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_GRO_HW_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         /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
 199         bool xdp_enabled;
 200
 201         /* I like... big packets and I cannot lie! */
 202         bool big_packets;
 203
 204         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 205         bool mergeable_rx_bufs;
 206
 207         /* Has control virtqueue */
 208         bool has_cvq;
 209
 210         /* Host can handle any s/g split between our header and packet data */
 211         bool any_header_sg;
 212
 213         /* Packet virtio header size */
 214         u8 hdr_len;
 215
 216         /* Work struct for refilling if we run low on memory. */
 217         struct delayed_work refill;
 218
 219         /* Work struct for config space updates */
 220         struct work_struct config_work;
 221
 222         /* Does the affinity hint is set for virtqueues? */
 223         bool affinity_hint_set;
 224
 225         /* CPU hotplug instances for online & dead */
 226         struct hlist_node node;
 227         struct hlist_node node_dead;
 228
 229         struct control_buf *ctrl;
 230
 231         /* Ethtool settings */
 232         u8 duplex;
 233         u32 speed;
 234
 235         unsigned long guest_offloads;
 236         unsigned long guest_offloads_capable;
 237
 238         /* failover when STANDBY feature enabled */
 239         struct failover *failover;
 240 };
 241
 242 struct padded_vnet_hdr {
 243         struct virtio_net_hdr_mrg_rxbuf hdr;
 244         /*
 245          * hdr is in a separate sg buffer, and data sg buffer shares same page
 246          * with this header sg. This padding makes next sg 16 byte aligned
 247          * after the header.
 248          */
 249         char padding[4];
 250 };
 251
 252 static bool is_xdp_frame(void *ptr)
 253 {
 254         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 255 }
 256
 257 static void *xdp_to_ptr(struct xdp_frame *ptr)
 258 {
 259         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 260 }
 261
 262 static struct xdp_frame *ptr_to_xdp(void *ptr)
 263 {
 264         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 265 }
 266
 267 /* Converting between virtqueue no. and kernel tx/rx queue no.
 268  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 269  */
 270 static int vq2txq(struct virtqueue *vq)
 271 {
 272         return (vq->index - 1) / 2;
 273 }
 274
 275 static int txq2vq(int txq)
 276 {
 277         return txq * 2 + 1;
 278 }
 279
 280 static int vq2rxq(struct virtqueue *vq)
 281 {
 282         return vq->index / 2;
 283 }
 284
 285 static int rxq2vq(int rxq)
 286 {
 287         return rxq * 2;
 288 }
 289
 290 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 291 {
 292         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 293 }
 294
 295 /*
 296  * private is used to chain pages for big packets, put the whole
 297  * most recent used list in the beginning for reuse
 298  */
 299 static void give_pages(struct receive_queue *rq, struct page *page)
 300 {
 301         struct page *end;
 302
 303         /* Find end of list, sew whole thing into vi->rq.pages. */
 304         for (end = page; end->private; end = (struct page *)end->private);
 305         end->private = (unsigned long)rq->pages;
 306         rq->pages = page;
 307 }
 308
 309 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 310 {
 311         struct page *p = rq->pages;
 312
 313         if (p) {
 314                 rq->pages = (struct page *)p->private;
 315                 /* clear private here, it is used to chain pages */
 316                 p->private = 0;
 317         } else
 318                 p = alloc_page(gfp_mask);
 319         return p;
 320 }
 321
 322 static void virtqueue_napi_schedule(struct napi_struct *napi,
 323                                     struct virtqueue *vq)
 324 {
 325         if (napi_schedule_prep(napi)) {
 326                 virtqueue_disable_cb(vq);
 327                 __napi_schedule(napi);
 328         }
 329 }
 330
 331 static void virtqueue_napi_complete(struct napi_struct *napi,
 332                                     struct virtqueue *vq, int processed)
 333 {
 334         int opaque;
 335
 336         opaque = virtqueue_enable_cb_prepare(vq);
 337         if (napi_complete_done(napi, processed)) {
 338                 if (unlikely(virtqueue_poll(vq, opaque)))
 339                         virtqueue_napi_schedule(napi, vq);
 340         } else {
 341                 virtqueue_disable_cb(vq);
 342         }
 343 }
 344
 345 static void skb_xmit_done(struct virtqueue *vq)
 346 {
 347         struct virtnet_info *vi = vq->vdev->priv;
 348         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 349
 350         /* Suppress further interrupts. */
 351         virtqueue_disable_cb(vq);
 352
 353         if (napi->weight)
 354                 virtqueue_napi_schedule(napi, vq);
 355         else
 356                 /* We were probably waiting for more output buffers. */
 357                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 358 }
 359
 360 #define MRG_CTX_HEADER_SHIFT 22
 361 static void *mergeable_len_to_ctx(unsigned int truesize,
 362                                   unsigned int headroom)
 363 {
 364         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 365 }
 366
 367 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 368 {
 369         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 370 }
 371
 372 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 373 {
 374         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 375 }
 376
 377 /* Called from bottom half context */
 378 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 379                                    struct receive_queue *rq,
 380                                    struct page *page, unsigned int offset,
 381                                    unsigned int len, unsigned int truesize,
 382                                    bool hdr_valid, unsigned int metasize,
 383                                    unsigned int headroom)
 384 {
 385         struct sk_buff *skb;
 386         struct virtio_net_hdr_mrg_rxbuf *hdr;
 387         unsigned int copy, hdr_len, hdr_padded_len;
 388         struct page *page_to_free = NULL;
 389         int tailroom, shinfo_size;
 390         char *p, *hdr_p, *buf;
 391
 392         p = page_address(page) + offset;
 393         hdr_p = p;
 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         /* If headroom is not 0, there is an offset between the beginning of the
 402          * data and the allocated space, otherwise the data and the allocated
 403          * space are aligned.
 404          *
 405          * Buffers with headroom use PAGE_SIZE as alloc size, see
 406          * add_recvbuf_mergeable() + get_mergeable_buf_len()
 407          */
 408         truesize = headroom ? PAGE_SIZE : truesize;
 409         tailroom = truesize - len - headroom;
 410         buf = p - headroom;
 411
 412         len -= hdr_len;
 413         offset += hdr_padded_len;
 414         p += hdr_padded_len;
 415
 416         shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 417
 418         /* copy small packet so we can reuse these pages */
 419         if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
 420                 skb = build_skb(buf, truesize);
 421                 if (unlikely(!skb))
 422                         return NULL;
 423
 424                 skb_reserve(skb, p - buf);
 425                 skb_put(skb, len);
 426                 goto ok;
 427         }
 428
 429         /* copy small packet so we can reuse these pages for small data */
 430         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 431         if (unlikely(!skb))
 432                 return NULL;
 433
 434         /* Copy all frame if it fits skb->head, otherwise
 435          * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
 436          */
 437         if (len <= skb_tailroom(skb))
 438                 copy = len;
 439         else
 440                 copy = ETH_HLEN + metasize;
 441         skb_put_data(skb, p, copy);
 442
 443         len -= copy;
 444         offset += copy;
 445
 446         if (vi->mergeable_rx_bufs) {
 447                 if (len)
 448                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 449                 else
 450                         page_to_free = page;
 451                 goto ok;
 452         }
 453
 454         /*
 455          * Verify that we can indeed put this data into a skb.
 456          * This is here to handle cases when the device erroneously
 457          * tries to receive more than is possible. This is usually
 458          * the case of a broken device.
 459          */
 460         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 461                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 462                 dev_kfree_skb(skb);
 463                 return NULL;
 464         }
 465         BUG_ON(offset >= PAGE_SIZE);
 466         while (len) {
 467                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 468                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 469                                 frag_size, truesize);
 470                 len -= frag_size;
 471                 page = (struct page *)page->private;
 472                 offset = 0;
 473         }
 474
 475         if (page)
 476                 give_pages(rq, page);
 477
 478 ok:
 479         /* hdr_valid means no XDP, so we can copy the vnet header */
 480         if (hdr_valid) {
 481                 hdr = skb_vnet_hdr(skb);
 482                 memcpy(hdr, hdr_p, hdr_len);
 483         }
 484         if (page_to_free)
 485                 put_page(page_to_free);
 486
 487         if (metasize) {
 488                 __skb_pull(skb, metasize);
 489                 skb_metadata_set(skb, metasize);
 490         }
 491
 492         return skb;
 493 }
 494
 495 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 496                                    struct send_queue *sq,
 497                                    struct xdp_frame *xdpf)
 498 {
 499         struct virtio_net_hdr_mrg_rxbuf *hdr;
 500         int err;
 501
 502         if (unlikely(xdpf->headroom < vi->hdr_len))
 503                 return -EOVERFLOW;
 504
 505         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 506         xdpf->data -= vi->hdr_len;
 507         /* Zero header and leave csum up to XDP layers */
 508         hdr = xdpf->data;
 509         memset(hdr, 0, vi->hdr_len);
 510         xdpf->len   += vi->hdr_len;
 511
 512         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 513
 514         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
 515                                    GFP_ATOMIC);
 516         if (unlikely(err))
 517                 return -ENOSPC; /* Caller handle free/refcnt */
 518
 519         return 0;
 520 }
 521
 522 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
 523  * the current cpu, so it does not need to be locked.
 524  *
 525  * Here we use marco instead of inline functions because we have to deal with
 526  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
 527  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
 528  * functions to perfectly solve these three problems at the same time.
 529  */
 530 #define virtnet_xdp_get_sq(vi) ({                                       \
 531         int cpu = smp_processor_id();                                   \
 532         struct netdev_queue *txq;                                       \
 533         typeof(vi) v = (vi);                                            \
 534         unsigned int qp;                                                \
 535                                                                         \
 536         if (v->curr_queue_pairs > nr_cpu_ids) {                         \
 537                 qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
 538                 qp += cpu;                                              \
 539                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 540                 __netif_tx_acquire(txq);                                \
 541         } else {                                                        \
 542                 qp = cpu % v->curr_queue_pairs;                         \
 543                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 544                 __netif_tx_lock(txq, cpu);                              \
 545         }                                                               \
 546         v->sq + qp;                                                     \
 547 })
 548
 549 #define virtnet_xdp_put_sq(vi, q) {                                     \
 550         struct netdev_queue *txq;                                       \
 551         typeof(vi) v = (vi);                                            \
 552                                                                         \
 553         txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
 554         if (v->curr_queue_pairs > nr_cpu_ids)                           \
 555                 __netif_tx_release(txq);                                \
 556         else                                                            \
 557                 __netif_tx_unlock(txq);                                 \
 558 }
 559
 560 static int virtnet_xdp_xmit(struct net_device *dev,
 561                             int n, struct xdp_frame **frames, u32 flags)
 562 {
 563         struct virtnet_info *vi = netdev_priv(dev);
 564         struct receive_queue *rq = vi->rq;
 565         struct bpf_prog *xdp_prog;
 566         struct send_queue *sq;
 567         unsigned int len;
 568         int packets = 0;
 569         int bytes = 0;
 570         int nxmit = 0;
 571         int kicks = 0;
 572         void *ptr;
 573         int ret;
 574         int i;
 575
 576         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 577          * indicate XDP resources have been successfully allocated.
 578          */
 579         xdp_prog = rcu_access_pointer(rq->xdp_prog);
 580         if (!xdp_prog)
 581                 return -ENXIO;
 582
 583         sq = virtnet_xdp_get_sq(vi);
 584
 585         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 586                 ret = -EINVAL;
 587                 goto out;
 588         }
 589
 590         /* Free up any pending old buffers before queueing new ones. */
 591         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 592                 if (likely(is_xdp_frame(ptr))) {
 593                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 594
 595                         bytes += frame->len;
 596                         xdp_return_frame(frame);
 597                 } else {
 598                         struct sk_buff *skb = ptr;
 599
 600                         bytes += skb->len;
 601                         napi_consume_skb(skb, false);
 602                 }
 603                 packets++;
 604         }
 605
 606         for (i = 0; i < n; i++) {
 607                 struct xdp_frame *xdpf = frames[i];
 608
 609                 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
 610                         break;
 611                 nxmit++;
 612         }
 613         ret = nxmit;
 614
 615         if (flags & XDP_XMIT_FLUSH) {
 616                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
 617                         kicks = 1;
 618         }
 619 out:
 620         u64_stats_update_begin(&sq->stats.syncp);
 621         sq->stats.bytes += bytes;
 622         sq->stats.packets += packets;
 623         sq->stats.xdp_tx += n;
 624         sq->stats.xdp_tx_drops += n - nxmit;
 625         sq->stats.kicks += kicks;
 626         u64_stats_update_end(&sq->stats.syncp);
 627
 628         virtnet_xdp_put_sq(vi, sq);
 629         return ret;
 630 }
 631
 632 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 633 {
 634         return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
 635 }
 636
 637 /* We copy the packet for XDP in the following cases:
 638  *
 639  * 1) Packet is scattered across multiple rx buffers.
 640  * 2) Headroom space is insufficient.
 641  *
 642  * This is inefficient but it's a temporary condition that
 643  * we hit right after XDP is enabled and until queue is refilled
 644  * with large buffers with sufficient headroom - so it should affect
 645  * at most queue size packets.
 646  * Afterwards, the conditions to enable
 647  * XDP should preclude the underlying device from sending packets
 648  * across multiple buffers (num_buf > 1), and we make sure buffers
 649  * have enough headroom.
 650  */
 651 static struct page *xdp_linearize_page(struct receive_queue *rq,
 652                                        u16 *num_buf,
 653                                        struct page *p,
 654                                        int offset,
 655                                        int page_off,
 656                                        unsigned int *len)
 657 {
 658         struct page *page = alloc_page(GFP_ATOMIC);
 659
 660         if (!page)
 661                 return NULL;
 662
 663         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 664         page_off += *len;
 665
 666         while (--*num_buf) {
 667                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 668                 unsigned int buflen;
 669                 void *buf;
 670                 int off;
 671
 672                 buf = virtqueue_get_buf(rq->vq, &buflen);
 673                 if (unlikely(!buf))
 674                         goto err_buf;
 675
 676                 p = virt_to_head_page(buf);
 677                 off = buf - page_address(p);
 678
 679                 /* guard against a misconfigured or uncooperative backend that
 680                  * is sending packet larger than the MTU.
 681                  */
 682                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 683                         put_page(p);
 684                         goto err_buf;
 685                 }
 686
 687                 memcpy(page_address(page) + page_off,
 688                        page_address(p) + off, buflen);
 689                 page_off += buflen;
 690                 put_page(p);
 691         }
 692
 693         /* Headroom does not contribute to packet length */
 694         *len = page_off - VIRTIO_XDP_HEADROOM;
 695         return page;
 696 err_buf:
 697         __free_pages(page, 0);
 698         return NULL;
 699 }
 700
 701 static struct sk_buff *receive_small(struct net_device *dev,
 702                                      struct virtnet_info *vi,
 703                                      struct receive_queue *rq,
 704                                      void *buf, void *ctx,
 705                                      unsigned int len,
 706                                      unsigned int *xdp_xmit,
 707                                      struct virtnet_rq_stats *stats)
 708 {
 709         struct sk_buff *skb;
 710         struct bpf_prog *xdp_prog;
 711         unsigned int xdp_headroom = (unsigned long)ctx;
 712         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 713         unsigned int headroom = vi->hdr_len + header_offset;
 714         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 715                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 716         struct page *page = virt_to_head_page(buf);
 717         unsigned int delta = 0;
 718         struct page *xdp_page;
 719         int err;
 720         unsigned int metasize = 0;
 721
 722         len -= vi->hdr_len;
 723         stats->bytes += len;
 724
 725         if (unlikely(len > GOOD_PACKET_LEN)) {
 726                 pr_debug("%s: rx error: len %u exceeds max size %d\n",
 727                          dev->name, len, GOOD_PACKET_LEN);
 728                 dev->stats.rx_length_errors++;
 729                 goto err_len;
 730         }
 731         rcu_read_lock();
 732         xdp_prog = rcu_dereference(rq->xdp_prog);
 733         if (xdp_prog) {
 734                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 735                 struct xdp_frame *xdpf;
 736                 struct xdp_buff xdp;
 737                 void *orig_data;
 738                 u32 act;
 739
 740                 if (unlikely(hdr->hdr.gso_type))
 741                         goto err_xdp;
 742
 743                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 744                         int offset = buf - page_address(page) + header_offset;
 745                         unsigned int tlen = len + vi->hdr_len;
 746                         u16 num_buf = 1;
 747
 748                         xdp_headroom = virtnet_get_headroom(vi);
 749                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 750                         headroom = vi->hdr_len + header_offset;
 751                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 752                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 753                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 754                                                       offset, header_offset,
 755                                                       &tlen);
 756                         if (!xdp_page)
 757                                 goto err_xdp;
 758
 759                         buf = page_address(xdp_page);
 760                         put_page(page);
 761                         page = xdp_page;
 762                 }
 763
 764                 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
 765                 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
 766                                  xdp_headroom, len, true);
 767                 orig_data = xdp.data;
 768                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 769                 stats->xdp_packets++;
 770
 771                 switch (act) {
 772                 case XDP_PASS:
 773                         /* Recalculate length in case bpf program changed it */
 774                         delta = orig_data - xdp.data;
 775                         len = xdp.data_end - xdp.data;
 776                         metasize = xdp.data - xdp.data_meta;
 777                         break;
 778                 case XDP_TX:
 779                         stats->xdp_tx++;
 780                         xdpf = xdp_convert_buff_to_frame(&xdp);
 781                         if (unlikely(!xdpf))
 782                                 goto err_xdp;
 783                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 784                         if (unlikely(!err)) {
 785                                 xdp_return_frame_rx_napi(xdpf);
 786                         } else if (unlikely(err < 0)) {
 787                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 788                                 goto err_xdp;
 789                         }
 790                         *xdp_xmit |= VIRTIO_XDP_TX;
 791                         rcu_read_unlock();
 792                         goto xdp_xmit;
 793                 case XDP_REDIRECT:
 794                         stats->xdp_redirects++;
 795                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 796                         if (err)
 797                                 goto err_xdp;
 798                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 799                         rcu_read_unlock();
 800                         goto xdp_xmit;
 801                 default:
 802                         bpf_warn_invalid_xdp_action(act);
 803                         fallthrough;
 804                 case XDP_ABORTED:
 805                         trace_xdp_exception(vi->dev, xdp_prog, act);
 806                         goto err_xdp;
 807                 case XDP_DROP:
 808                         goto err_xdp;
 809                 }
 810         }
 811         rcu_read_unlock();
 812
 813         skb = build_skb(buf, buflen);
 814         if (!skb) {
 815                 put_page(page);
 816                 goto err;
 817         }
 818         skb_reserve(skb, headroom - delta);
 819         skb_put(skb, len);
 820         if (!xdp_prog) {
 821                 buf += header_offset;
 822                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 823         } /* keep zeroed vnet hdr since XDP is loaded */
 824
 825         if (metasize)
 826                 skb_metadata_set(skb, metasize);
 827
 828 err:
 829         return skb;
 830
 831 err_xdp:
 832         rcu_read_unlock();
 833         stats->xdp_drops++;
 834 err_len:
 835         stats->drops++;
 836         put_page(page);
 837 xdp_xmit:
 838         return NULL;
 839 }
 840
 841 static struct sk_buff *receive_big(struct net_device *dev,
 842                                    struct virtnet_info *vi,
 843                                    struct receive_queue *rq,
 844                                    void *buf,
 845                                    unsigned int len,
 846                                    struct virtnet_rq_stats *stats)
 847 {
 848         struct page *page = buf;
 849         struct sk_buff *skb =
 850                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
 851
 852         stats->bytes += len - vi->hdr_len;
 853         if (unlikely(!skb))
 854                 goto err;
 855
 856         return skb;
 857
 858 err:
 859         stats->drops++;
 860         give_pages(rq, page);
 861         return NULL;
 862 }
 863
 864 static struct sk_buff *receive_mergeable(struct net_device *dev,
 865                                          struct virtnet_info *vi,
 866                                          struct receive_queue *rq,
 867                                          void *buf,
 868                                          void *ctx,
 869                                          unsigned int len,
 870                                          unsigned int *xdp_xmit,
 871                                          struct virtnet_rq_stats *stats)
 872 {
 873         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 874         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 875         struct page *page = virt_to_head_page(buf);
 876         int offset = buf - page_address(page);
 877         struct sk_buff *head_skb, *curr_skb;
 878         struct bpf_prog *xdp_prog;
 879         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
 880         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 881         unsigned int metasize = 0;
 882         unsigned int frame_sz;
 883         int err;
 884
 885         head_skb = NULL;
 886         stats->bytes += len - vi->hdr_len;
 887
 888         if (unlikely(len > truesize)) {
 889                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 890                          dev->name, len, (unsigned long)ctx);
 891                 dev->stats.rx_length_errors++;
 892                 goto err_skb;
 893         }
 894         rcu_read_lock();
 895         xdp_prog = rcu_dereference(rq->xdp_prog);
 896         if (xdp_prog) {
 897                 struct xdp_frame *xdpf;
 898                 struct page *xdp_page;
 899                 struct xdp_buff xdp;
 900                 void *data;
 901                 u32 act;
 902
 903                 /* Transient failure which in theory could occur if
 904                  * in-flight packets from before XDP was enabled reach
 905                  * the receive path after XDP is loaded.
 906                  */
 907                 if (unlikely(hdr->hdr.gso_type))
 908                         goto err_xdp;
 909
 910                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 911                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 912                  */
 913                 frame_sz = headroom ? PAGE_SIZE : truesize;
 914
 915                 /* This happens when rx buffer size is underestimated
 916                  * or headroom is not enough because of the buffer
 917                  * was refilled before XDP is set. This should only
 918                  * happen for the first several packets, so we don't
 919                  * care much about its performance.
 920                  */
 921                 if (unlikely(num_buf > 1 ||
 922                              headroom < virtnet_get_headroom(vi))) {
 923                         /* linearize data for XDP */
 924                         xdp_page = xdp_linearize_page(rq, &num_buf,
 925                                                       page, offset,
 926                                                       VIRTIO_XDP_HEADROOM,
 927                                                       &len);
 928                         frame_sz = PAGE_SIZE;
 929
 930                         if (!xdp_page)
 931                                 goto err_xdp;
 932                         offset = VIRTIO_XDP_HEADROOM;
 933                 } else {
 934                         xdp_page = page;
 935                 }
 936
 937                 /* Allow consuming headroom but reserve enough space to push
 938                  * the descriptor on if we get an XDP_TX return code.
 939                  */
 940                 data = page_address(xdp_page) + offset;
 941                 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
 942                 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
 943                                  VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
 944
 945                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 946                 stats->xdp_packets++;
 947
 948                 switch (act) {
 949                 case XDP_PASS:
 950                         metasize = xdp.data - xdp.data_meta;
 951
 952                         /* recalculate offset to account for any header
 953                          * adjustments and minus the metasize to copy the
 954                          * metadata in page_to_skb(). Note other cases do not
 955                          * build an skb and avoid using offset
 956                          */
 957                         offset = xdp.data - page_address(xdp_page) -
 958                                  vi->hdr_len - metasize;
 959
 960                         /* recalculate len if xdp.data, xdp.data_end or
 961                          * xdp.data_meta were adjusted
 962                          */
 963                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 964                         /* We can only create skb based on xdp_page. */
 965                         if (unlikely(xdp_page != page)) {
 966                                 rcu_read_unlock();
 967                                 put_page(page);
 968                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
 969                                                        len, PAGE_SIZE, false,
 970                                                        metasize,
 971                                                        VIRTIO_XDP_HEADROOM);
 972                                 return head_skb;
 973                         }
 974                         break;
 975                 case XDP_TX:
 976                         stats->xdp_tx++;
 977                         xdpf = xdp_convert_buff_to_frame(&xdp);
 978                         if (unlikely(!xdpf))
 979                                 goto err_xdp;
 980                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 981                         if (unlikely(!err)) {
 982                                 xdp_return_frame_rx_napi(xdpf);
 983                         } else if (unlikely(err < 0)) {
 984                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 985                                 if (unlikely(xdp_page != page))
 986                                         put_page(xdp_page);
 987                                 goto err_xdp;
 988                         }
 989                         *xdp_xmit |= VIRTIO_XDP_TX;
 990                         if (unlikely(xdp_page != page))
 991                                 put_page(page);
 992                         rcu_read_unlock();
 993                         goto xdp_xmit;
 994                 case XDP_REDIRECT:
 995                         stats->xdp_redirects++;
 996                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 997                         if (err) {
 998                                 if (unlikely(xdp_page != page))
 999                                         put_page(xdp_page);
1000                                 goto err_xdp;
1001                         }
1002                         *xdp_xmit |= VIRTIO_XDP_REDIR;
1003                         if (unlikely(xdp_page != page))
1004                                 put_page(page);
1005                         rcu_read_unlock();
1006                         goto xdp_xmit;
1007                 default:
1008                         bpf_warn_invalid_xdp_action(act);
1009                         fallthrough;
1010                 case XDP_ABORTED:
1011                         trace_xdp_exception(vi->dev, xdp_prog, act);
1012                         fallthrough;
1013                 case XDP_DROP:
1014                         if (unlikely(xdp_page != page))
1015                                 __free_pages(xdp_page, 0);
1016                         goto err_xdp;
1017                 }
1018         }
1019         rcu_read_unlock();
1020
1021         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1022                                metasize, headroom);
1023         curr_skb = head_skb;
1024
1025         if (unlikely(!curr_skb))
1026                 goto err_skb;
1027         while (--num_buf) {
1028                 int num_skb_frags;
1029
1030                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1031                 if (unlikely(!buf)) {
1032                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1033                                  dev->name, num_buf,
1034                                  virtio16_to_cpu(vi->vdev,
1035                                                  hdr->num_buffers));
1036                         dev->stats.rx_length_errors++;
1037                         goto err_buf;
1038                 }
1039
1040                 stats->bytes += len;
1041                 page = virt_to_head_page(buf);
1042
1043                 truesize = mergeable_ctx_to_truesize(ctx);
1044                 if (unlikely(len > truesize)) {
1045                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1046                                  dev->name, len, (unsigned long)ctx);
1047                         dev->stats.rx_length_errors++;
1048                         goto err_skb;
1049                 }
1050
1051                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1052                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1053                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1054
1055                         if (unlikely(!nskb))
1056                                 goto err_skb;
1057                         if (curr_skb == head_skb)
1058                                 skb_shinfo(curr_skb)->frag_list = nskb;
1059                         else
1060                                 curr_skb->next = nskb;
1061                         curr_skb = nskb;
1062                         head_skb->truesize += nskb->truesize;
1063                         num_skb_frags = 0;
1064                 }
1065                 if (curr_skb != head_skb) {
1066                         head_skb->data_len += len;
1067                         head_skb->len += len;
1068                         head_skb->truesize += truesize;
1069                 }
1070                 offset = buf - page_address(page);
1071                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1072                         put_page(page);
1073                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1074                                              len, truesize);
1075                 } else {
1076                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1077                                         offset, len, truesize);
1078                 }
1079         }
1080
1081         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1082         return head_skb;
1083
1084 err_xdp:
1085         rcu_read_unlock();
1086         stats->xdp_drops++;
1087 err_skb:
1088         put_page(page);
1089         while (num_buf-- > 1) {
1090                 buf = virtqueue_get_buf(rq->vq, &len);
1091                 if (unlikely(!buf)) {
1092                         pr_debug("%s: rx error: %d buffers missing\n",
1093                                  dev->name, num_buf);
1094                         dev->stats.rx_length_errors++;
1095                         break;
1096                 }
1097                 stats->bytes += len;
1098                 page = virt_to_head_page(buf);
1099                 put_page(page);
1100         }
1101 err_buf:
1102         stats->drops++;
1103         dev_kfree_skb(head_skb);
1104 xdp_xmit:
1105         return NULL;
1106 }
1107
1108 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1109                         void *buf, unsigned int len, void **ctx,
1110                         unsigned int *xdp_xmit,
1111                         struct virtnet_rq_stats *stats)
1112 {
1113         struct net_device *dev = vi->dev;
1114         struct sk_buff *skb;
1115         struct virtio_net_hdr_mrg_rxbuf *hdr;
1116
1117         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1118                 pr_debug("%s: short packet %i\n", dev->name, len);
1119                 dev->stats.rx_length_errors++;
1120                 if (vi->mergeable_rx_bufs) {
1121                         put_page(virt_to_head_page(buf));
1122                 } else if (vi->big_packets) {
1123                         give_pages(rq, buf);
1124                 } else {
1125                         put_page(virt_to_head_page(buf));
1126                 }
1127                 return;
1128         }
1129
1130         if (vi->mergeable_rx_bufs)
1131                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1132                                         stats);
1133         else if (vi->big_packets)
1134                 skb = receive_big(dev, vi, rq, buf, len, stats);
1135         else
1136                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1137
1138         if (unlikely(!skb))
1139                 return;
1140
1141         hdr = skb_vnet_hdr(skb);
1142
1143         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1144                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1145
1146         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1147                                   virtio_is_little_endian(vi->vdev))) {
1148                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1149                                      dev->name, hdr->hdr.gso_type,
1150                                      hdr->hdr.gso_size);
1151                 goto frame_err;
1152         }
1153
1154         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1155         skb->protocol = eth_type_trans(skb, dev);
1156         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1157                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1158
1159         napi_gro_receive(&rq->napi, skb);
1160         return;
1161
1162 frame_err:
1163         dev->stats.rx_frame_errors++;
1164         dev_kfree_skb(skb);
1165 }
1166
1167 /* Unlike mergeable buffers, all buffers are allocated to the
1168  * same size, except for the headroom. For this reason we do
1169  * not need to use  mergeable_len_to_ctx here - it is enough
1170  * to store the headroom as the context ignoring the truesize.
1171  */
1172 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1173                              gfp_t gfp)
1174 {
1175         struct page_frag *alloc_frag = &rq->alloc_frag;
1176         char *buf;
1177         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1178         void *ctx = (void *)(unsigned long)xdp_headroom;
1179         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1180         int err;
1181
1182         len = SKB_DATA_ALIGN(len) +
1183               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1184         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1185                 return -ENOMEM;
1186
1187         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1188         get_page(alloc_frag->page);
1189         alloc_frag->offset += len;
1190         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1191                     vi->hdr_len + GOOD_PACKET_LEN);
1192         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1193         if (err < 0)
1194                 put_page(virt_to_head_page(buf));
1195         return err;
1196 }
1197
1198 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1199                            gfp_t gfp)
1200 {
1201         struct page *first, *list = NULL;
1202         char *p;
1203         int i, err, offset;
1204
1205         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1206
1207         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1208         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1209                 first = get_a_page(rq, gfp);
1210                 if (!first) {
1211                         if (list)
1212                                 give_pages(rq, list);
1213                         return -ENOMEM;
1214                 }
1215                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1216
1217                 /* chain new page in list head to match sg */
1218                 first->private = (unsigned long)list;
1219                 list = first;
1220         }
1221
1222         first = get_a_page(rq, gfp);
1223         if (!first) {
1224                 give_pages(rq, list);
1225                 return -ENOMEM;
1226         }
1227         p = page_address(first);
1228
1229         /* rq->sg[0], rq->sg[1] share the same page */
1230         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1231         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1232
1233         /* rq->sg[1] for data packet, from offset */
1234         offset = sizeof(struct padded_vnet_hdr);
1235         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1236
1237         /* chain first in list head */
1238         first->private = (unsigned long)list;
1239         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1240                                   first, gfp);
1241         if (err < 0)
1242                 give_pages(rq, first);
1243
1244         return err;
1245 }
1246
1247 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1248                                           struct ewma_pkt_len *avg_pkt_len,
1249                                           unsigned int room)
1250 {
1251         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1252         unsigned int len;
1253
1254         if (room)
1255                 return PAGE_SIZE - room;
1256
1257         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1258                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1259
1260         return ALIGN(len, L1_CACHE_BYTES);
1261 }
1262
1263 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1264                                  struct receive_queue *rq, gfp_t gfp)
1265 {
1266         struct page_frag *alloc_frag = &rq->alloc_frag;
1267         unsigned int headroom = virtnet_get_headroom(vi);
1268         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1269         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1270         char *buf;
1271         void *ctx;
1272         int err;
1273         unsigned int len, hole;
1274
1275         /* Extra tailroom is needed to satisfy XDP's assumption. This
1276          * means rx frags coalescing won't work, but consider we've
1277          * disabled GSO for XDP, it won't be a big issue.
1278          */
1279         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1280         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1281                 return -ENOMEM;
1282
1283         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1284         buf += headroom; /* advance address leaving hole at front of pkt */
1285         get_page(alloc_frag->page);
1286         alloc_frag->offset += len + room;
1287         hole = alloc_frag->size - alloc_frag->offset;
1288         if (hole < len + room) {
1289                 /* To avoid internal fragmentation, if there is very likely not
1290                  * enough space for another buffer, add the remaining space to
1291                  * the current buffer.
1292                  */
1293                 len += hole;
1294                 alloc_frag->offset += hole;
1295         }
1296
1297         sg_init_one(rq->sg, buf, len);
1298         ctx = mergeable_len_to_ctx(len, headroom);
1299         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1300         if (err < 0)
1301                 put_page(virt_to_head_page(buf));
1302
1303         return err;
1304 }
1305
1306 /*
1307  * Returns false if we couldn't fill entirely (OOM).
1308  *
1309  * Normally run in the receive path, but can also be run from ndo_open
1310  * before we're receiving packets, or from refill_work which is
1311  * careful to disable receiving (using napi_disable).
1312  */
1313 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1314                           gfp_t gfp)
1315 {
1316         int err;
1317         bool oom;
1318
1319         do {
1320                 if (vi->mergeable_rx_bufs)
1321                         err = add_recvbuf_mergeable(vi, rq, gfp);
1322                 else if (vi->big_packets)
1323                         err = add_recvbuf_big(vi, rq, gfp);
1324                 else
1325                         err = add_recvbuf_small(vi, rq, gfp);
1326
1327                 oom = err == -ENOMEM;
1328                 if (err)
1329                         break;
1330         } while (rq->vq->num_free);
1331         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1332                 unsigned long flags;
1333
1334                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1335                 rq->stats.kicks++;
1336                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1337         }
1338
1339         return !oom;
1340 }
1341
1342 static void skb_recv_done(struct virtqueue *rvq)
1343 {
1344         struct virtnet_info *vi = rvq->vdev->priv;
1345         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1346
1347         virtqueue_napi_schedule(&rq->napi, rvq);
1348 }
1349
1350 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1351 {
1352         napi_enable(napi);
1353
1354         /* If all buffers were filled by other side before we napi_enabled, we
1355          * won't get another interrupt, so process any outstanding packets now.
1356          * Call local_bh_enable after to trigger softIRQ processing.
1357          */
1358         local_bh_disable();
1359         virtqueue_napi_schedule(napi, vq);
1360         local_bh_enable();
1361 }
1362
1363 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1364                                    struct virtqueue *vq,
1365                                    struct napi_struct *napi)
1366 {
1367         if (!napi->weight)
1368                 return;
1369
1370         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1371          * enable the feature if this is likely affine with the transmit path.
1372          */
1373         if (!vi->affinity_hint_set) {
1374                 napi->weight = 0;
1375                 return;
1376         }
1377
1378         return virtnet_napi_enable(vq, napi);
1379 }
1380
1381 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1382 {
1383         if (napi->weight)
1384                 napi_disable(napi);
1385 }
1386
1387 static void refill_work(struct work_struct *work)
1388 {
1389         struct virtnet_info *vi =
1390                 container_of(work, struct virtnet_info, refill.work);
1391         bool still_empty;
1392         int i;
1393
1394         for (i = 0; i < vi->curr_queue_pairs; i++) {
1395                 struct receive_queue *rq = &vi->rq[i];
1396
1397                 napi_disable(&rq->napi);
1398                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1399                 virtnet_napi_enable(rq->vq, &rq->napi);
1400
1401                 /* In theory, this can happen: if we don't get any buffers in
1402                  * we will *never* try to fill again.
1403                  */
1404                 if (still_empty)
1405                         schedule_delayed_work(&vi->refill, HZ/2);
1406         }
1407 }
1408
1409 static int virtnet_receive(struct receive_queue *rq, int budget,
1410                            unsigned int *xdp_xmit)
1411 {
1412         struct virtnet_info *vi = rq->vq->vdev->priv;
1413         struct virtnet_rq_stats stats = {};
1414         unsigned int len;
1415         void *buf;
1416         int i;
1417
1418         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1419                 void *ctx;
1420
1421                 while (stats.packets < budget &&
1422                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1423                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1424                         stats.packets++;
1425                 }
1426         } else {
1427                 while (stats.packets < budget &&
1428                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1429                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1430                         stats.packets++;
1431                 }
1432         }
1433
1434         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1435                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1436                         schedule_delayed_work(&vi->refill, 0);
1437         }
1438
1439         u64_stats_update_begin(&rq->stats.syncp);
1440         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1441                 size_t offset = virtnet_rq_stats_desc[i].offset;
1442                 u64 *item;
1443
1444                 item = (u64 *)((u8 *)&rq->stats + offset);
1445                 *item += *(u64 *)((u8 *)&stats + offset);
1446         }
1447         u64_stats_update_end(&rq->stats.syncp);
1448
1449         return stats.packets;
1450 }
1451
1452 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1453 {
1454         unsigned int len;
1455         unsigned int packets = 0;
1456         unsigned int bytes = 0;
1457         void *ptr;
1458
1459         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1460                 if (likely(!is_xdp_frame(ptr))) {
1461                         struct sk_buff *skb = ptr;
1462
1463                         pr_debug("Sent skb %p\n", skb);
1464
1465                         bytes += skb->len;
1466                         napi_consume_skb(skb, in_napi);
1467                 } else {
1468                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1469
1470                         bytes += frame->len;
1471                         xdp_return_frame(frame);
1472                 }
1473                 packets++;
1474         }
1475
1476         /* Avoid overhead when no packets have been processed
1477          * happens when called speculatively from start_xmit.
1478          */
1479         if (!packets)
1480                 return;
1481
1482         u64_stats_update_begin(&sq->stats.syncp);
1483         sq->stats.bytes += bytes;
1484         sq->stats.packets += packets;
1485         u64_stats_update_end(&sq->stats.syncp);
1486 }
1487
1488 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1489 {
1490         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1491                 return false;
1492         else if (q < vi->curr_queue_pairs)
1493                 return true;
1494         else
1495                 return false;
1496 }
1497
1498 static void virtnet_poll_cleantx(struct receive_queue *rq)
1499 {
1500         struct virtnet_info *vi = rq->vq->vdev->priv;
1501         unsigned int index = vq2rxq(rq->vq);
1502         struct send_queue *sq = &vi->sq[index];
1503         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1504
1505         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1506                 return;
1507
1508         if (__netif_tx_trylock(txq)) {
1509                 do {
1510                         virtqueue_disable_cb(sq->vq);
1511                         free_old_xmit_skbs(sq, true);
1512                 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1513
1514                 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1515                         netif_tx_wake_queue(txq);
1516
1517                 __netif_tx_unlock(txq);
1518         }
1519 }
1520
1521 static int virtnet_poll(struct napi_struct *napi, int budget)
1522 {
1523         struct receive_queue *rq =
1524                 container_of(napi, struct receive_queue, napi);
1525         struct virtnet_info *vi = rq->vq->vdev->priv;
1526         struct send_queue *sq;
1527         unsigned int received;
1528         unsigned int xdp_xmit = 0;
1529
1530         virtnet_poll_cleantx(rq);
1531
1532         received = virtnet_receive(rq, budget, &xdp_xmit);
1533
1534         /* Out of packets? */
1535         if (received < budget)
1536                 virtqueue_napi_complete(napi, rq->vq, received);
1537
1538         if (xdp_xmit & VIRTIO_XDP_REDIR)
1539                 xdp_do_flush();
1540
1541         if (xdp_xmit & VIRTIO_XDP_TX) {
1542                 sq = virtnet_xdp_get_sq(vi);
1543                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1544                         u64_stats_update_begin(&sq->stats.syncp);
1545                         sq->stats.kicks++;
1546                         u64_stats_update_end(&sq->stats.syncp);
1547                 }
1548                 virtnet_xdp_put_sq(vi, sq);
1549         }
1550
1551         return received;
1552 }
1553
1554 static int virtnet_open(struct net_device *dev)
1555 {
1556         struct virtnet_info *vi = netdev_priv(dev);
1557         int i, err;
1558
1559         for (i = 0; i < vi->max_queue_pairs; i++) {
1560                 if (i < vi->curr_queue_pairs)
1561                         /* Make sure we have some buffers: if oom use wq. */
1562                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1563                                 schedule_delayed_work(&vi->refill, 0);
1564
1565                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1566                 if (err < 0)
1567                         return err;
1568
1569                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1570                                                  MEM_TYPE_PAGE_SHARED, NULL);
1571                 if (err < 0) {
1572                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1573                         return err;
1574                 }
1575
1576                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1577                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1578         }
1579
1580         return 0;
1581 }
1582
1583 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1584 {
1585         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1586         struct virtnet_info *vi = sq->vq->vdev->priv;
1587         unsigned int index = vq2txq(sq->vq);
1588         struct netdev_queue *txq;
1589         int opaque;
1590         bool done;
1591
1592         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1593                 /* We don't need to enable cb for XDP */
1594                 napi_complete_done(napi, 0);
1595                 return 0;
1596         }
1597
1598         txq = netdev_get_tx_queue(vi->dev, index);
1599         __netif_tx_lock(txq, raw_smp_processor_id());
1600         virtqueue_disable_cb(sq->vq);
1601         free_old_xmit_skbs(sq, true);
1602
1603         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1604                 netif_tx_wake_queue(txq);
1605
1606         opaque = virtqueue_enable_cb_prepare(sq->vq);
1607
1608         done = napi_complete_done(napi, 0);
1609
1610         if (!done)
1611                 virtqueue_disable_cb(sq->vq);
1612
1613         __netif_tx_unlock(txq);
1614
1615         if (done) {
1616                 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1617                         if (napi_schedule_prep(napi)) {
1618                                 __netif_tx_lock(txq, raw_smp_processor_id());
1619                                 virtqueue_disable_cb(sq->vq);
1620                                 __netif_tx_unlock(txq);
1621                                 __napi_schedule(napi);
1622                         }
1623                 }
1624         }
1625
1626         return 0;
1627 }
1628
1629 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1630 {
1631         struct virtio_net_hdr_mrg_rxbuf *hdr;
1632         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1633         struct virtnet_info *vi = sq->vq->vdev->priv;
1634         int num_sg;
1635         unsigned hdr_len = vi->hdr_len;
1636         bool can_push;
1637
1638         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1639
1640         can_push = vi->any_header_sg &&
1641                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1642                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1643         /* Even if we can, don't push here yet as this would skew
1644          * csum_start offset below. */
1645         if (can_push)
1646                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1647         else
1648                 hdr = skb_vnet_hdr(skb);
1649
1650         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1651                                     virtio_is_little_endian(vi->vdev), false,
1652                                     0))
1653                 return -EPROTO;
1654
1655         if (vi->mergeable_rx_bufs)
1656                 hdr->num_buffers = 0;
1657
1658         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1659         if (can_push) {
1660                 __skb_push(skb, hdr_len);
1661                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1662                 if (unlikely(num_sg < 0))
1663                         return num_sg;
1664                 /* Pull header back to avoid skew in tx bytes calculations. */
1665                 __skb_pull(skb, hdr_len);
1666         } else {
1667                 sg_set_buf(sq->sg, hdr, hdr_len);
1668                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1669                 if (unlikely(num_sg < 0))
1670                         return num_sg;
1671                 num_sg++;
1672         }
1673         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1674 }
1675
1676 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1677 {
1678         struct virtnet_info *vi = netdev_priv(dev);
1679         int qnum = skb_get_queue_mapping(skb);
1680         struct send_queue *sq = &vi->sq[qnum];
1681         int err;
1682         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1683         bool kick = !netdev_xmit_more();
1684         bool use_napi = sq->napi.weight;
1685
1686         /* Free up any pending old buffers before queueing new ones. */
1687         do {
1688                 if (use_napi)
1689                         virtqueue_disable_cb(sq->vq);
1690
1691                 free_old_xmit_skbs(sq, false);
1692
1693         } while (use_napi && kick &&
1694                unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1695
1696         /* timestamp packet in software */
1697         skb_tx_timestamp(skb);
1698
1699         /* Try to transmit */
1700         err = xmit_skb(sq, skb);
1701
1702         /* This should not happen! */
1703         if (unlikely(err)) {
1704                 dev->stats.tx_fifo_errors++;
1705                 if (net_ratelimit())
1706                         dev_warn(&dev->dev,
1707                                  "Unexpected TXQ (%d) queue failure: %d\n",
1708                                  qnum, err);
1709                 dev->stats.tx_dropped++;
1710                 dev_kfree_skb_any(skb);
1711                 return NETDEV_TX_OK;
1712         }
1713
1714         /* Don't wait up for transmitted skbs to be freed. */
1715         if (!use_napi) {
1716                 skb_orphan(skb);
1717                 nf_reset_ct(skb);
1718         }
1719
1720         /* If running out of space, stop queue to avoid getting packets that we
1721          * are then unable to transmit.
1722          * An alternative would be to force queuing layer to requeue the skb by
1723          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1724          * returned in a normal path of operation: it means that driver is not
1725          * maintaining the TX queue stop/start state properly, and causes
1726          * the stack to do a non-trivial amount of useless work.
1727          * Since most packets only take 1 or 2 ring slots, stopping the queue
1728          * early means 16 slots are typically wasted.
1729          */
1730         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1731                 netif_stop_subqueue(dev, qnum);
1732                 if (!use_napi &&
1733                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1734                         /* More just got used, free them then recheck. */
1735                         free_old_xmit_skbs(sq, false);
1736                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1737                                 netif_start_subqueue(dev, qnum);
1738                                 virtqueue_disable_cb(sq->vq);
1739                         }
1740                 }
1741         }
1742
1743         if (kick || netif_xmit_stopped(txq)) {
1744                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1745                         u64_stats_update_begin(&sq->stats.syncp);
1746                         sq->stats.kicks++;
1747                         u64_stats_update_end(&sq->stats.syncp);
1748                 }
1749         }
1750
1751         return NETDEV_TX_OK;
1752 }
1753
1754 /*
1755  * Send command via the control virtqueue and check status.  Commands
1756  * supported by the hypervisor, as indicated by feature bits, should
1757  * never fail unless improperly formatted.
1758  */
1759 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1760                                  struct scatterlist *out)
1761 {
1762         struct scatterlist *sgs[4], hdr, stat;
1763         unsigned out_num = 0, tmp;
1764         int ret;
1765
1766         /* Caller should know better */
1767         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1768
1769         vi->ctrl->status = ~0;
1770         vi->ctrl->hdr.class = class;
1771         vi->ctrl->hdr.cmd = cmd;
1772         /* Add header */
1773         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1774         sgs[out_num++] = &hdr;
1775
1776         if (out)
1777                 sgs[out_num++] = out;
1778
1779         /* Add return status. */
1780         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1781         sgs[out_num] = &stat;
1782
1783         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1784         ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1785         if (ret < 0) {
1786                 dev_warn(&vi->vdev->dev,
1787                          "Failed to add sgs for command vq: %d\n.", ret);
1788                 return false;
1789         }
1790
1791         if (unlikely(!virtqueue_kick(vi->cvq)))
1792                 return vi->ctrl->status == VIRTIO_NET_OK;
1793
1794         /* Spin for a response, the kick causes an ioport write, trapping
1795          * into the hypervisor, so the request should be handled immediately.
1796          */
1797         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1798                !virtqueue_is_broken(vi->cvq))
1799                 cpu_relax();
1800
1801         return vi->ctrl->status == VIRTIO_NET_OK;
1802 }
1803
1804 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1805 {
1806         struct virtnet_info *vi = netdev_priv(dev);
1807         struct virtio_device *vdev = vi->vdev;
1808         int ret;
1809         struct sockaddr *addr;
1810         struct scatterlist sg;
1811
1812         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1813                 return -EOPNOTSUPP;
1814
1815         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1816         if (!addr)
1817                 return -ENOMEM;
1818
1819         ret = eth_prepare_mac_addr_change(dev, addr);
1820         if (ret)
1821                 goto out;
1822
1823         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1824                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1825                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1826                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1827                         dev_warn(&vdev->dev,
1828                                  "Failed to set mac address by vq command.\n");
1829                         ret = -EINVAL;
1830                         goto out;
1831                 }
1832         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1833                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1834                 unsigned int i;
1835
1836                 /* Naturally, this has an atomicity problem. */
1837                 for (i = 0; i < dev->addr_len; i++)
1838                         virtio_cwrite8(vdev,
1839                                        offsetof(struct virtio_net_config, mac) +
1840                                        i, addr->sa_data[i]);
1841         }
1842
1843         eth_commit_mac_addr_change(dev, p);
1844         ret = 0;
1845
1846 out:
1847         kfree(addr);
1848         return ret;
1849 }
1850
1851 static void virtnet_stats(struct net_device *dev,
1852                           struct rtnl_link_stats64 *tot)
1853 {
1854         struct virtnet_info *vi = netdev_priv(dev);
1855         unsigned int start;
1856         int i;
1857
1858         for (i = 0; i < vi->max_queue_pairs; i++) {
1859                 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1860                 struct receive_queue *rq = &vi->rq[i];
1861                 struct send_queue *sq = &vi->sq[i];
1862
1863                 do {
1864                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1865                         tpackets = sq->stats.packets;
1866                         tbytes   = sq->stats.bytes;
1867                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1868
1869                 do {
1870                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1871                         rpackets = rq->stats.packets;
1872                         rbytes   = rq->stats.bytes;
1873                         rdrops   = rq->stats.drops;
1874                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1875
1876                 tot->rx_packets += rpackets;
1877                 tot->tx_packets += tpackets;
1878                 tot->rx_bytes   += rbytes;
1879                 tot->tx_bytes   += tbytes;
1880                 tot->rx_dropped += rdrops;
1881         }
1882
1883         tot->tx_dropped = dev->stats.tx_dropped;
1884         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1885         tot->rx_length_errors = dev->stats.rx_length_errors;
1886         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1887 }
1888
1889 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1890 {
1891         rtnl_lock();
1892         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1893                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1894                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1895         rtnl_unlock();
1896 }
1897
1898 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1899 {
1900         struct scatterlist sg;
1901         struct net_device *dev = vi->dev;
1902
1903         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1904                 return 0;
1905
1906         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1907         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1908
1909         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1910                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1911                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1912                          queue_pairs);
1913                 return -EINVAL;
1914         } else {
1915                 vi->curr_queue_pairs = queue_pairs;
1916                 /* virtnet_open() will refill when device is going to up. */
1917                 if (dev->flags & IFF_UP)
1918                         schedule_delayed_work(&vi->refill, 0);
1919         }
1920
1921         return 0;
1922 }
1923
1924 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1925 {
1926         int err;
1927
1928         rtnl_lock();
1929         err = _virtnet_set_queues(vi, queue_pairs);
1930         rtnl_unlock();
1931         return err;
1932 }
1933
1934 static int virtnet_close(struct net_device *dev)
1935 {
1936         struct virtnet_info *vi = netdev_priv(dev);
1937         int i;
1938
1939         /* Make sure refill_work doesn't re-enable napi! */
1940         cancel_delayed_work_sync(&vi->refill);
1941
1942         for (i = 0; i < vi->max_queue_pairs; i++) {
1943                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1944                 napi_disable(&vi->rq[i].napi);
1945                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1946         }
1947
1948         return 0;
1949 }
1950
1951 static void virtnet_set_rx_mode(struct net_device *dev)
1952 {
1953         struct virtnet_info *vi = netdev_priv(dev);
1954         struct scatterlist sg[2];
1955         struct virtio_net_ctrl_mac *mac_data;
1956         struct netdev_hw_addr *ha;
1957         int uc_count;
1958         int mc_count;
1959         void *buf;
1960         int i;
1961
1962         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1963         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1964                 return;
1965
1966         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1967         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1968
1969         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1970
1971         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1972                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1973                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1974                          vi->ctrl->promisc ? "en" : "dis");
1975
1976         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1977
1978         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1979                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1980                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1981                          vi->ctrl->allmulti ? "en" : "dis");
1982
1983         uc_count = netdev_uc_count(dev);
1984         mc_count = netdev_mc_count(dev);
1985         /* MAC filter - use one buffer for both lists */
1986         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1987                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1988         mac_data = buf;
1989         if (!buf)
1990                 return;
1991
1992         sg_init_table(sg, 2);
1993
1994         /* Store the unicast list and count in the front of the buffer */
1995         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1996         i = 0;
1997         netdev_for_each_uc_addr(ha, dev)
1998                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1999
2000         sg_set_buf(&sg[0], mac_data,
2001                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2002
2003         /* multicast list and count fill the end */
2004         mac_data = (void *)&mac_data->macs[uc_count][0];
2005
2006         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2007         i = 0;
2008         netdev_for_each_mc_addr(ha, dev)
2009                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2010
2011         sg_set_buf(&sg[1], mac_data,
2012                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2013
2014         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2015                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2016                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2017
2018         kfree(buf);
2019 }
2020
2021 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2022                                    __be16 proto, u16 vid)
2023 {
2024         struct virtnet_info *vi = netdev_priv(dev);
2025         struct scatterlist sg;
2026
2027         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2028         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2029
2030         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2031                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2032                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2033         return 0;
2034 }
2035
2036 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2037                                     __be16 proto, u16 vid)
2038 {
2039         struct virtnet_info *vi = netdev_priv(dev);
2040         struct scatterlist sg;
2041
2042         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2043         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2044
2045         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2046                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2047                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2048         return 0;
2049 }
2050
2051 static void virtnet_clean_affinity(struct virtnet_info *vi)
2052 {
2053         int i;
2054
2055         if (vi->affinity_hint_set) {
2056                 for (i = 0; i < vi->max_queue_pairs; i++) {
2057                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
2058                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
2059                 }
2060
2061                 vi->affinity_hint_set = false;
2062         }
2063 }
2064
2065 static void virtnet_set_affinity(struct virtnet_info *vi)
2066 {
2067         cpumask_var_t mask;
2068         int stragglers;
2069         int group_size;
2070         int i, j, cpu;
2071         int num_cpu;
2072         int stride;
2073
2074         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2075                 virtnet_clean_affinity(vi);
2076                 return;
2077         }
2078
2079         num_cpu = num_online_cpus();
2080         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2081         stragglers = num_cpu >= vi->curr_queue_pairs ?
2082                         num_cpu % vi->curr_queue_pairs :
2083                         0;
2084         cpu = cpumask_next(-1, cpu_online_mask);
2085
2086         for (i = 0; i < vi->curr_queue_pairs; i++) {
2087                 group_size = stride + (i < stragglers ? 1 : 0);
2088
2089                 for (j = 0; j < group_size; j++) {
2090                         cpumask_set_cpu(cpu, mask);
2091                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2092                                                 nr_cpu_ids, false);
2093                 }
2094                 virtqueue_set_affinity(vi->rq[i].vq, mask);
2095                 virtqueue_set_affinity(vi->sq[i].vq, mask);
2096                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2097                 cpumask_clear(mask);
2098         }
2099
2100         vi->affinity_hint_set = true;
2101         free_cpumask_var(mask);
2102 }
2103
2104 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2105 {
2106         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2107                                                    node);
2108         virtnet_set_affinity(vi);
2109         return 0;
2110 }
2111
2112 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2113 {
2114         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2115                                                    node_dead);
2116         virtnet_set_affinity(vi);
2117         return 0;
2118 }
2119
2120 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2121 {
2122         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2123                                                    node);
2124
2125         virtnet_clean_affinity(vi);
2126         return 0;
2127 }
2128
2129 static enum cpuhp_state virtionet_online;
2130
2131 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2132 {
2133         int ret;
2134
2135         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2136         if (ret)
2137                 return ret;
2138         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2139                                                &vi->node_dead);
2140         if (!ret)
2141                 return ret;
2142         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2143         return ret;
2144 }
2145
2146 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2147 {
2148         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2149         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2150                                             &vi->node_dead);
2151 }
2152
2153 static void virtnet_get_ringparam(struct net_device *dev,
2154                                 struct ethtool_ringparam *ring)
2155 {
2156         struct virtnet_info *vi = netdev_priv(dev);
2157
2158         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2159         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2160         ring->rx_pending = ring->rx_max_pending;
2161         ring->tx_pending = ring->tx_max_pending;
2162 }
2163
2164
2165 static void virtnet_get_drvinfo(struct net_device *dev,
2166                                 struct ethtool_drvinfo *info)
2167 {
2168         struct virtnet_info *vi = netdev_priv(dev);
2169         struct virtio_device *vdev = vi->vdev;
2170
2171         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2172         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2173         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2174
2175 }
2176
2177 /* TODO: Eliminate OOO packets during switching */
2178 static int virtnet_set_channels(struct net_device *dev,
2179                                 struct ethtool_channels *channels)
2180 {
2181         struct virtnet_info *vi = netdev_priv(dev);
2182         u16 queue_pairs = channels->combined_count;
2183         int err;
2184
2185         /* We don't support separate rx/tx channels.
2186          * We don't allow setting 'other' channels.
2187          */
2188         if (channels->rx_count || channels->tx_count || channels->other_count)
2189                 return -EINVAL;
2190
2191         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2192                 return -EINVAL;
2193
2194         /* For now we don't support modifying channels while XDP is loaded
2195          * also when XDP is loaded all RX queues have XDP programs so we only
2196          * need to check a single RX queue.
2197          */
2198         if (vi->rq[0].xdp_prog)
2199                 return -EINVAL;
2200
2201         cpus_read_lock();
2202         err = _virtnet_set_queues(vi, queue_pairs);
2203         if (err) {
2204                 cpus_read_unlock();
2205                 goto err;
2206         }
2207         virtnet_set_affinity(vi);
2208         cpus_read_unlock();
2209
2210         netif_set_real_num_tx_queues(dev, queue_pairs);
2211         netif_set_real_num_rx_queues(dev, queue_pairs);
2212  err:
2213         return err;
2214 }
2215
2216 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2217 {
2218         struct virtnet_info *vi = netdev_priv(dev);
2219         unsigned int i, j;
2220         u8 *p = data;
2221
2222         switch (stringset) {
2223         case ETH_SS_STATS:
2224                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2225                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2226                                 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2227                                                 virtnet_rq_stats_desc[j].desc);
2228                 }
2229
2230                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2231                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2232                                 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2233                                                 virtnet_sq_stats_desc[j].desc);
2234                 }
2235                 break;
2236         }
2237 }
2238
2239 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2240 {
2241         struct virtnet_info *vi = netdev_priv(dev);
2242
2243         switch (sset) {
2244         case ETH_SS_STATS:
2245                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2246                                                VIRTNET_SQ_STATS_LEN);
2247         default:
2248                 return -EOPNOTSUPP;
2249         }
2250 }
2251
2252 static void virtnet_get_ethtool_stats(struct net_device *dev,
2253                                       struct ethtool_stats *stats, u64 *data)
2254 {
2255         struct virtnet_info *vi = netdev_priv(dev);
2256         unsigned int idx = 0, start, i, j;
2257         const u8 *stats_base;
2258         size_t offset;
2259
2260         for (i = 0; i < vi->curr_queue_pairs; i++) {
2261                 struct receive_queue *rq = &vi->rq[i];
2262
2263                 stats_base = (u8 *)&rq->stats;
2264                 do {
2265                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2266                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2267                                 offset = virtnet_rq_stats_desc[j].offset;
2268                                 data[idx + j] = *(u64 *)(stats_base + offset);
2269                         }
2270                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2271                 idx += VIRTNET_RQ_STATS_LEN;
2272         }
2273
2274         for (i = 0; i < vi->curr_queue_pairs; i++) {
2275                 struct send_queue *sq = &vi->sq[i];
2276
2277                 stats_base = (u8 *)&sq->stats;
2278                 do {
2279                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2280                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2281                                 offset = virtnet_sq_stats_desc[j].offset;
2282                                 data[idx + j] = *(u64 *)(stats_base + offset);
2283                         }
2284                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2285                 idx += VIRTNET_SQ_STATS_LEN;
2286         }
2287 }
2288
2289 static void virtnet_get_channels(struct net_device *dev,
2290                                  struct ethtool_channels *channels)
2291 {
2292         struct virtnet_info *vi = netdev_priv(dev);
2293
2294         channels->combined_count = vi->curr_queue_pairs;
2295         channels->max_combined = vi->max_queue_pairs;
2296         channels->max_other = 0;
2297         channels->rx_count = 0;
2298         channels->tx_count = 0;
2299         channels->other_count = 0;
2300 }
2301
2302 static int virtnet_set_link_ksettings(struct net_device *dev,
2303                                       const struct ethtool_link_ksettings *cmd)
2304 {
2305         struct virtnet_info *vi = netdev_priv(dev);
2306
2307         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2308                                                   &vi->speed, &vi->duplex);
2309 }
2310
2311 static int virtnet_get_link_ksettings(struct net_device *dev,
2312                                       struct ethtool_link_ksettings *cmd)
2313 {
2314         struct virtnet_info *vi = netdev_priv(dev);
2315
2316         cmd->base.speed = vi->speed;
2317         cmd->base.duplex = vi->duplex;
2318         cmd->base.port = PORT_OTHER;
2319
2320         return 0;
2321 }
2322
2323 static int virtnet_set_coalesce(struct net_device *dev,
2324                                 struct ethtool_coalesce *ec,
2325                                 struct kernel_ethtool_coalesce *kernel_coal,
2326                                 struct netlink_ext_ack *extack)
2327 {
2328         struct virtnet_info *vi = netdev_priv(dev);
2329         int i, napi_weight;
2330
2331         if (ec->tx_max_coalesced_frames > 1 ||
2332             ec->rx_max_coalesced_frames != 1)
2333                 return -EINVAL;
2334
2335         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2336         if (napi_weight ^ vi->sq[0].napi.weight) {
2337                 if (dev->flags & IFF_UP)
2338                         return -EBUSY;
2339                 for (i = 0; i < vi->max_queue_pairs; i++)
2340                         vi->sq[i].napi.weight = napi_weight;
2341         }
2342
2343         return 0;
2344 }
2345
2346 static int virtnet_get_coalesce(struct net_device *dev,
2347                                 struct ethtool_coalesce *ec,
2348                                 struct kernel_ethtool_coalesce *kernel_coal,
2349                                 struct netlink_ext_ack *extack)
2350 {
2351         struct ethtool_coalesce ec_default = {
2352                 .cmd = ETHTOOL_GCOALESCE,
2353                 .rx_max_coalesced_frames = 1,
2354         };
2355         struct virtnet_info *vi = netdev_priv(dev);
2356
2357         memcpy(ec, &ec_default, sizeof(ec_default));
2358
2359         if (vi->sq[0].napi.weight)
2360                 ec->tx_max_coalesced_frames = 1;
2361
2362         return 0;
2363 }
2364
2365 static void virtnet_init_settings(struct net_device *dev)
2366 {
2367         struct virtnet_info *vi = netdev_priv(dev);
2368
2369         vi->speed = SPEED_UNKNOWN;
2370         vi->duplex = DUPLEX_UNKNOWN;
2371 }
2372
2373 static void virtnet_update_settings(struct virtnet_info *vi)
2374 {
2375         u32 speed;
2376         u8 duplex;
2377
2378         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2379                 return;
2380
2381         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2382
2383         if (ethtool_validate_speed(speed))
2384                 vi->speed = speed;
2385
2386         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2387
2388         if (ethtool_validate_duplex(duplex))
2389                 vi->duplex = duplex;
2390 }
2391
2392 static const struct ethtool_ops virtnet_ethtool_ops = {
2393         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2394         .get_drvinfo = virtnet_get_drvinfo,
2395         .get_link = ethtool_op_get_link,
2396         .get_ringparam = virtnet_get_ringparam,
2397         .get_strings = virtnet_get_strings,
2398         .get_sset_count = virtnet_get_sset_count,
2399         .get_ethtool_stats = virtnet_get_ethtool_stats,
2400         .set_channels = virtnet_set_channels,
2401         .get_channels = virtnet_get_channels,
2402         .get_ts_info = ethtool_op_get_ts_info,
2403         .get_link_ksettings = virtnet_get_link_ksettings,
2404         .set_link_ksettings = virtnet_set_link_ksettings,
2405         .set_coalesce = virtnet_set_coalesce,
2406         .get_coalesce = virtnet_get_coalesce,
2407 };
2408
2409 static void virtnet_freeze_down(struct virtio_device *vdev)
2410 {
2411         struct virtnet_info *vi = vdev->priv;
2412         int i;
2413
2414         /* Make sure no work handler is accessing the device */
2415         flush_work(&vi->config_work);
2416
2417         netif_tx_lock_bh(vi->dev);
2418         netif_device_detach(vi->dev);
2419         netif_tx_unlock_bh(vi->dev);
2420         cancel_delayed_work_sync(&vi->refill);
2421
2422         if (netif_running(vi->dev)) {
2423                 for (i = 0; i < vi->max_queue_pairs; i++) {
2424                         napi_disable(&vi->rq[i].napi);
2425                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2426                 }
2427         }
2428 }
2429
2430 static int init_vqs(struct virtnet_info *vi);
2431
2432 static int virtnet_restore_up(struct virtio_device *vdev)
2433 {
2434         struct virtnet_info *vi = vdev->priv;
2435         int err, i;
2436
2437         err = init_vqs(vi);
2438         if (err)
2439                 return err;
2440
2441         virtio_device_ready(vdev);
2442
2443         if (netif_running(vi->dev)) {
2444                 for (i = 0; i < vi->curr_queue_pairs; i++)
2445                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2446                                 schedule_delayed_work(&vi->refill, 0);
2447
2448                 for (i = 0; i < vi->max_queue_pairs; i++) {
2449                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2450                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2451                                                &vi->sq[i].napi);
2452                 }
2453         }
2454
2455         netif_tx_lock_bh(vi->dev);
2456         netif_device_attach(vi->dev);
2457         netif_tx_unlock_bh(vi->dev);
2458         return err;
2459 }
2460
2461 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2462 {
2463         struct scatterlist sg;
2464         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2465
2466         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2467
2468         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2469                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2470                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2471                 return -EINVAL;
2472         }
2473
2474         return 0;
2475 }
2476
2477 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2478 {
2479         u64 offloads = 0;
2480
2481         if (!vi->guest_offloads)
2482                 return 0;
2483
2484         return virtnet_set_guest_offloads(vi, offloads);
2485 }
2486
2487 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2488 {
2489         u64 offloads = vi->guest_offloads;
2490
2491         if (!vi->guest_offloads)
2492                 return 0;
2493
2494         return virtnet_set_guest_offloads(vi, offloads);
2495 }
2496
2497 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2498                            struct netlink_ext_ack *extack)
2499 {
2500         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2501         struct virtnet_info *vi = netdev_priv(dev);
2502         struct bpf_prog *old_prog;
2503         u16 xdp_qp = 0, curr_qp;
2504         int i, err;
2505
2506         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2507             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2508                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2509                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2510                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2511                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2512                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
2513                 return -EOPNOTSUPP;
2514         }
2515
2516         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2517                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2518                 return -EINVAL;
2519         }
2520
2521         if (dev->mtu > max_sz) {
2522                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2523                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2524                 return -EINVAL;
2525         }
2526
2527         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2528         if (prog)
2529                 xdp_qp = nr_cpu_ids;
2530
2531         /* XDP requires extra queues for XDP_TX */
2532         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2533                 netdev_warn(dev, "XDP request %i queues but max is %i. XDP_TX and XDP_REDIRECT will operate in a slower locked tx mode.\n",
2534                             curr_qp + xdp_qp, vi->max_queue_pairs);
2535                 xdp_qp = 0;
2536         }
2537
2538         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2539         if (!prog && !old_prog)
2540                 return 0;
2541
2542         if (prog)
2543                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2544
2545         /* Make sure NAPI is not using any XDP TX queues for RX. */
2546         if (netif_running(dev)) {
2547                 for (i = 0; i < vi->max_queue_pairs; i++) {
2548                         napi_disable(&vi->rq[i].napi);
2549                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2550                 }
2551         }
2552
2553         if (!prog) {
2554                 for (i = 0; i < vi->max_queue_pairs; i++) {
2555                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2556                         if (i == 0)
2557                                 virtnet_restore_guest_offloads(vi);
2558                 }
2559                 synchronize_net();
2560         }
2561
2562         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2563         if (err)
2564                 goto err;
2565         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2566         vi->xdp_queue_pairs = xdp_qp;
2567
2568         if (prog) {
2569                 vi->xdp_enabled = true;
2570                 for (i = 0; i < vi->max_queue_pairs; i++) {
2571                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2572                         if (i == 0 && !old_prog)
2573                                 virtnet_clear_guest_offloads(vi);
2574                 }
2575         } else {
2576                 vi->xdp_enabled = false;
2577         }
2578
2579         for (i = 0; i < vi->max_queue_pairs; i++) {
2580                 if (old_prog)
2581                         bpf_prog_put(old_prog);
2582                 if (netif_running(dev)) {
2583                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2584                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2585                                                &vi->sq[i].napi);
2586                 }
2587         }
2588
2589         return 0;
2590
2591 err:
2592         if (!prog) {
2593                 virtnet_clear_guest_offloads(vi);
2594                 for (i = 0; i < vi->max_queue_pairs; i++)
2595                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2596         }
2597
2598         if (netif_running(dev)) {
2599                 for (i = 0; i < vi->max_queue_pairs; i++) {
2600                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2601                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2602                                                &vi->sq[i].napi);
2603                 }
2604         }
2605         if (prog)
2606                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2607         return err;
2608 }
2609
2610 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2611 {
2612         switch (xdp->command) {
2613         case XDP_SETUP_PROG:
2614                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2615         default:
2616                 return -EINVAL;
2617         }
2618 }
2619
2620 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2621                                       size_t len)
2622 {
2623         struct virtnet_info *vi = netdev_priv(dev);
2624         int ret;
2625
2626         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2627                 return -EOPNOTSUPP;
2628
2629         ret = snprintf(buf, len, "sby");
2630         if (ret >= len)
2631                 return -EOPNOTSUPP;
2632
2633         return 0;
2634 }
2635
2636 static int virtnet_set_features(struct net_device *dev,
2637                                 netdev_features_t features)
2638 {
2639         struct virtnet_info *vi = netdev_priv(dev);
2640         u64 offloads;
2641         int err;
2642
2643         if ((dev->features ^ features) & NETIF_F_GRO_HW) {
2644                 if (vi->xdp_enabled)
2645                         return -EBUSY;
2646
2647                 if (features & NETIF_F_GRO_HW)
2648                         offloads = vi->guest_offloads_capable;
2649                 else
2650                         offloads = vi->guest_offloads_capable &
2651                                    ~GUEST_OFFLOAD_GRO_HW_MASK;
2652
2653                 err = virtnet_set_guest_offloads(vi, offloads);
2654                 if (err)
2655                         return err;
2656                 vi->guest_offloads = offloads;
2657         }
2658
2659         return 0;
2660 }
2661
2662 static const struct net_device_ops virtnet_netdev = {
2663         .ndo_open            = virtnet_open,
2664         .ndo_stop            = virtnet_close,
2665         .ndo_start_xmit      = start_xmit,
2666         .ndo_validate_addr   = eth_validate_addr,
2667         .ndo_set_mac_address = virtnet_set_mac_address,
2668         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2669         .ndo_get_stats64     = virtnet_stats,
2670         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2671         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2672         .ndo_bpf                = virtnet_xdp,
2673         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2674         .ndo_features_check     = passthru_features_check,
2675         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2676         .ndo_set_features       = virtnet_set_features,
2677 };
2678
2679 static void virtnet_config_changed_work(struct work_struct *work)
2680 {
2681         struct virtnet_info *vi =
2682                 container_of(work, struct virtnet_info, config_work);
2683         u16 v;
2684
2685         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2686                                  struct virtio_net_config, status, &v) < 0)
2687                 return;
2688
2689         if (v & VIRTIO_NET_S_ANNOUNCE) {
2690                 netdev_notify_peers(vi->dev);
2691                 virtnet_ack_link_announce(vi);
2692         }
2693
2694         /* Ignore unknown (future) status bits */
2695         v &= VIRTIO_NET_S_LINK_UP;
2696
2697         if (vi->status == v)
2698                 return;
2699
2700         vi->status = v;
2701
2702         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2703                 virtnet_update_settings(vi);
2704                 netif_carrier_on(vi->dev);
2705                 netif_tx_wake_all_queues(vi->dev);
2706         } else {
2707                 netif_carrier_off(vi->dev);
2708                 netif_tx_stop_all_queues(vi->dev);
2709         }
2710 }
2711
2712 static void virtnet_config_changed(struct virtio_device *vdev)
2713 {
2714         struct virtnet_info *vi = vdev->priv;
2715
2716         schedule_work(&vi->config_work);
2717 }
2718
2719 static void virtnet_free_queues(struct virtnet_info *vi)
2720 {
2721         int i;
2722
2723         for (i = 0; i < vi->max_queue_pairs; i++) {
2724                 __netif_napi_del(&vi->rq[i].napi);
2725                 __netif_napi_del(&vi->sq[i].napi);
2726         }
2727
2728         /* We called __netif_napi_del(),
2729          * we need to respect an RCU grace period before freeing vi->rq
2730          */
2731         synchronize_net();
2732
2733         kfree(vi->rq);
2734         kfree(vi->sq);
2735         kfree(vi->ctrl);
2736 }
2737
2738 static void _free_receive_bufs(struct virtnet_info *vi)
2739 {
2740         struct bpf_prog *old_prog;
2741         int i;
2742
2743         for (i = 0; i < vi->max_queue_pairs; i++) {
2744                 while (vi->rq[i].pages)
2745                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2746
2747                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2748                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2749                 if (old_prog)
2750                         bpf_prog_put(old_prog);
2751         }
2752 }
2753
2754 static void free_receive_bufs(struct virtnet_info *vi)
2755 {
2756         rtnl_lock();
2757         _free_receive_bufs(vi);
2758         rtnl_unlock();
2759 }
2760
2761 static void free_receive_page_frags(struct virtnet_info *vi)
2762 {
2763         int i;
2764         for (i = 0; i < vi->max_queue_pairs; i++)
2765                 if (vi->rq[i].alloc_frag.page)
2766                         put_page(vi->rq[i].alloc_frag.page);
2767 }
2768
2769 static void free_unused_bufs(struct virtnet_info *vi)
2770 {
2771         void *buf;
2772         int i;
2773
2774         for (i = 0; i < vi->max_queue_pairs; i++) {
2775                 struct virtqueue *vq = vi->sq[i].vq;
2776                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2777                         if (!is_xdp_frame(buf))
2778                                 dev_kfree_skb(buf);
2779                         else
2780                                 xdp_return_frame(ptr_to_xdp(buf));
2781                 }
2782         }
2783
2784         for (i = 0; i < vi->max_queue_pairs; i++) {
2785                 struct virtqueue *vq = vi->rq[i].vq;
2786
2787                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2788                         if (vi->mergeable_rx_bufs) {
2789                                 put_page(virt_to_head_page(buf));
2790                         } else if (vi->big_packets) {
2791                                 give_pages(&vi->rq[i], buf);
2792                         } else {
2793                                 put_page(virt_to_head_page(buf));
2794                         }
2795                 }
2796         }
2797 }
2798
2799 static void virtnet_del_vqs(struct virtnet_info *vi)
2800 {
2801         struct virtio_device *vdev = vi->vdev;
2802
2803         virtnet_clean_affinity(vi);
2804
2805         vdev->config->del_vqs(vdev);
2806
2807         virtnet_free_queues(vi);
2808 }
2809
2810 /* How large should a single buffer be so a queue full of these can fit at
2811  * least one full packet?
2812  * Logic below assumes the mergeable buffer header is used.
2813  */
2814 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2815 {
2816         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2817         unsigned int rq_size = virtqueue_get_vring_size(vq);
2818         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2819         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2820         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2821
2822         return max(max(min_buf_len, hdr_len) - hdr_len,
2823                    (unsigned int)GOOD_PACKET_LEN);
2824 }
2825
2826 static int virtnet_find_vqs(struct virtnet_info *vi)
2827 {
2828         vq_callback_t **callbacks;
2829         struct virtqueue **vqs;
2830         int ret = -ENOMEM;
2831         int i, total_vqs;
2832         const char **names;
2833         bool *ctx;
2834
2835         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2836          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2837          * possible control vq.
2838          */
2839         total_vqs = vi->max_queue_pairs * 2 +
2840                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2841
2842         /* Allocate space for find_vqs parameters */
2843         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2844         if (!vqs)
2845                 goto err_vq;
2846         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2847         if (!callbacks)
2848                 goto err_callback;
2849         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2850         if (!names)
2851                 goto err_names;
2852         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2853                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2854                 if (!ctx)
2855                         goto err_ctx;
2856         } else {
2857                 ctx = NULL;
2858         }
2859
2860         /* Parameters for control virtqueue, if any */
2861         if (vi->has_cvq) {
2862                 callbacks[total_vqs - 1] = NULL;
2863                 names[total_vqs - 1] = "control";
2864         }
2865
2866         /* Allocate/initialize parameters for send/receive virtqueues */
2867         for (i = 0; i < vi->max_queue_pairs; i++) {
2868                 callbacks[rxq2vq(i)] = skb_recv_done;
2869                 callbacks[txq2vq(i)] = skb_xmit_done;
2870                 sprintf(vi->rq[i].name, "input.%d", i);
2871                 sprintf(vi->sq[i].name, "output.%d", i);
2872                 names[rxq2vq(i)] = vi->rq[i].name;
2873                 names[txq2vq(i)] = vi->sq[i].name;
2874                 if (ctx)
2875                         ctx[rxq2vq(i)] = true;
2876         }
2877
2878         ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
2879                                   names, ctx, NULL);
2880         if (ret)
2881                 goto err_find;
2882
2883         if (vi->has_cvq) {
2884                 vi->cvq = vqs[total_vqs - 1];
2885                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2886                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2887         }
2888
2889         for (i = 0; i < vi->max_queue_pairs; i++) {
2890                 vi->rq[i].vq = vqs[rxq2vq(i)];
2891                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2892                 vi->sq[i].vq = vqs[txq2vq(i)];
2893         }
2894
2895         /* run here: ret == 0. */
2896
2897
2898 err_find:
2899         kfree(ctx);
2900 err_ctx:
2901         kfree(names);
2902 err_names:
2903         kfree(callbacks);
2904 err_callback:
2905         kfree(vqs);
2906 err_vq:
2907         return ret;
2908 }
2909
2910 static int virtnet_alloc_queues(struct virtnet_info *vi)
2911 {
2912         int i;
2913
2914         if (vi->has_cvq) {
2915                 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2916                 if (!vi->ctrl)
2917                         goto err_ctrl;
2918         } else {
2919                 vi->ctrl = NULL;
2920         }
2921         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2922         if (!vi->sq)
2923                 goto err_sq;
2924         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2925         if (!vi->rq)
2926                 goto err_rq;
2927
2928         INIT_DELAYED_WORK(&vi->refill, refill_work);
2929         for (i = 0; i < vi->max_queue_pairs; i++) {
2930                 vi->rq[i].pages = NULL;
2931                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2932                                napi_weight);
2933                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2934                                   napi_tx ? napi_weight : 0);
2935
2936                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2937                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2938                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2939
2940                 u64_stats_init(&vi->rq[i].stats.syncp);
2941                 u64_stats_init(&vi->sq[i].stats.syncp);
2942         }
2943
2944         return 0;
2945
2946 err_rq:
2947         kfree(vi->sq);
2948 err_sq:
2949         kfree(vi->ctrl);
2950 err_ctrl:
2951         return -ENOMEM;
2952 }
2953
2954 static int init_vqs(struct virtnet_info *vi)
2955 {
2956         int ret;
2957
2958         /* Allocate send & receive queues */
2959         ret = virtnet_alloc_queues(vi);
2960         if (ret)
2961                 goto err;
2962
2963         ret = virtnet_find_vqs(vi);
2964         if (ret)
2965                 goto err_free;
2966
2967         cpus_read_lock();
2968         virtnet_set_affinity(vi);
2969         cpus_read_unlock();
2970
2971         return 0;
2972
2973 err_free:
2974         virtnet_free_queues(vi);
2975 err:
2976         return ret;
2977 }
2978
2979 #ifdef CONFIG_SYSFS
2980 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2981                 char *buf)
2982 {
2983         struct virtnet_info *vi = netdev_priv(queue->dev);
2984         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2985         unsigned int headroom = virtnet_get_headroom(vi);
2986         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2987         struct ewma_pkt_len *avg;
2988
2989         BUG_ON(queue_index >= vi->max_queue_pairs);
2990         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2991         return sprintf(buf, "%u\n",
2992                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
2993                                        SKB_DATA_ALIGN(headroom + tailroom)));
2994 }
2995
2996 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2997         __ATTR_RO(mergeable_rx_buffer_size);
2998
2999 static struct attribute *virtio_net_mrg_rx_attrs[] = {
3000         &mergeable_rx_buffer_size_attribute.attr,
3001         NULL
3002 };
3003
3004 static const struct attribute_group virtio_net_mrg_rx_group = {
3005         .name = "virtio_net",
3006         .attrs = virtio_net_mrg_rx_attrs
3007 };
3008 #endif
3009
3010 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3011                                     unsigned int fbit,
3012                                     const char *fname, const char *dname)
3013 {
3014         if (!virtio_has_feature(vdev, fbit))
3015                 return false;
3016
3017         dev_err(&vdev->dev, "device advertises feature %s but not %s",
3018                 fname, dname);
3019
3020         return true;
3021 }
3022
3023 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
3024         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3025
3026 static bool virtnet_validate_features(struct virtio_device *vdev)
3027 {
3028         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3029             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3030                              "VIRTIO_NET_F_CTRL_VQ") ||
3031              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3032                              "VIRTIO_NET_F_CTRL_VQ") ||
3033              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3034                              "VIRTIO_NET_F_CTRL_VQ") ||
3035              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3036              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3037                              "VIRTIO_NET_F_CTRL_VQ"))) {
3038                 return false;
3039         }
3040
3041         return true;
3042 }
3043
3044 #define MIN_MTU ETH_MIN_MTU
3045 #define MAX_MTU ETH_MAX_MTU
3046
3047 static int virtnet_validate(struct virtio_device *vdev)
3048 {
3049         if (!vdev->config->get) {
3050                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3051                         __func__);
3052                 return -EINVAL;
3053         }
3054
3055         if (!virtnet_validate_features(vdev))
3056                 return -EINVAL;
3057
3058         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3059                 int mtu = virtio_cread16(vdev,
3060                                          offsetof(struct virtio_net_config,
3061                                                   mtu));
3062                 if (mtu < MIN_MTU)
3063                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3064         }
3065
3066         return 0;
3067 }
3068
3069 static int virtnet_probe(struct virtio_device *vdev)
3070 {
3071         int i, err = -ENOMEM;
3072         struct net_device *dev;
3073         struct virtnet_info *vi;
3074         u16 max_queue_pairs;
3075         int mtu;
3076
3077         /* Find if host supports multiqueue virtio_net device */
3078         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3079                                    struct virtio_net_config,
3080                                    max_virtqueue_pairs, &max_queue_pairs);
3081
3082         /* We need at least 2 queue's */
3083         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3084             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3085             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3086                 max_queue_pairs = 1;
3087
3088         /* Allocate ourselves a network device with room for our info */
3089         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3090         if (!dev)
3091                 return -ENOMEM;
3092
3093         /* Set up network device as normal. */
3094         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3095                            IFF_TX_SKB_NO_LINEAR;
3096         dev->netdev_ops = &virtnet_netdev;
3097         dev->features = NETIF_F_HIGHDMA;
3098
3099         dev->ethtool_ops = &virtnet_ethtool_ops;
3100         SET_NETDEV_DEV(dev, &vdev->dev);
3101
3102         /* Do we support "hardware" checksums? */
3103         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3104                 /* This opens up the world of extra features. */
3105                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3106                 if (csum)
3107                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3108
3109                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3110                         dev->hw_features |= NETIF_F_TSO
3111                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3112                 }
3113                 /* Individual feature bits: what can host handle? */
3114                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3115                         dev->hw_features |= NETIF_F_TSO;
3116                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3117                         dev->hw_features |= NETIF_F_TSO6;
3118                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3119                         dev->hw_features |= NETIF_F_TSO_ECN;
3120
3121                 dev->features |= NETIF_F_GSO_ROBUST;
3122
3123                 if (gso)
3124                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3125                 /* (!csum && gso) case will be fixed by register_netdev() */
3126         }
3127         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3128                 dev->features |= NETIF_F_RXCSUM;
3129         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3130             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3131                 dev->features |= NETIF_F_GRO_HW;
3132         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3133                 dev->hw_features |= NETIF_F_GRO_HW;
3134
3135         dev->vlan_features = dev->features;
3136
3137         /* MTU range: 68 - 65535 */
3138         dev->min_mtu = MIN_MTU;
3139         dev->max_mtu = MAX_MTU;
3140
3141         /* Configuration may specify what MAC to use.  Otherwise random. */
3142         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3143                 virtio_cread_bytes(vdev,
3144                                    offsetof(struct virtio_net_config, mac),
3145                                    dev->dev_addr, dev->addr_len);
3146         else
3147                 eth_hw_addr_random(dev);
3148
3149         /* Set up our device-specific information */
3150         vi = netdev_priv(dev);
3151         vi->dev = dev;
3152         vi->vdev = vdev;
3153         vdev->priv = vi;
3154
3155         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3156
3157         /* If we can receive ANY GSO packets, we must allocate large ones. */
3158         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3159             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3160             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3161             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3162                 vi->big_packets = true;
3163
3164         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3165                 vi->mergeable_rx_bufs = true;
3166
3167         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3168             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3169                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3170         else
3171                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3172
3173         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3174             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3175                 vi->any_header_sg = true;
3176
3177         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3178                 vi->has_cvq = true;
3179
3180         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3181                 mtu = virtio_cread16(vdev,
3182                                      offsetof(struct virtio_net_config,
3183                                               mtu));
3184                 if (mtu < dev->min_mtu) {
3185                         /* Should never trigger: MTU was previously validated
3186                          * in virtnet_validate.
3187                          */
3188                         dev_err(&vdev->dev,
3189                                 "device MTU appears to have changed it is now %d < %d",
3190                                 mtu, dev->min_mtu);
3191                         err = -EINVAL;
3192                         goto free;
3193                 }
3194
3195                 dev->mtu = mtu;
3196                 dev->max_mtu = mtu;
3197
3198                 /* TODO: size buffers correctly in this case. */
3199                 if (dev->mtu > ETH_DATA_LEN)
3200                         vi->big_packets = true;
3201         }
3202
3203         if (vi->any_header_sg)
3204                 dev->needed_headroom = vi->hdr_len;
3205
3206         /* Enable multiqueue by default */
3207         if (num_online_cpus() >= max_queue_pairs)
3208                 vi->curr_queue_pairs = max_queue_pairs;
3209         else
3210                 vi->curr_queue_pairs = num_online_cpus();
3211         vi->max_queue_pairs = max_queue_pairs;
3212
3213         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3214         err = init_vqs(vi);
3215         if (err)
3216                 goto free;
3217
3218 #ifdef CONFIG_SYSFS
3219         if (vi->mergeable_rx_bufs)
3220                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3221 #endif
3222         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3223         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3224
3225         virtnet_init_settings(dev);
3226
3227         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3228                 vi->failover = net_failover_create(vi->dev);
3229                 if (IS_ERR(vi->failover)) {
3230                         err = PTR_ERR(vi->failover);
3231                         goto free_vqs;
3232                 }
3233         }
3234
3235         err = register_netdev(dev);
3236         if (err) {
3237                 pr_debug("virtio_net: registering device failed\n");
3238                 goto free_failover;
3239         }
3240
3241         virtio_device_ready(vdev);
3242
3243         err = virtnet_cpu_notif_add(vi);
3244         if (err) {
3245                 pr_debug("virtio_net: registering cpu notifier failed\n");
3246                 goto free_unregister_netdev;
3247         }
3248
3249         virtnet_set_queues(vi, vi->curr_queue_pairs);
3250
3251         /* Assume link up if device can't report link status,
3252            otherwise get link status from config. */
3253         netif_carrier_off(dev);
3254         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3255                 schedule_work(&vi->config_work);
3256         } else {
3257                 vi->status = VIRTIO_NET_S_LINK_UP;
3258                 virtnet_update_settings(vi);
3259                 netif_carrier_on(dev);
3260         }
3261
3262         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3263                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3264                         set_bit(guest_offloads[i], &vi->guest_offloads);
3265         vi->guest_offloads_capable = vi->guest_offloads;
3266
3267         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3268                  dev->name, max_queue_pairs);
3269
3270         return 0;
3271
3272 free_unregister_netdev:
3273         vi->vdev->config->reset(vdev);
3274
3275         unregister_netdev(dev);
3276 free_failover:
3277         net_failover_destroy(vi->failover);
3278 free_vqs:
3279         cancel_delayed_work_sync(&vi->refill);
3280         free_receive_page_frags(vi);
3281         virtnet_del_vqs(vi);
3282 free:
3283         free_netdev(dev);
3284         return err;
3285 }
3286
3287 static void remove_vq_common(struct virtnet_info *vi)
3288 {
3289         vi->vdev->config->reset(vi->vdev);
3290
3291         /* Free unused buffers in both send and recv, if any. */
3292         free_unused_bufs(vi);
3293
3294         free_receive_bufs(vi);
3295
3296         free_receive_page_frags(vi);
3297
3298         virtnet_del_vqs(vi);
3299 }
3300
3301 static void virtnet_remove(struct virtio_device *vdev)
3302 {
3303         struct virtnet_info *vi = vdev->priv;
3304
3305         virtnet_cpu_notif_remove(vi);
3306
3307         /* Make sure no work handler is accessing the device. */
3308         flush_work(&vi->config_work);
3309
3310         unregister_netdev(vi->dev);
3311
3312         net_failover_destroy(vi->failover);
3313
3314         remove_vq_common(vi);
3315
3316         free_netdev(vi->dev);
3317 }
3318
3319 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3320 {
3321         struct virtnet_info *vi = vdev->priv;
3322
3323         virtnet_cpu_notif_remove(vi);
3324         virtnet_freeze_down(vdev);
3325         remove_vq_common(vi);
3326
3327         return 0;
3328 }
3329
3330 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3331 {
3332         struct virtnet_info *vi = vdev->priv;
3333         int err;
3334
3335         err = virtnet_restore_up(vdev);
3336         if (err)
3337                 return err;
3338         virtnet_set_queues(vi, vi->curr_queue_pairs);
3339
3340         err = virtnet_cpu_notif_add(vi);
3341         if (err) {
3342                 virtnet_freeze_down(vdev);
3343                 remove_vq_common(vi);
3344                 return err;
3345         }
3346
3347         return 0;
3348 }
3349
3350 static struct virtio_device_id id_table[] = {
3351         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3352         { 0 },
3353 };
3354
3355 #define VIRTNET_FEATURES \
3356         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3357         VIRTIO_NET_F_MAC, \
3358         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3359         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3360         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3361         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3362         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3363         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3364         VIRTIO_NET_F_CTRL_MAC_ADDR, \
3365         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3366         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3367
3368 static unsigned int features[] = {
3369         VIRTNET_FEATURES,
3370 };
3371
3372 static unsigned int features_legacy[] = {
3373         VIRTNET_FEATURES,
3374         VIRTIO_NET_F_GSO,
3375         VIRTIO_F_ANY_LAYOUT,
3376 };
3377
3378 static struct virtio_driver virtio_net_driver = {
3379         .feature_table = features,
3380         .feature_table_size = ARRAY_SIZE(features),
3381         .feature_table_legacy = features_legacy,
3382         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3383         .driver.name =  KBUILD_MODNAME,
3384         .driver.owner = THIS_MODULE,
3385         .id_table =     id_table,
3386         .validate =     virtnet_validate,
3387         .probe =        virtnet_probe,
3388         .remove =       virtnet_remove,
3389         .config_changed = virtnet_config_changed,
3390 #ifdef CONFIG_PM_SLEEP
3391         .freeze =       virtnet_freeze,
3392         .restore =      virtnet_restore,
3393 #endif
3394 };
3395
3396 static __init int virtio_net_driver_init(void)
3397 {
3398         int ret;
3399
3400         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3401                                       virtnet_cpu_online,
3402                                       virtnet_cpu_down_prep);
3403         if (ret < 0)
3404                 goto out;
3405         virtionet_online = ret;
3406         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3407                                       NULL, virtnet_cpu_dead);
3408         if (ret)
3409                 goto err_dead;
3410
3411         ret = register_virtio_driver(&virtio_net_driver);
3412         if (ret)
3413                 goto err_virtio;
3414         return 0;
3415 err_virtio:
3416         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3417 err_dead:
3418         cpuhp_remove_multi_state(virtionet_online);
3419 out:
3420         return ret;
3421 }
3422 module_init(virtio_net_driver_init);
3423
3424 static __exit void virtio_net_driver_exit(void)
3425 {
3426         unregister_virtio_driver(&virtio_net_driver);
3427         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3428         cpuhp_remove_multi_state(virtionet_online);
3429 }
3430 module_exit(virtio_net_driver_exit);
3431
3432 MODULE_DEVICE_TABLE(virtio, id_table);
3433 MODULE_DESCRIPTION("Virtio network driver");
3434 MODULE_LICENSE("GPL");