy;
 422         offset += copy;
 423
 424         if (vi->mergeable_rx_bufs) {
 425                 if (len)
 426                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 427                 else
 428                         put_page(page);
 429                 return skb;
 430         }
 431
 432         /*
 433          * Verify that we can indeed put this data into a skb.
 434          * This is here to handle cases when the device erroneously
 435          * tries to receive more than is possible. This is usually
 436          * the case of a broken device.
 437          */
 438         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 439                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 440                 dev_kfree_skb(skb);
 441                 return NULL;
 442         }
 443         BUG_ON(offset >= PAGE_SIZE);
 444         while (len) {
 445                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 446                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 447                                 frag_size, truesize);
 448                 len -= frag_size;
 449                 page = (struct page *)page->private;
 450                 offset = 0;
 451         }
 452
 453         if (page)
 454                 give_pages(rq, page);
 455
 456         return skb;
 457 }
 458
 459 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 460                                    struct send_queue *sq,
 461                                    struct xdp_frame *xdpf)
 462 {
 463         struct virtio_net_hdr_mrg_rxbuf *hdr;
 464         int err;
 465
 466         if (unlikely(xdpf->headroom < vi->hdr_len))
 467                 return -EOVERFLOW;
 468
 469         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 470         xdpf->data -= vi->hdr_len;
 471         /* Zero header and leave csum up to XDP layers */
 472         hdr = xdpf->data;
 473         memset(hdr, 0, vi->hdr_len);
 474         xdpf->len   += vi->hdr_len;
 475
 476         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 477
 478         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
 479                                    GFP_ATOMIC);
 480         if (unlikely(err))
 481                 return -ENOSPC; /* Caller handle free/refcnt */
 482
 483         return 0;
 484 }
 485
 486 static struct send_queue *virtnet_xdp_sq(struct virtnet_info *vi)
 487 {
 488         unsigned int qp;
 489
 490         qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
 491         return &vi->sq[qp];
 492 }
 493
 494 static int virtnet_xdp_xmit(struct net_device *dev,
 495                             int n, struct xdp_frame **frames, u32 flags)
 496 {
 497         struct virtnet_info *vi = netdev_priv(dev);
 498         struct receive_queue *rq = vi->rq;
 499         struct bpf_prog *xdp_prog;
 500         struct send_queue *sq;
 501         unsigned int len;
 502         int packets = 0;
 503         int bytes = 0;
 504         int drops = 0;
 505         int kicks = 0;
 506         int ret, err;
 507         void *ptr;
 508         int i;
 509
 510         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 511          * indicate XDP resources have been successfully allocated.
 512          */
 513         xdp_prog = rcu_access_pointer(rq->xdp_prog);
 514         if (!xdp_prog)
 515                 return -ENXIO;
 516
 517         sq = virtnet_xdp_sq(vi);
 518
 519         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 520                 ret = -EINVAL;
 521                 drops = n;
 522                 goto out;
 523         }
 524
 525         /* Free up any pending old buffers before queueing new ones. */
 526         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 527                 if (likely(is_xdp_frame(ptr))) {
 528                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 529
 530                         bytes += frame->len;
 531                         xdp_return_frame(frame);
 532                 } else {
 533                         struct sk_buff *skb = ptr;
 534
 535                         bytes += skb->len;
 536                         napi_consume_skb(skb, false);
 537                 }
 538                 packets++;
 539         }
 540
 541         for (i = 0; i < n; i++) {
 542                 struct xdp_frame *xdpf = frames[i];
 543
 544                 err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
 545                 if (err) {
 546                         xdp_return_frame_rx_napi(xdpf);
 547                         drops++;
 548                 }
 549         }
 550         ret = n - drops;
 551
 552         if (flags & XDP_XMIT_FLUSH) {
 553                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
 554                         kicks = 1;
 555         }
 556 out:
 557         u64_stats_update_begin(&sq->stats.syncp);
 558         sq->stats.bytes += bytes;
 559         sq->stats.packets += packets;
 560         sq->stats.xdp_tx += n;
 561         sq->stats.xdp_tx_drops += drops;
 562         sq->stats.kicks += kicks;
 563         u64_stats_update_end(&sq->stats.syncp);
 564
 565         return ret;
 566 }
 567
 568 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 569 {
 570         return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
 571 }
 572
 573 /* We copy the packet for XDP in the following cases:
 574  *
 575  * 1) Packet is scattered across multiple rx buffers.
 576  * 2) Headroom space is insufficient.
 577  *
 578  * This is inefficient but it's a temporary condition that
 579  * we hit right after XDP is enabled and until queue is refilled
 580  * with large buffers with sufficient headroom - so it should affect
 581  * at most queue size packets.
 582  * Afterwards, the conditions to enable
 583  * XDP should preclude the underlying device from sending packets
 584  * across multiple buffers (num_buf > 1), and we make sure buffers
 585  * have enough headroom.
 586  */
 587 static struct page *xdp_linearize_page(struct receive_queue *rq,
 588                                        u16 *num_buf,
 589                                        struct page *p,
 590                                        int offset,
 591                                        int page_off,
 592                                        unsigned int *len)
 593 {
 594         struct page *page = alloc_page(GFP_ATOMIC);
 595
 596         if (!page)
 597                 return NULL;
 598
 599         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 600         page_off += *len;
 601
 602         while (--*num_buf) {
 603                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 604                 unsigned int buflen;
 605                 void *buf;
 606                 int off;
 607
 608                 buf = virtqueue_get_buf(rq->vq, &buflen);
 609                 if (unlikely(!buf))
 610                         goto err_buf;
 611
 612                 p = virt_to_head_page(buf);
 613                 off = buf - page_address(p);
 614
 615                 /* guard against a misconfigured or uncooperative backend that
 616                  * is sending packet larger than the MTU.
 617                  */
 618                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 619                         put_page(p);
 620                         goto err_buf;
 621                 }
 622
 623                 memcpy(page_address(page) + page_off,
 624                        page_address(p) + off, buflen);
 625                 page_off += buflen;
 626                 put_page(p);
 627         }
 628
 629         /* Headroom does not contribute to packet length */
 630         *len = page_off - VIRTIO_XDP_HEADROOM;
 631         return page;
 632 err_buf:
 633         __free_pages(page, 0);
 634         return NULL;
 635 }
 636
 637 static struct sk_buff *receive_small(struct net_device *dev,
 638                                      struct virtnet_info *vi,
 639                                      struct receive_queue *rq,
 640                                      void *buf, void *ctx,
 641                                      unsigned int len,
 642                                      unsigned int *xdp_xmit,
 643                                      struct virtnet_rq_stats *stats)
 644 {
 645         struct sk_buff *skb;
 646         struct bpf_prog *xdp_prog;
 647         unsigned int xdp_headroom = (unsigned long)ctx;
 648         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 649         unsigned int headroom = vi->hdr_len + header_offset;
 650         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 651                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 652         struct page *page = virt_to_head_page(buf);
 653         unsigned int delta = 0;
 654         struct page *xdp_page;
 655         int err;
 656         unsigned int metasize = 0;
 657
 658         len -= vi->hdr_len;
 659         stats->bytes += len;
 660
 661         rcu_read_lock();
 662         xdp_prog = rcu_dereference(rq->xdp_prog);
 663         if (xdp_prog) {
 664                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 665                 struct xdp_frame *xdpf;
 666                 struct xdp_buff xdp;
 667                 void *orig_data;
 668                 u32 act;
 669
 670                 if (unlikely(hdr->hdr.gso_type))
 671                         goto err_xdp;
 672
 673                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 674                         int offset = buf - page_address(page) + header_offset;
 675                         unsigned int tlen = len + vi->hdr_len;
 676                         u16 num_buf = 1;
 677
 678                         xdp_headroom = virtnet_get_headroom(vi);
 679                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 680                         headroom = vi->hdr_len + header_offset;
 681                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 682                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 683                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 684                                                       offset, header_offset,
 685                                                       &tlen);
 686                         if (!xdp_page)
 687                                 goto err_xdp;
 688
 689                         buf = page_address(xdp_page);
 690                         put_page(page);
 691                         page = xdp_page;
 692                 }
 693
 694                 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
 695                 xdp.data = xdp.data_hard_start + xdp_headroom;
 696                 xdp.data_end = xdp.data + len;
 697                 xdp.data_meta = xdp.data;
 698                 xdp.rxq = &rq->xdp_rxq;
 699                 xdp.frame_sz = buflen;
 700                 orig_data = xdp.data;
 701                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 702                 stats->xdp_packets++;
 703
 704                 switch (act) {
 705                 case XDP_PASS:
 706                         /* Recalculate length in case bpf program changed it */
 707                         delta = orig_data - xdp.data;
 708                         len = xdp.data_end - xdp.data;
 709                         metasize = xdp.data - xdp.data_meta;
 710                         break;
 711                 case XDP_TX:
 712                         stats->xdp_tx++;
 713                         xdpf = xdp_convert_buff_to_frame(&xdp);
 714                         if (unlikely(!xdpf))
 715                                 goto err_xdp;
 716                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 717                         if (unlikely(err < 0)) {
 718                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 719                                 goto err_xdp;
 720                         }
 721                         *xdp_xmit |= VIRTIO_XDP_TX;
 722                         rcu_read_unlock();
 723                         goto xdp_xmit;
 724                 case XDP_REDIRECT:
 725                         stats->xdp_redirects++;
 726                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 727                         if (err)
 728                                 goto err_xdp;
 729                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 730                         rcu_read_unlock();
 731                         goto xdp_xmit;
 732                 default:
 733                         bpf_warn_invalid_xdp_action(act);
 734                         fallthrough;
 735                 case XDP_ABORTED:
 736                         trace_xdp_exception(vi->dev, xdp_prog, act);
 737                 case XDP_DROP:
 738                         goto err_xdp;
 739                 }
 740         }
 741         rcu_read_unlock();
 742
 743         skb = build_skb(buf, buflen);
 744         if (!skb) {
 745                 put_page(page);
 746                 goto err;
 747         }
 748         skb_reserve(skb, headroom - delta);
 749         skb_put(skb, len);
 750         if (!xdp_prog) {
 751                 buf += header_offset;
 752                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 753         } /* keep zeroed vnet hdr since XDP is loaded */
 754
 755         if (metasize)
 756                 skb_metadata_set(skb, metasize);
 757
 758 err:
 759         return skb;
 760
 761 err_xdp:
 762         rcu_read_unlock();
 763         stats->xdp_drops++;
 764         stats->drops++;
 765         put_page(page);
 766 xdp_xmit:
 767         return NULL;
 768 }
 769
 770 static struct sk_buff *receive_big(struct net_device *dev,
 771                                    struct virtnet_info *vi,
 772                                    struct receive_queue *rq,
 773                                    void *buf,
 774                                    unsigned int len,
 775                                    struct virtnet_rq_stats *stats)
 776 {
 777         struct page *page = buf;
 778         struct sk_buff *skb =
 779                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 780
 781         stats->bytes += len - vi->hdr_len;
 782         if (unlikely(!skb))
 783                 goto err;
 784
 785         return skb;
 786
 787 err:
 788         stats->drops++;
 789         give_pages(rq, page);
 790         return NULL;
 791 }
 792
 793 static struct sk_buff *receive_mergeable(struct net_device *dev,
 794                                          struct virtnet_info *vi,
 795                                          struct receive_queue *rq,
 796                                          void *buf,
 797                                          void *ctx,
 798                                          unsigned int len,
 799                                          unsigned int *xdp_xmit,
 800                                          struct virtnet_rq_stats *stats)
 801 {
 802         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 803         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 804         struct page *page = virt_to_head_page(buf);
 805         int offset = buf - page_address(page);
 806         struct sk_buff *head_skb, *curr_skb;
 807         struct bpf_prog *xdp_prog;
 808         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
 809         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 810         unsigned int metasize = 0;
 811         unsigned int frame_sz;
 812         int err;
 813
 814         head_skb = NULL;
 815         stats->bytes += len - vi->hdr_len;
 816
 817         rcu_read_lock();
 818         xdp_prog = rcu_dereference(rq->xdp_prog);
 819         if (xdp_prog) {
 820                 struct xdp_frame *xdpf;
 821                 struct page *xdp_page;
 822                 struct xdp_buff xdp;
 823                 void *data;
 824                 u32 act;
 825
 826                 /* Transient failure which in theory could occur if
 827                  * in-flight packets from before XDP was enabled reach
 828                  * the receive path after XDP is loaded.
 829                  */
 830                 if (unlikely(hdr->hdr.gso_type))
 831                         goto err_xdp;
 832
 833                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 834                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 835                  */
 836                 frame_sz = headroom ? PAGE_SIZE : truesize;
 837
 838                 /* This happens when rx buffer size is underestimated
 839                  * or headroom is not enough because of the buffer
 840                  * was refilled before XDP is set. This should only
 841                  * happen for the first several packets, so we don't
 842                  * care much about its performance.
 843                  */
 844                 if (unlikely(num_buf > 1 ||
 845                              headroom < virtnet_get_headroom(vi))) {
 846                         /* linearize data for XDP */
 847                         xdp_page = xdp_linearize_page(rq, &num_buf,
 848                                                       page, offset,
 849                                                       VIRTIO_XDP_HEADROOM,
 850                                                       &len);
 851                         frame_sz = PAGE_SIZE;
 852
 853                         if (!xdp_page)
 854                                 goto err_xdp;
 855                         offset = VIRTIO_XDP_HEADROOM;
 856                 } else {
 857                         xdp_page = page;
 858                 }
 859
 860                 /* Allow consuming headroom but reserve enough space to push
 861                  * the descriptor on if we get an XDP_TX return code.
 862                  */
 863                 data = page_address(xdp_page) + offset;
 864                 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
 865                 xdp.data = data + vi->hdr_len;
 866                 xdp.data_end = xdp.data + (len - vi->hdr_len);
 867                 xdp.data_meta = xdp.data;
 868                 xdp.rxq = &rq->xdp_rxq;
 869                 xdp.frame_sz = frame_sz - vi->hdr_len;
 870
 871                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 872                 stats->xdp_packets++;
 873
 874                 switch (act) {
 875                 case XDP_PASS:
 876                         metasize = xdp.data - xdp.data_meta;
 877
 878                         /* recalculate offset to account for any header
 879                          * adjustments and minus the metasize to copy the
 880                          * metadata in page_to_skb(). Note other cases do not
 881                          * build an skb and avoid using offset
 882                          */
 883                         offset = xdp.data - page_address(xdp_page) -
 884                                  vi->hdr_len - metasize;
 885
 886                         /* recalculate len if xdp.data, xdp.data_end or
 887                          * xdp.data_meta were adjusted
 888                          */
 889                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 890                         /* We can only create skb based on xdp_page. */
 891                         if (unlikely(xdp_page != page)) {
 892                                 rcu_read_unlock();
 893                                 put_page(page);
 894                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
 895                                                        len, PAGE_SIZE, false,
 896                                                        metasize);
 897                                 return head_skb;
 898                         }
 899                         break;
 900                 case XDP_TX:
 901                         stats->xdp_tx++;
 902                         xdpf = xdp_convert_buff_to_frame(&xdp);
 903                         if (unlikely(!xdpf))
 904                                 goto err_xdp;
 905                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 906                         if (unlikely(err < 0)) {
 907                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 908                                 if (unlikely(xdp_page != page))
 909                                         put_page(xdp_page);
 910                                 goto err_xdp;
 911                         }
 912                         *xdp_xmit |= VIRTIO_XDP_TX;
 913                         if (unlikely(xdp_page != page))
 914                                 put_page(page);
 915                         rcu_read_unlock();
 916                         goto xdp_xmit;
 917                 case XDP_REDIRECT:
 918                         stats->xdp_redirects++;
 919                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 920                         if (err) {
 921                                 if (unlikely(xdp_page != page))
 922                                         put_page(xdp_page);
 923                                 goto err_xdp;
 924                         }
 925                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 926                         if (unlikely(xdp_page != page))
 927                                 put_page(page);
 928                         rcu_read_unlock();
 929                         goto xdp_xmit;
 930                 default:
 931                         bpf_warn_invalid_xdp_action(act);
 932                         fallthrough;
 933                 case XDP_ABORTED:
 934                         trace_xdp_exception(vi->dev, xdp_prog, act);
 935                         fallthrough;
 936                 case XDP_DROP:
 937                         if (unlikely(xdp_page != page))
 938                                 __free_pages(xdp_page, 0);
 939                         goto err_xdp;
 940                 }
 941         }
 942         rcu_read_unlock();
 943
 944         if (unlikely(len > truesize)) {
 945                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 946                          dev->name, len, (unsigned long)ctx);
 947                 dev->stats.rx_length_errors++;
 948                 goto err_skb;
 949         }
 950
 951         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
 952                                metasize);
 953         curr_skb = head_skb;
 954
 955         if (unlikely(!curr_skb))
 956                 goto err_skb;
 957         while (--num_buf) {
 958                 int num_skb_frags;
 959
 960                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
 961                 if (unlikely(!buf)) {
 962                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 963                                  dev->name, num_buf,
 964                                  virtio16_to_cpu(vi->vdev,
 965                                                  hdr->num_buffers));
 966                         dev->stats.rx_length_errors++;
 967                         goto err_buf;
 968                 }
 969
 970                 stats->bytes += len;
 971                 page = virt_to_head_page(buf);
 972
 973                 truesize = mergeable_ctx_to_truesize(ctx);
 974                 if (unlikely(len > truesize)) {
 975                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 976                                  dev->name, len, (unsigned long)ctx);
 977                         dev->stats.rx_length_errors++;
 978                         goto err_skb;
 979                 }
 980
 981                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 982                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 983                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 984
 985                         if (unlikely(!nskb))
 986                                 goto err_skb;
 987                         if (curr_skb == head_skb)
 988                                 skb_shinfo(curr_skb)->frag_list = nskb;
 989                         else
 990                                 curr_skb->next = nskb;
 991                         curr_skb = nskb;
 992                         head_skb->truesize += nskb->truesize;
 993                         num_skb_frags = 0;
 994                 }
 995                 if (curr_skb != head_skb) {
 996                         head_skb->data_len += len;
 997                         head_skb->len += len;
 998                         head_skb->truesize += truesize;
 999                 }
1000                 offset = buf - page_address(page);
1001                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1002                         put_page(page);
1003                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1004                                              len, truesize);
1005                 } else {
1006                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1007                                         offset, len, truesize);
1008                 }
1009         }
1010
1011         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1012         return head_skb;
1013
1014 err_xdp:
1015         rcu_read_unlock();
1016         stats->xdp_drops++;
1017 err_skb:
1018         put_page(page);
1019         while (num_buf-- > 1) {
1020                 buf = virtqueue_get_buf(rq->vq, &len);
1021                 if (unlikely(!buf)) {
1022                         pr_debug("%s: rx error: %d buffers missing\n",
1023                                  dev->name, num_buf);
1024                         dev->stats.rx_length_errors++;
1025                         break;
1026                 }
1027                 stats->bytes += len;
1028                 page = virt_to_head_page(buf);
1029                 put_page(page);
1030         }
1031 err_buf:
1032         stats->drops++;
1033         dev_kfree_skb(head_skb);
1034 xdp_xmit:
1035         return NULL;
1036 }
1037
1038 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1039                         void *buf, unsigned int len, void **ctx,
1040                         unsigned int *xdp_xmit,
1041                         struct virtnet_rq_stats *stats)
1042 {
1043         struct net_device *dev = vi->dev;
1044         struct sk_buff *skb;
1045         struct virtio_net_hdr_mrg_rxbuf *hdr;
1046
1047         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1048                 pr_debug("%s: short packet %i\n", dev->name, len);
1049                 dev->stats.rx_length_errors++;
1050                 if (vi->mergeable_rx_bufs) {
1051                         put_page(virt_to_head_page(buf));
1052                 } else if (vi->big_packets) {
1053                         give_pages(rq, buf);
1054                 } else {
1055                         put_page(virt_to_head_page(buf));
1056                 }
1057                 return;
1058         }
1059
1060         if (vi->mergeable_rx_bufs)
1061                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1062                                         stats);
1063         else if (vi->big_packets)
1064                 skb = receive_big(dev, vi, rq, buf, len, stats);
1065         else
1066                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1067
1068         if (unlikely(!skb))
1069                 return;
1070
1071         hdr = skb_vnet_hdr(skb);
1072
1073         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1074                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1075
1076         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1077                                   virtio_is_little_endian(vi->vdev))) {
1078                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1079                                      dev->name, hdr->hdr.gso_type,
1080                                      hdr->hdr.gso_size);
1081                 goto frame_err;
1082         }
1083
1084         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1085         skb->protocol = eth_type_trans(skb, dev);
1086         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1087                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1088
1089         napi_gro_receive(&rq->napi, skb);
1090         return;
1091
1092 frame_err:
1093         dev->stats.rx_frame_errors++;
1094         dev_kfree_skb(skb);
1095 }
1096
1097 /* Unlike mergeable buffers, all buffers are allocated to the
1098  * same size, except for the headroom. For this reason we do
1099  * not need to use  mergeable_len_to_ctx here - it is enough
1100  * to store the headroom as the context ignoring the truesize.
1101  */
1102 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1103                              gfp_t gfp)
1104 {
1105         struct page_frag *alloc_frag = &rq->alloc_frag;
1106         char *buf;
1107         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1108         void *ctx = (void *)(unsigned long)xdp_headroom;
1109         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1110         int err;
1111
1112         len = SKB_DATA_ALIGN(len) +
1113               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1114         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1115                 return -ENOMEM;
1116
1117         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1118         get_page(alloc_frag->page);
1119         alloc_frag->offset += len;
1120         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1121                     vi->hdr_len + GOOD_PACKET_LEN);
1122         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1123         if (err < 0)
1124                 put_page(virt_to_head_page(buf));
1125         return err;
1126 }
1127
1128 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1129                            gfp_t gfp)
1130 {
1131         struct page *first, *list = NULL;
1132         char *p;
1133         int i, err, offset;
1134
1135         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1136
1137         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1138         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1139                 first = get_a_page(rq, gfp);
1140                 if (!first) {
1141                         if (list)
1142                                 give_pages(rq, list);
1143                         return -ENOMEM;
1144                 }
1145                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1146
1147                 /* chain new page in list head to match sg */
1148                 first->private = (unsigned long)list;
1149                 list = first;
1150         }
1151
1152         first = get_a_page(rq, gfp);
1153         if (!first) {
1154                 give_pages(rq, list);
1155                 return -ENOMEM;
1156         }
1157         p = page_address(first);
1158
1159         /* rq->sg[0], rq->sg[1] share the same page */
1160         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1161         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1162
1163         /* rq->sg[1] for data packet, from offset */
1164         offset = sizeof(struct padded_vnet_hdr);
1165         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1166
1167         /* chain first in list head */
1168         first->private = (unsigned long)list;
1169         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1170                                   first, gfp);
1171         if (err < 0)
1172                 give_pages(rq, first);
1173
1174         return err;
1175 }
1176
1177 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1178                                           struct ewma_pkt_len *avg_pkt_len,
1179                                           unsigned int room)
1180 {
1181         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1182         unsigned int len;
1183
1184         if (room)
1185                 return PAGE_SIZE - room;
1186
1187         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1188                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1189
1190         return ALIGN(len, L1_CACHE_BYTES);
1191 }
1192
1193 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1194                                  struct receive_queue *rq, gfp_t gfp)
1195 {
1196         struct page_frag *alloc_frag = &rq->alloc_frag;
1197         unsigned int headroom = virtnet_get_headroom(vi);
1198         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1199         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1200         char *buf;
1201         void *ctx;
1202         int err;
1203         unsigned int len, hole;
1204
1205         /* Extra tailroom is needed to satisfy XDP's assumption. This
1206          * means rx frags coalescing won't work, but consider we've
1207          * disabled GSO for XDP, it won't be a big issue.
1208          */
1209         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1210         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1211                 return -ENOMEM;
1212
1213         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1214         buf += headroom; /* advance address leaving hole at front of pkt */
1215         get_page(alloc_frag->page);
1216         alloc_frag->offset += len + room;
1217         hole = alloc_frag->size - alloc_frag->offset;
1218         if (hole < len + room) {
1219                 /* To avoid internal fragmentation, if there is very likely not
1220                  * enough space for another buffer, add the remaining space to
1221                  * the current buffer.
1222                  */
1223                 len += hole;
1224                 alloc_frag->offset += hole;
1225         }
1226
1227         sg_init_one(rq->sg, buf, len);
1228         ctx = mergeable_len_to_ctx(len, headroom);
1229         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1230         if (err < 0)
1231                 put_page(virt_to_head_page(buf));
1232
1233         return err;
1234 }
1235
1236 /*
1237  * Returns false if we couldn't fill entirely (OOM).
1238  *
1239  * Normally run in the receive path, but can also be run from ndo_open
1240  * before we're receiving packets, or from refill_work which is
1241  * careful to disable receiving (using napi_disable).
1242  */
1243 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1244                           gfp_t gfp)
1245 {
1246         int err;
1247         bool oom;
1248
1249         do {
1250                 if (vi->mergeable_rx_bufs)
1251                         err = add_recvbuf_mergeable(vi, rq, gfp);
1252                 else if (vi->big_packets)
1253                         err = add_recvbuf_big(vi, rq, gfp);
1254                 else
1255                         err = add_recvbuf_small(vi, rq, gfp);
1256
1257                 oom = err == -ENOMEM;
1258                 if (err)
1259                         break;
1260         } while (rq->vq->num_free);
1261         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1262                 unsigned long flags;
1263
1264                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1265                 rq->stats.kicks++;
1266                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1267         }
1268
1269         return !oom;
1270 }
1271
1272 static void skb_recv_done(struct virtqueue *rvq)
1273 {
1274         struct virtnet_info *vi = rvq->vdev->priv;
1275         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1276
1277         virtqueue_napi_schedule(&rq->napi, rvq);
1278 }
1279
1280 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1281 {
1282         napi_enable(napi);
1283
1284         /* If all buffers were filled by other side before we napi_enabled, we
1285          * won't get another interrupt, so process any outstanding packets now.
1286          * Call local_bh_enable after to trigger softIRQ processing.
1287          */
1288         local_bh_disable();
1289         virtqueue_napi_schedule(napi, vq);
1290         local_bh_enable();
1291 }
1292
1293 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1294                                    struct virtqueue *vq,
1295                                    struct napi_struct *napi)
1296 {
1297         if (!napi->weight)
1298                 return;
1299
1300         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1301          * enable the feature if this is likely affine with the transmit path.
1302          */
1303         if (!vi->affinity_hint_set) {
1304                 napi->weight = 0;
1305                 return;
1306         }
1307
1308         return virtnet_napi_enable(vq, napi);
1309 }
1310
1311 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1312 {
1313         if (napi->weight)
1314                 napi_disable(napi);
1315 }
1316
1317 static void refill_work(struct work_struct *work)
1318 {
1319         struct virtnet_info *vi =
1320                 container_of(work, struct virtnet_info, refill.work);
1321         bool still_empty;
1322         int i;
1323
1324         for (i = 0; i < vi->curr_queue_pairs; i++) {
1325                 struct receive_queue *rq = &vi->rq[i];
1326
1327                 napi_disable(&rq->napi);
1328                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1329                 virtnet_napi_enable(rq->vq, &rq->napi);
1330
1331                 /* In theory, this can happen: if we don't get any buffers in
1332                  * we will *never* try to fill again.
1333                  */
1334                 if (still_empty)
1335                         schedule_delayed_work(&vi->refill, HZ/2);
1336         }
1337 }
1338
1339 static int virtnet_receive(struct receive_queue *rq, int budget,
1340                            unsigned int *xdp_xmit)
1341 {
1342         struct virtnet_info *vi = rq->vq->vdev->priv;
1343         struct virtnet_rq_stats stats = {};
1344         unsigned int len;
1345         void *buf;
1346         int i;
1347
1348         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1349                 void *ctx;
1350
1351                 while (stats.packets < budget &&
1352                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1353                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1354                         stats.packets++;
1355                 }
1356         } else {
1357                 while (stats.packets < budget &&
1358                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1359                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1360                         stats.packets++;
1361                 }
1362         }
1363
1364         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1365                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1366                         schedule_delayed_work(&vi->refill, 0);
1367         }
1368
1369         u64_stats_update_begin(&rq->stats.syncp);
1370         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1371                 size_t offset = virtnet_rq_stats_desc[i].offset;
1372                 u64 *item;
1373
1374                 item = (u64 *)((u8 *)&rq->stats + offset);
1375                 *item += *(u64 *)((u8 *)&stats + offset);
1376         }
1377         u64_stats_update_end(&rq->stats.syncp);
1378
1379         return stats.packets;
1380 }
1381
1382 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1383 {
1384         unsigned int len;
1385         unsigned int packets = 0;
1386         unsigned int bytes = 0;
1387         void *ptr;
1388
1389         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1390                 if (likely(!is_xdp_frame(ptr))) {
1391                         struct sk_buff *skb = ptr;
1392
1393                         pr_debug("Sent skb %p\n", skb);
1394
1395                         bytes += skb->len;
1396                         napi_consume_skb(skb, in_napi);
1397                 } else {
1398                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1399
1400                         bytes += frame->len;
1401                         xdp_return_frame(frame);
1402                 }
1403                 packets++;
1404         }
1405
1406         /* Avoid overhead when no packets have been processed
1407          * happens when called speculatively from start_xmit.
1408          */
1409         if (!packets)
1410                 return;
1411
1412         u64_stats_update_begin(&sq->stats.syncp);
1413         sq->stats.bytes += bytes;
1414         sq->stats.packets += packets;
1415         u64_stats_update_end(&sq->stats.syncp);
1416 }
1417
1418 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1419 {
1420         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1421                 return false;
1422         else if (q < vi->curr_queue_pairs)
1423                 return true;
1424         else
1425                 return false;
1426 }
1427
1428 static void virtnet_poll_cleantx(struct receive_queue *rq)
1429 {
1430         struct virtnet_info *vi = rq->vq->vdev->priv;
1431         unsigned int index = vq2rxq(rq->vq);
1432         struct send_queue *sq = &vi->sq[index];
1433         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1434
1435         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1436                 return;
1437
1438         if (__netif_tx_trylock(txq)) {
1439                 free_old_xmit_skbs(sq, true);
1440                 __netif_tx_unlock(txq);
1441         }
1442
1443         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1444                 netif_tx_wake_queue(txq);
1445 }
1446
1447 static int virtnet_poll(struct napi_struct *napi, int budget)
1448 {
1449         struct receive_queue *rq =
1450                 container_of(napi, struct receive_queue, napi);
1451         struct virtnet_info *vi = rq->vq->vdev->priv;
1452         struct send_queue *sq;
1453         unsigned int received;
1454         unsigned int xdp_xmit = 0;
1455
1456         virtnet_poll_cleantx(rq);
1457
1458         received = virtnet_receive(rq, budget, &xdp_xmit);
1459
1460         /* Out of packets? */
1461         if (received < budget)
1462                 virtqueue_napi_complete(napi, rq->vq, received);
1463
1464         if (xdp_xmit & VIRTIO_XDP_REDIR)
1465                 xdp_do_flush();
1466
1467         if (xdp_xmit & VIRTIO_XDP_TX) {
1468                 sq = virtnet_xdp_sq(vi);
1469                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1470                         u64_stats_update_begin(&sq->stats.syncp);
1471                         sq->stats.kicks++;
1472                         u64_stats_update_end(&sq->stats.syncp);
1473                 }
1474         }
1475
1476         return received;
1477 }
1478
1479 static int virtnet_open(struct net_device *dev)
1480 {
1481         struct virtnet_info *vi = netdev_priv(dev);
1482         int i, err;
1483
1484         for (i = 0; i < vi->max_queue_pairs; i++) {
1485                 if (i < vi->curr_queue_pairs)
1486                         /* Make sure we have some buffers: if oom use wq. */
1487                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1488                                 schedule_delayed_work(&vi->refill, 0);
1489
1490                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1491                 if (err < 0)
1492                         return err;
1493
1494                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1495                                                  MEM_TYPE_PAGE_SHARED, NULL);
1496                 if (err < 0) {
1497                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1498                         return err;
1499                 }
1500
1501                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1502                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1503         }
1504
1505         return 0;
1506 }
1507
1508 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1509 {
1510         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1511         struct virtnet_info *vi = sq->vq->vdev->priv;
1512         unsigned int index = vq2txq(sq->vq);
1513         struct netdev_queue *txq;
1514
1515         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1516                 /* We don't need to enable cb for XDP */
1517                 napi_complete_done(napi, 0);
1518                 return 0;
1519         }
1520
1521         txq = netdev_get_tx_queue(vi->dev, index);
1522         __netif_tx_lock(txq, raw_smp_processor_id());
1523         free_old_xmit_skbs(sq, true);
1524         __netif_tx_unlock(txq);
1525
1526         virtqueue_napi_complete(napi, sq->vq, 0);
1527
1528         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1529                 netif_tx_wake_queue(txq);
1530
1531         return 0;
1532 }
1533
1534 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1535 {
1536         struct virtio_net_hdr_mrg_rxbuf *hdr;
1537         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1538         struct virtnet_info *vi = sq->vq->vdev->priv;
1539         int num_sg;
1540         unsigned hdr_len = vi->hdr_len;
1541         bool can_push;
1542
1543         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1544
1545         can_push = vi->any_header_sg &&
1546                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1547                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1548         /* Even if we can, don't push here yet as this would skew
1549          * csum_start offset below. */
1550         if (can_push)
1551                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1552         else
1553                 hdr = skb_vnet_hdr(skb);
1554
1555         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1556                                     virtio_is_little_endian(vi->vdev), false,
1557                                     0))
1558                 BUG();
1559
1560         if (vi->mergeable_rx_bufs)
1561                 hdr->num_buffers = 0;
1562
1563         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1564         if (can_push) {
1565                 __skb_push(skb, hdr_len);
1566                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1567                 if (unlikely(num_sg < 0))
1568                         return num_sg;
1569                 /* Pull header back to avoid skew in tx bytes calculations. */
1570                 __skb_pull(skb, hdr_len);
1571         } else {
1572                 sg_set_buf(sq->sg, hdr, hdr_len);
1573                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1574                 if (unlikely(num_sg < 0))
1575                         return num_sg;
1576                 num_sg++;
1577         }
1578         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1579 }
1580
1581 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1582 {
1583         struct virtnet_info *vi = netdev_priv(dev);
1584         int qnum = skb_get_queue_mapping(skb);
1585         struct send_queue *sq = &vi->sq[qnum];
1586         int err;
1587         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1588         bool kick = !netdev_xmit_more();
1589         bool use_napi = sq->napi.weight;
1590
1591         /* Free up any pending old buffers before queueing new ones. */
1592         free_old_xmit_skbs(sq, false);
1593
1594         if (use_napi && kick)
1595                 virtqueue_enable_cb_delayed(sq->vq);
1596
1597         /* timestamp packet in software */
1598         skb_tx_timestamp(skb);
1599
1600         /* Try to transmit */
1601         err = xmit_skb(sq, skb);
1602
1603         /* This should not happen! */
1604         if (unlikely(err)) {
1605                 dev->stats.tx_fifo_errors++;
1606                 if (net_ratelimit())
1607                         dev_warn(&dev->dev,
1608                                  "Unexpected TXQ (%d) queue failure: %d\n",
1609                                  qnum, err);
1610                 dev->stats.tx_dropped++;
1611                 dev_kfree_skb_any(skb);
1612                 return NETDEV_TX_OK;
1613         }
1614
1615         /* Don't wait up for transmitted skbs to be freed. */
1616         if (!use_napi) {
1617                 skb_orphan(skb);
1618                 nf_reset_ct(skb);
1619         }
1620
1621         /* If running out of space, stop queue to avoid getting packets that we
1622          * are then unable to transmit.
1623          * An alternative would be to force queuing layer to requeue the skb by
1624          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1625          * returned in a normal path of operation: it means that driver is not
1626          * maintaining the TX queue stop/start state properly, and causes
1627          * the stack to do a non-trivial amount of useless work.
1628          * Since most packets only take 1 or 2 ring slots, stopping the queue
1629          * early means 16 slots are typically wasted.
1630          */
1631         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1632                 netif_stop_subqueue(dev, qnum);
1633                 if (!use_napi &&
1634                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1635                         /* More just got used, free them then recheck. */
1636                         free_old_xmit_skbs(sq, false);
1637                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1638                                 netif_start_subqueue(dev, qnum);
1639                                 virtqueue_disable_cb(sq->vq);
1640                         }
1641                 }
1642         }
1643
1644         if (kick || netif_xmit_stopped(txq)) {
1645                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1646                         u64_stats_update_begin(&sq->stats.syncp);
1647                         sq->stats.kicks++;
1648                         u64_stats_update_end(&sq->stats.syncp);
1649                 }
1650         }
1651
1652         return NETDEV_TX_OK;
1653 }
1654
1655 /*
1656  * Send command via the control virtqueue and check status.  Commands
1657  * supported by the hypervisor, as indicated by feature bits, should
1658  * never fail unless improperly formatted.
1659  */
1660 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1661                                  struct scatterlist *out)
1662 {
1663         struct scatterlist *sgs[4], hdr, stat;
1664         unsigned out_num = 0, tmp;
1665
1666         /* Caller should know better */
1667         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1668
1669         vi->ctrl->status = ~0;
1670         vi->ctrl->hdr.class = class;
1671         vi->ctrl->hdr.cmd = cmd;
1672         /* Add header */
1673         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1674         sgs[out_num++] = &hdr;
1675
1676         if (out)
1677                 sgs[out_num++] = out;
1678
1679         /* Add return status. */
1680         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1681         sgs[out_num] = &stat;
1682
1683         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1684         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1685
1686         if (unlikely(!virtqueue_kick(vi->cvq)))
1687                 return vi->ctrl->status == VIRTIO_NET_OK;
1688
1689         /* Spin for a response, the kick causes an ioport write, trapping
1690          * into the hypervisor, so the request should be handled immediately.
1691          */
1692         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1693                !virtqueue_is_broken(vi->cvq))
1694                 cpu_relax();
1695
1696         return vi->ctrl->status == VIRTIO_NET_OK;
1697 }
1698
1699 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1700 {
1701         struct virtnet_info *vi = netdev_priv(dev);
1702         struct virtio_device *vdev = vi->vdev;
1703         int ret;
1704         struct sockaddr *addr;
1705         struct scatterlist sg;
1706
1707         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1708                 return -EOPNOTSUPP;
1709
1710         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1711         if (!addr)
1712                 return -ENOMEM;
1713
1714         ret = eth_prepare_mac_addr_change(dev, addr);
1715         if (ret)
1716                 goto out;
1717
1718         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1719                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1720                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1721                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1722                         dev_warn(&vdev->dev,
1723                                  "Failed to set mac address by vq command.\n");
1724                         ret = -EINVAL;
1725                         goto out;
1726                 }
1727         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1728                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1729                 unsigned int i;
1730
1731                 /* Naturally, this has an atomicity problem. */
1732                 for (i = 0; i < dev->addr_len; i++)
1733                         virtio_cwrite8(vdev,
1734                                        offsetof(struct virtio_net_config, mac) +
1735                                        i, addr->sa_data[i]);
1736         }
1737
1738         eth_commit_mac_addr_change(dev, p);
1739         ret = 0;
1740
1741 out:
1742         kfree(addr);
1743         return ret;
1744 }
1745
1746 static void virtnet_stats(struct net_device *dev,
1747                           struct rtnl_link_stats64 *tot)
1748 {
1749         struct virtnet_info *vi = netdev_priv(dev);
1750         unsigned int start;
1751         int i;
1752
1753         for (i = 0; i < vi->max_queue_pairs; i++) {
1754                 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1755                 struct receive_queue *rq = &vi->rq[i];
1756                 struct send_queue *sq = &vi->sq[i];
1757
1758                 do {
1759                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1760                         tpackets = sq->stats.packets;
1761                         tbytes   = sq->stats.bytes;
1762                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1763
1764                 do {
1765                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1766                         rpackets = rq->stats.packets;
1767                         rbytes   = rq->stats.bytes;
1768                         rdrops   = rq->stats.drops;
1769                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1770
1771                 tot->rx_packets += rpackets;
1772                 tot->tx_packets += tpackets;
1773                 tot->rx_bytes   += rbytes;
1774                 tot->tx_bytes   += tbytes;
1775                 tot->rx_dropped += rdrops;
1776         }
1777
1778         tot->tx_dropped = dev->stats.tx_dropped;
1779         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1780         tot->rx_length_errors = dev->stats.rx_length_errors;
1781         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1782 }
1783
1784 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1785 {
1786         rtnl_lock();
1787         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1788                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1789                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1790         rtnl_unlock();
1791 }
1792
1793 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1794 {
1795         struct scatterlist sg;
1796         struct net_device *dev = vi->dev;
1797
1798         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1799                 return 0;
1800
1801         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1802         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1803
1804         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1805                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1806                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1807                          queue_pairs);
1808                 return -EINVAL;
1809         } else {
1810                 vi->curr_queue_pairs = queue_pairs;
1811                 /* virtnet_open() will refill when device is going to up. */
1812                 if (dev->flags & IFF_UP)
1813                         schedule_delayed_work(&vi->refill, 0);
1814         }
1815
1816         return 0;
1817 }
1818
1819 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1820 {
1821         int err;
1822
1823         rtnl_lock();
1824         err = _virtnet_set_queues(vi, queue_pairs);
1825         rtnl_unlock();
1826         return err;
1827 }
1828
1829 static int virtnet_close(struct net_device *dev)
1830 {
1831         struct virtnet_info *vi = netdev_priv(dev);
1832         int i;
1833
1834         /* Make sure refill_work doesn't re-enable napi! */
1835         cancel_delayed_work_sync(&vi->refill);
1836
1837         for (i = 0; i < vi->max_queue_pairs; i++) {
1838                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1839                 napi_disable(&vi->rq[i].napi);
1840                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1841         }
1842
1843         return 0;
1844 }
1845
1846 static void virtnet_set_rx_mode(struct net_device *dev)
1847 {
1848         struct virtnet_info *vi = netdev_priv(dev);
1849         struct scatterlist sg[2];
1850         struct virtio_net_ctrl_mac *mac_data;
1851         struct netdev_hw_addr *ha;
1852         int uc_count;
1853         int mc_count;
1854         void *buf;
1855         int i;
1856
1857         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1858         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1859                 return;
1860
1861         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1862         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1863
1864         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1865
1866         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1867                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1868                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1869                          vi->ctrl->promisc ? "en" : "dis");
1870
1871         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1872
1873         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1874                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1875                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1876                          vi->ctrl->allmulti ? "en" : "dis");
1877
1878         uc_count = netdev_uc_count(dev);
1879         mc_count = netdev_mc_count(dev);
1880         /* MAC filter - use one buffer for both lists */
1881         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1882                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1883         mac_data = buf;
1884         if (!buf)
1885                 return;
1886
1887         sg_init_table(sg, 2);
1888
1889         /* Store the unicast list and count in the front of the buffer */
1890         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1891         i = 0;
1892         netdev_for_each_uc_addr(ha, dev)
1893                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1894
1895         sg_set_buf(&sg[0], mac_data,
1896                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1897
1898         /* multicast list and count fill the end */
1899         mac_data = (void *)&mac_data->macs[uc_count][0];
1900
1901         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1902         i = 0;
1903         netdev_for_each_mc_addr(ha, dev)
1904                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1905
1906         sg_set_buf(&sg[1], mac_data,
1907                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1908
1909         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1910                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1911                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1912
1913         kfree(buf);
1914 }
1915
1916 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1917                                    __be16 proto, u16 vid)
1918 {
1919         struct virtnet_info *vi = netdev_priv(dev);
1920         struct scatterlist sg;
1921
1922         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1923         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1924
1925         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1926                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1927                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1928         return 0;
1929 }
1930
1931 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1932                                     __be16 proto, u16 vid)
1933 {
1934         struct virtnet_info *vi = netdev_priv(dev);
1935         struct scatterlist sg;
1936
1937         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1938         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1939
1940         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1941                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1942                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1943         return 0;
1944 }
1945
1946 static void virtnet_clean_affinity(struct virtnet_info *vi)
1947 {
1948         int i;
1949
1950         if (vi->affinity_hint_set) {
1951                 for (i = 0; i < vi->max_queue_pairs; i++) {
1952                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
1953                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
1954                 }
1955
1956                 vi->affinity_hint_set = false;
1957         }
1958 }
1959
1960 static void virtnet_set_affinity(struct virtnet_info *vi)
1961 {
1962         cpumask_var_t mask;
1963         int stragglers;
1964         int group_size;
1965         int i, j, cpu;
1966         int num_cpu;
1967         int stride;
1968
1969         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
1970                 virtnet_clean_affinity(vi);
1971                 return;
1972         }
1973
1974         num_cpu = num_online_cpus();
1975         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
1976         stragglers = num_cpu >= vi->curr_queue_pairs ?
1977                         num_cpu % vi->curr_queue_pairs :
1978                         0;
1979         cpu = cpumask_next(-1, cpu_online_mask);
1980
1981         for (i = 0; i < vi->curr_queue_pairs; i++) {
1982                 group_size = stride + (i < stragglers ? 1 : 0);
1983
1984                 for (j = 0; j < group_size; j++) {
1985                         cpumask_set_cpu(cpu, mask);
1986                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
1987                                                 nr_cpu_ids, false);
1988                 }
1989                 virtqueue_set_affinity(vi->rq[i].vq, mask);
1990                 virtqueue_set_affinity(vi->sq[i].vq, mask);
1991                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, false);
1992                 cpumask_clear(mask);
1993         }
1994
1995         vi->affinity_hint_set = true;
1996         free_cpumask_var(mask);
1997 }
1998
1999 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2000 {
2001         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2002                                                    node);
2003         virtnet_set_affinity(vi);
2004         return 0;
2005 }
2006
2007 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2008 {
2009         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2010                                                    node_dead);
2011         virtnet_set_affinity(vi);
2012         return 0;
2013 }
2014
2015 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2016 {
2017         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2018                                                    node);
2019
2020         virtnet_clean_affinity(vi);
2021         return 0;
2022 }
2023
2024 static enum cpuhp_state virtionet_online;
2025
2026 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2027 {
2028         int ret;
2029
2030         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2031         if (ret)
2032                 return ret;
2033         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2034                                                &vi->node_dead);
2035         if (!ret)
2036                 return ret;
2037         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2038         return ret;
2039 }
2040
2041 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2042 {
2043         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2044         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2045                                             &vi->node_dead);
2046 }
2047
2048 static void virtnet_get_ringparam(struct net_device *dev,
2049                                 struct ethtool_ringparam *ring)
2050 {
2051         struct virtnet_info *vi = netdev_priv(dev);
2052
2053         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2054         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2055         ring->rx_pending = ring->rx_max_pending;
2056         ring->tx_pending = ring->tx_max_pending;
2057 }
2058
2059
2060 static void virtnet_get_drvinfo(struct net_device *dev,
2061                                 struct ethtool_drvinfo *info)
2062 {
2063         struct virtnet_info *vi = netdev_priv(dev);
2064         struct virtio_device *vdev = vi->vdev;
2065
2066         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2067         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2068         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2069
2070 }
2071
2072 /* TODO: Eliminate OOO packets during switching */
2073 static int virtnet_set_channels(struct net_device *dev,
2074                                 struct ethtool_channels *channels)
2075 {
2076         struct virtnet_info *vi = netdev_priv(dev);
2077         u16 queue_pairs = channels->combined_count;
2078         int err;
2079
2080         /* We don't support separate rx/tx channels.
2081          * We don't allow setting 'other' channels.
2082          */
2083         if (channels->rx_count || channels->tx_count || channels->other_count)
2084                 return -EINVAL;
2085
2086         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2087                 return -EINVAL;
2088
2089         /* For now we don't support modifying channels while XDP is loaded
2090          * also when XDP is loaded all RX queues have XDP programs so we only
2091          * need to check a single RX queue.
2092          */
2093         if (vi->rq[0].xdp_prog)
2094                 return -EINVAL;
2095
2096         get_online_cpus();
2097         err = _virtnet_set_queues(vi, queue_pairs);
2098         if (!err) {
2099                 netif_set_real_num_tx_queues(dev, queue_pairs);
2100                 netif_set_real_num_rx_queues(dev, queue_pairs);
2101
2102                 virtnet_set_affinity(vi);
2103         }
2104         put_online_cpus();
2105
2106         return err;
2107 }
2108
2109 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2110 {
2111         struct virtnet_info *vi = netdev_priv(dev);
2112         char *p = (char *)data;
2113         unsigned int i, j;
2114
2115         switch (stringset) {
2116         case ETH_SS_STATS:
2117                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2118                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2119                                 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
2120                                          i, virtnet_rq_stats_desc[j].desc);
2121                                 p += ETH_GSTRING_LEN;
2122                         }
2123                 }
2124
2125                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2126                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2127                                 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
2128                                          i, virtnet_sq_stats_desc[j].desc);
2129                                 p += ETH_GSTRING_LEN;
2130                         }
2131                 }
2132                 break;
2133         }
2134 }
2135
2136 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2137 {
2138         struct virtnet_info *vi = netdev_priv(dev);
2139
2140         switch (sset) {
2141         case ETH_SS_STATS:
2142                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2143                                                VIRTNET_SQ_STATS_LEN);
2144         default:
2145                 return -EOPNOTSUPP;
2146         }
2147 }
2148
2149 static void virtnet_get_ethtool_stats(struct net_device *dev,
2150                                       struct ethtool_stats *stats, u64 *data)
2151 {
2152         struct virtnet_info *vi = netdev_priv(dev);
2153         unsigned int idx = 0, start, i, j;
2154         const u8 *stats_base;
2155         size_t offset;
2156
2157         for (i = 0; i < vi->curr_queue_pairs; i++) {
2158                 struct receive_queue *rq = &vi->rq[i];
2159
2160                 stats_base = (u8 *)&rq->stats;
2161                 do {
2162                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2163                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2164                                 offset = virtnet_rq_stats_desc[j].offset;
2165                                 data[idx + j] = *(u64 *)(stats_base + offset);
2166                         }
2167                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2168                 idx += VIRTNET_RQ_STATS_LEN;
2169         }
2170
2171         for (i = 0; i < vi->curr_queue_pairs; i++) {
2172                 struct send_queue *sq = &vi->sq[i];
2173
2174                 stats_base = (u8 *)&sq->stats;
2175                 do {
2176                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2177                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2178                                 offset = virtnet_sq_stats_desc[j].offset;
2179                                 data[idx + j] = *(u64 *)(stats_base + offset);
2180                         }
2181                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2182                 idx += VIRTNET_SQ_STATS_LEN;
2183         }
2184 }
2185
2186 static void virtnet_get_channels(struct net_device *dev,
2187                                  struct ethtool_channels *channels)
2188 {
2189         struct virtnet_info *vi = netdev_priv(dev);
2190
2191         channels->combined_count = vi->curr_queue_pairs;
2192         channels->max_combined = vi->max_queue_pairs;
2193         channels->max_other = 0;
2194         channels->rx_count = 0;
2195         channels->tx_count = 0;
2196         channels->other_count = 0;
2197 }
2198
2199 static int virtnet_set_link_ksettings(struct net_device *dev,
2200                                       const struct ethtool_link_ksettings *cmd)
2201 {
2202         struct virtnet_info *vi = netdev_priv(dev);
2203
2204         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2205                                                   &vi->speed, &vi->duplex);
2206 }
2207
2208 static int virtnet_get_link_ksettings(struct net_device *dev,
2209                                       struct ethtool_link_ksettings *cmd)
2210 {
2211         struct virtnet_info *vi = netdev_priv(dev);
2212
2213         cmd->base.speed = vi->speed;
2214         cmd->base.duplex = vi->duplex;
2215         cmd->base.port = PORT_OTHER;
2216
2217         return 0;
2218 }
2219
2220 static int virtnet_set_coalesce(struct net_device *dev,
2221                                 struct ethtool_coalesce *ec)
2222 {
2223         struct virtnet_info *vi = netdev_priv(dev);
2224         int i, napi_weight;
2225
2226         if (ec->tx_max_coalesced_frames > 1 ||
2227             ec->rx_max_coalesced_frames != 1)
2228                 return -EINVAL;
2229
2230         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2231         if (napi_weight ^ vi->sq[0].napi.weight) {
2232                 if (dev->flags & IFF_UP)
2233                         return -EBUSY;
2234                 for (i = 0; i < vi->max_queue_pairs; i++)
2235                         vi->sq[i].napi.weight = napi_weight;
2236         }
2237
2238         return 0;
2239 }
2240
2241 static int virtnet_get_coalesce(struct net_device *dev,
2242                                 struct ethtool_coalesce *ec)
2243 {
2244         struct ethtool_coalesce ec_default = {
2245                 .cmd = ETHTOOL_GCOALESCE,
2246                 .rx_max_coalesced_frames = 1,
2247         };
2248         struct virtnet_info *vi = netdev_priv(dev);
2249
2250         memcpy(ec, &ec_default, sizeof(ec_default));
2251
2252         if (vi->sq[0].napi.weight)
2253                 ec->tx_max_coalesced_frames = 1;
2254
2255         return 0;
2256 }
2257
2258 static void virtnet_init_settings(struct net_device *dev)
2259 {
2260         struct virtnet_info *vi = netdev_priv(dev);
2261
2262         vi->speed = SPEED_UNKNOWN;
2263         vi->duplex = DUPLEX_UNKNOWN;
2264 }
2265
2266 static void virtnet_update_settings(struct virtnet_info *vi)
2267 {
2268         u32 speed;
2269         u8 duplex;
2270
2271         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2272                 return;
2273
2274         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2275
2276         if (ethtool_validate_speed(speed))
2277                 vi->speed = speed;
2278
2279         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2280
2281         if (ethtool_validate_duplex(duplex))
2282                 vi->duplex = duplex;
2283 }
2284
2285 static const struct ethtool_ops virtnet_ethtool_ops = {
2286         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2287         .get_drvinfo = virtnet_get_drvinfo,
2288         .get_link = ethtool_op_get_link,
2289         .get_ringparam = virtnet_get_ringparam,
2290         .get_strings = virtnet_get_strings,
2291         .get_sset_count = virtnet_get_sset_count,
2292         .get_ethtool_stats = virtnet_get_ethtool_stats,
2293         .set_channels = virtnet_set_channels,
2294         .get_channels = virtnet_get_channels,
2295         .get_ts_info = ethtool_op_get_ts_info,
2296         .get_link_ksettings = virtnet_get_link_ksettings,
2297         .set_link_ksettings = virtnet_set_link_ksettings,
2298         .set_coalesce = virtnet_set_coalesce,
2299         .get_coalesce = virtnet_get_coalesce,
2300 };
2301
2302 static void virtnet_freeze_down(struct virtio_device *vdev)
2303 {
2304         struct virtnet_info *vi = vdev->priv;
2305         int i;
2306
2307         /* Make sure no work handler is accessing the device */
2308         flush_work(&vi->config_work);
2309
2310         netif_tx_lock_bh(vi->dev);
2311         netif_device_detach(vi->dev);
2312         netif_tx_unlock_bh(vi->dev);
2313         cancel_delayed_work_sync(&vi->refill);
2314
2315         if (netif_running(vi->dev)) {
2316                 for (i = 0; i < vi->max_queue_pairs; i++) {
2317                         napi_disable(&vi->rq[i].napi);
2318                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2319                 }
2320         }
2321 }
2322
2323 static int init_vqs(struct virtnet_info *vi);
2324
2325 static int virtnet_restore_up(struct virtio_device *vdev)
2326 {
2327         struct virtnet_info *vi = vdev->priv;
2328         int err, i;
2329
2330         err = init_vqs(vi);
2331         if (err)
2332                 return err;
2333
2334         virtio_device_ready(vdev);
2335
2336         if (netif_running(vi->dev)) {
2337                 for (i = 0; i < vi->curr_queue_pairs; i++)
2338                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2339                                 schedule_delayed_work(&vi->refill, 0);
2340
2341                 for (i = 0; i < vi->max_queue_pairs; i++) {
2342                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2343                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2344                                                &vi->sq[i].napi);
2345                 }
2346         }
2347
2348         netif_tx_lock_bh(vi->dev);
2349         netif_device_attach(vi->dev);
2350         netif_tx_unlock_bh(vi->dev);
2351         return err;
2352 }
2353
2354 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2355 {
2356         struct scatterlist sg;
2357         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2358
2359         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2360
2361         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2362                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2363                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2364                 return -EINVAL;
2365         }
2366
2367         return 0;
2368 }
2369
2370 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2371 {
2372         u64 offloads = 0;
2373
2374         if (!vi->guest_offloads)
2375                 return 0;
2376
2377         return virtnet_set_guest_offloads(vi, offloads);
2378 }
2379
2380 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2381 {
2382         u64 offloads = vi->guest_offloads;
2383
2384         if (!vi->guest_offloads)
2385                 return 0;
2386
2387         return virtnet_set_guest_offloads(vi, offloads);
2388 }
2389
2390 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2391                            struct netlink_ext_ack *extack)
2392 {
2393         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2394         struct virtnet_info *vi = netdev_priv(dev);
2395         struct bpf_prog *old_prog;
2396         u16 xdp_qp = 0, curr_qp;
2397         int i, err;
2398
2399         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2400             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2401                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2402                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2403                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2404                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2405                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2406                 return -EOPNOTSUPP;
2407         }
2408
2409         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2410                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2411                 return -EINVAL;
2412         }
2413
2414         if (dev->mtu > max_sz) {
2415                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2416                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2417                 return -EINVAL;
2418         }
2419
2420         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2421         if (prog)
2422                 xdp_qp = nr_cpu_ids;
2423
2424         /* XDP requires extra queues for XDP_TX */
2425         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2426                 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2427                 netdev_warn(dev, "request %i queues but max is %i\n",
2428                             curr_qp + xdp_qp, vi->max_queue_pairs);
2429                 return -ENOMEM;
2430         }
2431
2432         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2433         if (!prog && !old_prog)
2434                 return 0;
2435
2436         if (prog)
2437                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2438
2439         /* Make sure NAPI is not using any XDP TX queues for RX. */
2440         if (netif_running(dev)) {
2441                 for (i = 0; i < vi->max_queue_pairs; i++) {
2442                         napi_disable(&vi->rq[i].napi);
2443                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2444                 }
2445         }
2446
2447         if (!prog) {
2448                 for (i = 0; i < vi->max_queue_pairs; i++) {
2449                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2450                         if (i == 0)
2451                                 virtnet_restore_guest_offloads(vi);
2452                 }
2453                 synchronize_net();
2454         }
2455
2456         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2457         if (err)
2458                 goto err;
2459         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2460         vi->xdp_queue_pairs = xdp_qp;
2461
2462         if (prog) {
2463                 for (i = 0; i < vi->max_queue_pairs; i++) {
2464                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2465                         if (i == 0 && !old_prog)
2466                                 virtnet_clear_guest_offloads(vi);
2467                 }
2468         }
2469
2470         for (i = 0; i < vi->max_queue_pairs; i++) {
2471                 if (old_prog)
2472                         bpf_prog_put(old_prog);
2473                 if (netif_running(dev)) {
2474                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2475                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2476                                                &vi->sq[i].napi);
2477                 }
2478         }
2479
2480         return 0;
2481
2482 err:
2483         if (!prog) {
2484                 virtnet_clear_guest_offloads(vi);
2485                 for (i = 0; i < vi->max_queue_pairs; i++)
2486                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2487         }
2488
2489         if (netif_running(dev)) {
2490                 for (i = 0; i < vi->max_queue_pairs; i++) {
2491                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2492                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2493                                                &vi->sq[i].napi);
2494                 }
2495         }
2496         if (prog)
2497                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2498         return err;
2499 }
2500
2501 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2502 {
2503         switch (xdp->command) {
2504         case XDP_SETUP_PROG:
2505                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2506         default:
2507                 return -EINVAL;
2508         }
2509 }
2510
2511 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2512                                       size_t len)
2513 {
2514         struct virtnet_info *vi = netdev_priv(dev);
2515         int ret;
2516
2517         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2518                 return -EOPNOTSUPP;
2519
2520         ret = snprintf(buf, len, "sby");
2521         if (ret >= len)
2522                 return -EOPNOTSUPP;
2523
2524         return 0;
2525 }
2526
2527 static netdev_features_t virtnet_fix_features(struct net_device *netdev,
2528                                               netdev_features_t features)
2529 {
2530         /* If Rx checksum is disabled, LRO should also be disabled. */
2531         if (!(features & NETIF_F_RXCSUM))
2532                 features &= ~NETIF_F_LRO;
2533
2534         return features;
2535 }
2536
2537 static int virtnet_set_features(struct net_device *dev,
2538                                 netdev_features_t features)
2539 {
2540         struct virtnet_info *vi = netdev_priv(dev);
2541         u64 offloads = vi->guest_offloads;
2542         int err;
2543
2544         /* Don't allow configuration while XDP is active. */
2545         if (vi->xdp_queue_pairs)
2546                 return -EBUSY;
2547
2548         if ((dev->features ^ features) & NETIF_F_LRO) {
2549                 if (features & NETIF_F_LRO)
2550                         offloads |= GUEST_OFFLOAD_LRO_MASK &
2551                                     vi->guest_offloads_capable;
2552                 else
2553                         offloads &= ~GUEST_OFFLOAD_LRO_MASK;
2554         }
2555
2556         if ((dev->features ^ features) & NETIF_F_RXCSUM) {
2557                 if (features & NETIF_F_RXCSUM)
2558                         offloads |= GUEST_OFFLOAD_CSUM_MASK &
2559                                     vi->guest_offloads_capable;
2560                 else
2561                         offloads &= ~GUEST_OFFLOAD_CSUM_MASK;
2562         }
2563
2564         err = virtnet_set_guest_offloads(vi, offloads);
2565         if (err)
2566                 return err;
2567
2568         vi->guest_offloads = offloads;
2569         return 0;
2570 }
2571
2572 static const struct net_device_ops virtnet_netdev = {
2573         .ndo_open            = virtnet_open,
2574         .ndo_stop            = virtnet_close,
2575         .ndo_start_xmit      = start_xmit,
2576         .ndo_validate_addr   = eth_validate_addr,
2577         .ndo_set_mac_address = virtnet_set_mac_address,
2578         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2579         .ndo_get_stats64     = virtnet_stats,
2580         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2581         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2582         .ndo_bpf                = virtnet_xdp,
2583         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2584         .ndo_features_check     = passthru_features_check,
2585         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2586         .ndo_set_features       = virtnet_set_features,
2587         .ndo_fix_features       = virtnet_fix_features,
2588 };
2589
2590 static void virtnet_config_changed_work(struct work_struct *work)
2591 {
2592         struct virtnet_info *vi =
2593                 container_of(work, struct virtnet_info, config_work);
2594         u16 v;
2595
2596         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2597                                  struct virtio_net_config, status, &v) < 0)
2598                 return;
2599
2600         if (v & VIRTIO_NET_S_ANNOUNCE) {
2601                 netdev_notify_peers(vi->dev);
2602                 virtnet_ack_link_announce(vi);
2603         }
2604
2605         /* Ignore unknown (future) status bits */
2606         v &= VIRTIO_NET_S_LINK_UP;
2607
2608         if (vi->status == v)
2609                 return;
2610
2611         vi->status = v;
2612
2613         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2614                 virtnet_update_settings(vi);
2615                 netif_carrier_on(vi->dev);
2616                 netif_tx_wake_all_queues(vi->dev);
2617         } else {
2618                 netif_carrier_off(vi->dev);
2619                 netif_tx_stop_all_queues(vi->dev);
2620         }
2621 }
2622
2623 static void virtnet_config_changed(struct virtio_device *vdev)
2624 {
2625         struct virtnet_info *vi = vdev->priv;
2626
2627         schedule_work(&vi->config_work);
2628 }
2629
2630 static void virtnet_free_queues(struct virtnet_info *vi)
2631 {
2632         int i;
2633
2634         for (i = 0; i < vi->max_queue_pairs; i++) {
2635                 __netif_napi_del(&vi->rq[i].napi);
2636                 __netif_napi_del(&vi->sq[i].napi);
2637         }
2638
2639         /* We called __netif_napi_del(),
2640          * we need to respect an RCU grace period before freeing vi->rq
2641          */
2642         synchronize_net();
2643
2644         kfree(vi->rq);
2645         kfree(vi->sq);
2646         kfree(vi->ctrl);
2647 }
2648
2649 static void _free_receive_bufs(struct virtnet_info *vi)
2650 {
2651         struct bpf_prog *old_prog;
2652         int i;
2653
2654         for (i = 0; i < vi->max_queue_pairs; i++) {
2655                 while (vi->rq[i].pages)
2656                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2657
2658                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2659                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2660                 if (old_prog)
2661                         bpf_prog_put(old_prog);
2662         }
2663 }
2664
2665 static void free_receive_bufs(struct virtnet_info *vi)
2666 {
2667         rtnl_lock();
2668         _free_receive_bufs(vi);
2669         rtnl_unlock();
2670 }
2671
2672 static void free_receive_page_frags(struct virtnet_info *vi)
2673 {
2674         int i;
2675         for (i = 0; i < vi->max_queue_pairs; i++)
2676                 if (vi->rq[i].alloc_frag.page)
2677                         put_page(vi->rq[i].alloc_frag.page);
2678 }
2679
2680 static void free_unused_bufs(struct virtnet_info *vi)
2681 {
2682         void *buf;
2683         int i;
2684
2685         for (i = 0; i < vi->max_queue_pairs; i++) {
2686                 struct virtqueue *vq = vi->sq[i].vq;
2687                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2688                         if (!is_xdp_frame(buf))
2689                                 dev_kfree_skb(buf);
2690                         else
2691                                 xdp_return_frame(ptr_to_xdp(buf));
2692                 }
2693         }
2694
2695         for (i = 0; i < vi->max_queue_pairs; i++) {
2696                 struct virtqueue *vq = vi->rq[i].vq;
2697
2698                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2699                         if (vi->mergeable_rx_bufs) {
2700                                 put_page(virt_to_head_page(buf));
2701                         } else if (vi->big_packets) {
2702                                 give_pages(&vi->rq[i], buf);
2703                         } else {
2704                                 put_page(virt_to_head_page(buf));
2705                         }
2706                 }
2707         }
2708 }
2709
2710 static void virtnet_del_vqs(struct virtnet_info *vi)
2711 {
2712         struct virtio_device *vdev = vi->vdev;
2713
2714         virtnet_clean_affinity(vi);
2715
2716         vdev->config->del_vqs(vdev);
2717
2718         virtnet_free_queues(vi);
2719 }
2720
2721 /* How large should a single buffer be so a queue full of these can fit at
2722  * least one full packet?
2723  * Logic below assumes the mergeable buffer header is used.
2724  */
2725 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2726 {
2727         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2728         unsigned int rq_size = virtqueue_get_vring_size(vq);
2729         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2730         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2731         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2732
2733         return max(max(min_buf_len, hdr_len) - hdr_len,
2734                    (unsigned int)GOOD_PACKET_LEN);
2735 }
2736
2737 static int virtnet_find_vqs(struct virtnet_info *vi)
2738 {
2739         vq_callback_t **callbacks;
2740         struct virtqueue **vqs;
2741         int ret = -ENOMEM;
2742         int i, total_vqs;
2743         const char **names;
2744         bool *ctx;
2745
2746         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2747          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2748          * possible control vq.
2749          */
2750         total_vqs = vi->max_queue_pairs * 2 +
2751                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2752
2753         /* Allocate space for find_vqs parameters */
2754         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2755         if (!vqs)
2756                 goto err_vq;
2757         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2758         if (!callbacks)
2759                 goto err_callback;
2760         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2761         if (!names)
2762                 goto err_names;
2763         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2764                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2765                 if (!ctx)
2766                         goto err_ctx;
2767         } else {
2768                 ctx = NULL;
2769         }
2770
2771         /* Parameters for control virtqueue, if any */
2772         if (vi->has_cvq) {
2773                 callbacks[total_vqs - 1] = NULL;
2774                 names[total_vqs - 1] = "control";
2775         }
2776
2777         /* Allocate/initialize parameters for send/receive virtqueues */
2778         for (i = 0; i < vi->max_queue_pairs; i++) {
2779                 callbacks[rxq2vq(i)] = skb_recv_done;
2780                 callbacks[txq2vq(i)] = skb_xmit_done;
2781                 sprintf(vi->rq[i].name, "input.%d", i);
2782                 sprintf(vi->sq[i].name, "output.%d", i);
2783                 names[rxq2vq(i)] = vi->rq[i].name;
2784                 names[txq2vq(i)] = vi->sq[i].name;
2785                 if (ctx)
2786                         ctx[rxq2vq(i)] = true;
2787         }
2788
2789         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2790                                          names, ctx, NULL);
2791         if (ret)
2792                 goto err_find;
2793
2794         if (vi->has_cvq) {
2795                 vi->cvq = vqs[total_vqs - 1];
2796                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2797                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2798         }
2799
2800         for (i = 0; i < vi->max_queue_pairs; i++) {
2801                 vi->rq[i].vq = vqs[rxq2vq(i)];
2802                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2803                 vi->sq[i].vq = vqs[txq2vq(i)];
2804         }
2805
2806         /* run here: ret == 0. */
2807
2808
2809 err_find:
2810         kfree(ctx);
2811 err_ctx:
2812         kfree(names);
2813 err_names:
2814         kfree(callbacks);
2815 err_callback:
2816         kfree(vqs);
2817 err_vq:
2818         return ret;
2819 }
2820
2821 static int virtnet_alloc_queues(struct virtnet_info *vi)
2822 {
2823         int i;
2824
2825         vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2826         if (!vi->ctrl)
2827                 goto err_ctrl;
2828         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2829         if (!vi->sq)
2830                 goto err_sq;
2831         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2832         if (!vi->rq)
2833                 goto err_rq;
2834
2835         INIT_DELAYED_WORK(&vi->refill, refill_work);
2836         for (i = 0; i < vi->max_queue_pairs; i++) {
2837                 vi->rq[i].pages = NULL;
2838                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2839                                napi_weight);
2840                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2841                                   napi_tx ? napi_weight : 0);
2842
2843                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2844                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2845                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2846
2847                 u64_stats_init(&vi->rq[i].stats.syncp);
2848                 u64_stats_init(&vi->sq[i].stats.syncp);
2849         }
2850
2851         return 0;
2852
2853 err_rq:
2854         kfree(vi->sq);
2855 err_sq:
2856         kfree(vi->ctrl);
2857 err_ctrl:
2858         return -ENOMEM;
2859 }
2860
2861 static int init_vqs(struct virtnet_info *vi)
2862 {
2863         int ret;
2864
2865         /* Allocate send & receive queues */
2866         ret = virtnet_alloc_queues(vi);
2867         if (ret)
2868                 goto err;
2869
2870         ret = virtnet_find_vqs(vi);
2871         if (ret)
2872                 goto err_free;
2873
2874         get_online_cpus();
2875         virtnet_set_affinity(vi);
2876         put_online_cpus();
2877
2878         return 0;
2879
2880 err_free:
2881         virtnet_free_queues(vi);
2882 err:
2883         return ret;
2884 }
2885
2886 #ifdef CONFIG_SYSFS
2887 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2888                 char *buf)
2889 {
2890         struct virtnet_info *vi = netdev_priv(queue->dev);
2891         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2892         unsigned int headroom = virtnet_get_headroom(vi);
2893         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2894         struct ewma_pkt_len *avg;
2895
2896         BUG_ON(queue_index >= vi->max_queue_pairs);
2897         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2898         return sprintf(buf, "%u\n",
2899                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
2900                                        SKB_DATA_ALIGN(headroom + tailroom)));
2901 }
2902
2903 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2904         __ATTR_RO(mergeable_rx_buffer_size);
2905
2906 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2907         &mergeable_rx_buffer_size_attribute.attr,
2908         NULL
2909 };
2910
2911 static const struct attribute_group virtio_net_mrg_rx_group = {
2912         .name = "virtio_net",
2913         .attrs = virtio_net_mrg_rx_attrs
2914 };
2915 #endif
2916
2917 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2918                                     unsigned int fbit,
2919                                     const char *fname, const char *dname)
2920 {
2921         if (!virtio_has_feature(vdev, fbit))
2922                 return false;
2923
2924         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2925                 fname, dname);
2926
2927         return true;
2928 }
2929
2930 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2931         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2932
2933 static bool virtnet_validate_features(struct virtio_device *vdev)
2934 {
2935         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2936             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2937                              "VIRTIO_NET_F_CTRL_VQ") ||
2938              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2939                              "VIRTIO_NET_F_CTRL_VQ") ||
2940              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2941                              "VIRTIO_NET_F_CTRL_VQ") ||
2942              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2943              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2944                              "VIRTIO_NET_F_CTRL_VQ"))) {
2945                 return false;
2946         }
2947
2948         return true;
2949 }
2950
2951 #define MIN_MTU ETH_MIN_MTU
2952 #define MAX_MTU ETH_MAX_MTU
2953
2954 static int virtnet_validate(struct virtio_device *vdev)
2955 {
2956         if (!vdev->config->get) {
2957                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2958                         __func__);
2959                 return -EINVAL;
2960         }
2961
2962         if (!virtnet_validate_features(vdev))
2963                 return -EINVAL;
2964
2965         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2966                 int mtu = virtio_cread16(vdev,
2967                                          offsetof(struct virtio_net_config,
2968                                                   mtu));
2969                 if (mtu < MIN_MTU)
2970                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2971         }
2972
2973         return 0;
2974 }
2975
2976 static int virtnet_probe(struct virtio_device *vdev)
2977 {
2978         int i, err = -ENOMEM;
2979         struct net_device *dev;
2980         struct virtnet_info *vi;
2981         u16 max_queue_pairs;
2982         int mtu;
2983
2984         /* Find if host supports multiqueue virtio_net device */
2985         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2986                                    struct virtio_net_config,
2987                                    max_virtqueue_pairs, &max_queue_pairs);
2988
2989         /* We need at least 2 queue's */
2990         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2991             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2992             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2993                 max_queue_pairs = 1;
2994
2995         /* Allocate ourselves a network device with room for our info */
2996         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2997         if (!dev)
2998                 return -ENOMEM;
2999
3000         /* Set up network device as normal. */
3001         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
3002         dev->netdev_ops = &virtnet_netdev;
3003         dev->features = NETIF_F_HIGHDMA;
3004
3005         dev->ethtool_ops = &virtnet_ethtool_ops;
3006         SET_NETDEV_DEV(dev, &vdev->dev);
3007
3008         /* Do we support "hardware" checksums? */
3009         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3010                 /* This opens up the world of extra features. */
3011                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3012                 if (csum)
3013                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3014
3015                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3016                         dev->hw_features |= NETIF_F_TSO
3017                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3018                 }
3019                 /* Individual feature bits: what can host handle? */
3020                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3021                         dev->hw_features |= NETIF_F_TSO;
3022                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3023                         dev->hw_features |= NETIF_F_TSO6;
3024                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3025                         dev->hw_features |= NETIF_F_TSO_ECN;
3026
3027                 dev->features |= NETIF_F_GSO_ROBUST;
3028
3029                 if (gso)
3030                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3031                 /* (!csum && gso) case will be fixed by register_netdev() */
3032         }
3033         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3034                 dev->features |= NETIF_F_RXCSUM;
3035         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3036             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3037                 dev->features |= NETIF_F_LRO;
3038         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) {
3039                 dev->hw_features |= NETIF_F_RXCSUM;
3040                 dev->hw_features |= NETIF_F_LRO;
3041         }
3042
3043         dev->vlan_features = dev->features;
3044
3045         /* MTU range: 68 - 65535 */
3046         dev->min_mtu = MIN_MTU;
3047         dev->max_mtu = MAX_MTU;
3048
3049         /* Configuration may specify what MAC to use.  Otherwise random. */
3050         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3051                 virtio_cread_bytes(vdev,
3052                                    offsetof(struct virtio_net_config, mac),
3053                                    dev->dev_addr, dev->addr_len);
3054         else
3055                 eth_hw_addr_random(dev);
3056
3057         /* Set up our device-specific information */
3058         vi = netdev_priv(dev);
3059         vi->dev = dev;
3060         vi->vdev = vdev;
3061         vdev->priv = vi;
3062
3063         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3064
3065         /* If we can receive ANY GSO packets, we must allocate large ones. */
3066         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3067             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3068             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3069             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3070                 vi->big_packets = true;
3071
3072         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3073                 vi->mergeable_rx_bufs = true;
3074
3075         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3076             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3077                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3078         else
3079                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3080
3081         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3082             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3083                 vi->any_header_sg = true;
3084
3085         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3086                 vi->has_cvq = true;
3087
3088         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3089                 mtu = virtio_cread16(vdev,
3090                                      offsetof(struct virtio_net_config,
3091                                               mtu));
3092                 if (mtu < dev->min_mtu) {
3093                         /* Should never trigger: MTU was previously validated
3094                          * in virtnet_validate.
3095                          */
3096                         dev_err(&vdev->dev,
3097                                 "device MTU appears to have changed it is now %d < %d",
3098                                 mtu, dev->min_mtu);
3099                         goto free;
3100                 }
3101
3102                 dev->mtu = mtu;
3103                 dev->max_mtu = mtu;
3104
3105                 /* TODO: size buffers correctly in this case. */
3106                 if (dev->mtu > ETH_DATA_LEN)
3107                         vi->big_packets = true;
3108         }
3109
3110         if (vi->any_header_sg)
3111                 dev->needed_headroom = vi->hdr_len;
3112
3113         /* Enable multiqueue by default */
3114         if (num_online_cpus() >= max_queue_pairs)
3115                 vi->curr_queue_pairs = max_queue_pairs;
3116         else
3117                 vi->curr_queue_pairs = num_online_cpus();
3118         vi->max_queue_pairs = max_queue_pairs;
3119
3120         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3121         err = init_vqs(vi);
3122         if (err)
3123                 goto free;
3124
3125 #ifdef CONFIG_SYSFS
3126         if (vi->mergeable_rx_bufs)
3127                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3128 #endif
3129         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3130         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3131
3132         virtnet_init_settings(dev);
3133
3134         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3135                 vi->failover = net_failover_create(vi->dev);
3136                 if (IS_ERR(vi->failover)) {
3137                         err = PTR_ERR(vi->failover);
3138                         goto free_vqs;
3139                 }
3140         }
3141
3142         err = register_netdev(dev);
3143         if (err) {
3144                 pr_debug("virtio_net: registering device failed\n");
3145                 goto free_failover;
3146         }
3147
3148         virtio_device_ready(vdev);
3149
3150         err = virtnet_cpu_notif_add(vi);
3151         if (err) {
3152                 pr_debug("virtio_net: registering cpu notifier failed\n");
3153                 goto free_unregister_netdev;
3154         }
3155
3156         virtnet_set_queues(vi, vi->curr_queue_pairs);
3157
3158         /* Assume link up if device can't report link status,
3159            otherwise get link status from config. */
3160         netif_carrier_off(dev);
3161         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3162                 schedule_work(&vi->config_work);
3163         } else {
3164                 vi->status = VIRTIO_NET_S_LINK_UP;
3165                 virtnet_update_settings(vi);
3166                 netif_carrier_on(dev);
3167         }
3168
3169         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3170                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3171                         set_bit(guest_offloads[i], &vi->guest_offloads);
3172         vi->guest_offloads_capable = vi->guest_offloads;
3173
3174         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3175                  dev->name, max_queue_pairs);
3176
3177         return 0;
3178
3179 free_unregister_netdev:
3180         vi->vdev->config->reset(vdev);
3181
3182         unregister_netdev(dev);
3183 free_failover:
3184         net_failover_destroy(vi->failover);
3185 free_vqs:
3186         cancel_delayed_work_sync(&vi->refill);
3187         free_receive_page_frags(vi);
3188         virtnet_del_vqs(vi);
3189 free:
3190         free_netdev(dev);
3191         return err;
3192 }
3193
3194 static void remove_vq_common(struct virtnet_info *vi)
3195 {
3196         vi->vdev->config->reset(vi->vdev);
3197
3198         /* Free unused buffers in both send and recv, if any. */
3199         free_unused_bufs(vi);
3200
3201         free_receive_bufs(vi);
3202
3203         free_receive_page_frags(vi);
3204
3205         virtnet_del_vqs(vi);
3206 }
3207
3208 static void virtnet_remove(struct virtio_device *vdev)
3209 {
3210         struct virtnet_info *vi = vdev->priv;
3211
3212         virtnet_cpu_notif_remove(vi);
3213
3214         /* Make sure no work handler is accessing the device. */
3215         flush_work(&vi->config_work);
3216
3217         unregister_netdev(vi->dev);
3218
3219         net_failover_destroy(vi->failover);
3220
3221         remove_vq_common(vi);
3222
3223         free_netdev(vi->dev);
3224 }
3225
3226 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3227 {
3228         struct virtnet_info *vi = vdev->priv;
3229
3230         virtnet_cpu_notif_remove(vi);
3231         virtnet_freeze_down(vdev);
3232         remove_vq_common(vi);
3233
3234         return 0;
3235 }
3236
3237 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3238 {
3239         struct virtnet_info *vi = vdev->priv;
3240         int err;
3241
3242         err = virtnet_restore_up(vdev);
3243         if (err)
3244                 return err;
3245         virtnet_set_queues(vi, vi->curr_queue_pairs);
3246
3247         err = virtnet_cpu_notif_add(vi);
3248         if (err)
3249                 return err;
3250
3251         return 0;
3252 }
3253
3254 static struct virtio_device_id id_table[] = {
3255         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3256         { 0 },
3257 };
3258
3259 #define VIRTNET_FEATURES \
3260         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3261         VIRTIO_NET_F_MAC, \
3262         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3263         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3264         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3265         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3266         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3267         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3268         VIRTIO_NET_F_CTRL_MAC_ADDR, \
3269         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3270         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3271
3272 static unsigned int features[] = {
3273         VIRTNET_FEATURES,
3274 };
3275
3276 static unsigned int features_legacy[] = {
3277         VIRTNET_FEATURES,
3278         VIRTIO_NET_F_GSO,
3279         VIRTIO_F_ANY_LAYOUT,
3280 };
3281
3282 static struct virtio_driver virtio_net_driver = {
3283         .feature_table = features,
3284         .feature_table_size = ARRAY_SIZE(features),
3285         .feature_table_legacy = features_legacy,
3286         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3287         .driver.name =  KBUILD_MODNAME,
3288         .driver.owner = THIS_MODULE,
3289         .id_table =     id_table,
3290         .validate =     virtnet_validate,
3291         .probe =        virtnet_probe,
3292         .remove =       virtnet_remove,
3293         .config_changed = virtnet_config_changed,
3294 #ifdef CONFIG_PM_SLEEP
3295         .freeze =       virtnet_freeze,
3296         .restore =      virtnet_restore,
3297 #endif
3298 };
3299
3300 static __init int virtio_net_driver_init(void)
3301 {
3302         int ret;
3303
3304         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3305                                       virtnet_cpu_online,
3306                                       virtnet_cpu_down_prep);
3307         if (ret < 0)
3308                 goto out;
3309         virtionet_online = ret;
3310         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3311                                       NULL, virtnet_cpu_dead);
3312         if (ret)
3313                 goto err_dead;
3314
3315         ret = register_virtio_driver(&virtio_net_driver);
3316         if (ret)
3317                 goto err_virtio;
3318         return 0;
3319 err_virtio:
3320         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3321 err_dead:
3322         cpuhp_remove_multi_state(virtionet_online);
3323 out:
3324         return ret;
3325 }
3326 module_init(virtio_net_driver_init);
3327
3328 static __exit void virtio_net_driver_exit(void)
3329 {
3330         unregister_virtio_driver(&virtio_net_driver);
3331         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3332         cpuhp_remove_multi_state(virtionet_online);
3333 }
3334 module_exit(virtio_net_driver_exit);
3335
3336 MODULE_DEVICE_TABLE(virtio, id_table);
3337 MODULE_DESCRIPTION("Virtio network driver");
3338 MODULE_LICENSE("GPL");