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