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 delayed refilling if we run low on memory. */
 217         struct delayed_work refill;
 218
 219         /* Is delayed refill enabled? */
 220         bool refill_enabled;
 221
 222         /* The lock to synchronize the access to refill_enabled */
 223         spinlock_t refill_lock;
 224
 225         /* Work struct for config space updates */
 226         struct work_struct config_work;
 227
 228         /* Does the affinity hint is set for virtqueues? */
 229         bool affinity_hint_set;
 230
 231         /* CPU hotplug instances for online & dead */
 232         struct hlist_node node;
 233         struct hlist_node node_dead;
 234
 235         struct control_buf *ctrl;
 236
 237         /* Ethtool settings */
 238         u8 duplex;
 239         u32 speed;
 240
 241         unsigned long guest_offloads;
 242         unsigned long guest_offloads_capable;
 243
 244         /* failover when STANDBY feature enabled */
 245         struct failover *failover;
 246 };
 247
 248 struct padded_vnet_hdr {
 249         struct virtio_net_hdr_mrg_rxbuf hdr;
 250         /*
 251          * hdr is in a separate sg buffer, and data sg buffer shares same page
 252          * with this header sg. This padding makes next sg 16 byte aligned
 253          * after the header.
 254          */
 255         char padding[4];
 256 };
 257
 258 static bool is_xdp_frame(void *ptr)
 259 {
 260         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 261 }
 262
 263 static void *xdp_to_ptr(struct xdp_frame *ptr)
 264 {
 265         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 266 }
 267
 268 static struct xdp_frame *ptr_to_xdp(void *ptr)
 269 {
 270         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 271 }
 272
 273 /* Converting between virtqueue no. and kernel tx/rx queue no.
 274  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 275  */
 276 static int vq2txq(struct virtqueue *vq)
 277 {
 278         return (vq->index - 1) / 2;
 279 }
 280
 281 static int txq2vq(int txq)
 282 {
 283         return txq * 2 + 1;
 284 }
 285
 286 static int vq2rxq(struct virtqueue *vq)
 287 {
 288         return vq->index / 2;
 289 }
 290
 291 static int rxq2vq(int rxq)
 292 {
 293         return rxq * 2;
 294 }
 295
 296 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 297 {
 298         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 299 }
 300
 301 /*
 302  * private is used to chain pages for big packets, put the whole
 303  * most recent used list in the beginning for reuse
 304  */
 305 static void give_pages(struct receive_queue *rq, struct page *page)
 306 {
 307         struct page *end;
 308
 309         /* Find end of list, sew whole thing into vi->rq.pages. */
 310         for (end = page; end->private; end = (struct page *)end->private);
 311         end->private = (unsigned long)rq->pages;
 312         rq->pages = page;
 313 }
 314
 315 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 316 {
 317         struct page *p = rq->pages;
 318
 319         if (p) {
 320                 rq->pages = (struct page *)p->private;
 321                 /* clear private here, it is used to chain pages */
 322                 p->private = 0;
 323         } else
 324                 p = alloc_page(gfp_mask);
 325         return p;
 326 }
 327
 328 static void enable_delayed_refill(struct virtnet_info *vi)
 329 {
 330         spin_lock_bh(&vi->refill_lock);
 331         vi->refill_enabled = true;
 332         spin_unlock_bh(&vi->refill_lock);
 333 }
 334
 335 static void disable_delayed_refill(struct virtnet_info *vi)
 336 {
 337         spin_lock_bh(&vi->refill_lock);
 338         vi->refill_enabled = false;
 339         spin_unlock_bh(&vi->refill_lock);
 340 }
 341
 342 static void virtqueue_napi_schedule(struct napi_struct *napi,
 343                                     struct virtqueue *vq)
 344 {
 345         if (napi_schedule_prep(napi)) {
 346                 virtqueue_disable_cb(vq);
 347                 __napi_schedule(napi);
 348         }
 349 }
 350
 351 static void virtqueue_napi_complete(struct napi_struct *napi,
 352                                     struct virtqueue *vq, int processed)
 353 {
 354         int opaque;
 355
 356         opaque = virtqueue_enable_cb_prepare(vq);
 357         if (napi_complete_done(napi, processed)) {
 358                 if (unlikely(virtqueue_poll(vq, opaque)))
 359                         virtqueue_napi_schedule(napi, vq);
 360         } else {
 361                 virtqueue_disable_cb(vq);
 362         }
 363 }
 364
 365 static void skb_xmit_done(struct virtqueue *vq)
 366 {
 367         struct virtnet_info *vi = vq->vdev->priv;
 368         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 369
 370         /* Suppress further interrupts. */
 371         virtqueue_disable_cb(vq);
 372
 373         if (napi->weight)
 374                 virtqueue_napi_schedule(napi, vq);
 375         else
 376                 /* We were probably waiting for more output buffers. */
 377                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 378 }
 379
 380 #define MRG_CTX_HEADER_SHIFT 22
 381 static void *mergeable_len_to_ctx(unsigned int truesize,
 382                                   unsigned int headroom)
 383 {
 384         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 385 }
 386
 387 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 388 {
 389         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 390 }
 391
 392 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 393 {
 394         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 395 }
 396
 397 /* Called from bottom half context */
 398 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 399                                    struct receive_queue *rq,
 400                                    struct page *page, unsigned int offset,
 401                                    unsigned int len, unsigned int truesize,
 402                                    bool hdr_valid, unsigned int metasize)
 403 {
 404         struct sk_buff *skb;
 405         struct virtio_net_hdr_mrg_rxbuf *hdr;
 406         unsigned int copy, hdr_len, hdr_padded_len;
 407         char *p;
 408
 409         p = page_address(page) + offset;
 410
 411         /* copy small packet so we can reuse these pages for small data */
 412         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 413         if (unlikely(!skb))
 414                 return NULL;
 415
 416         hdr = skb_vnet_hdr(skb);
 417
 418         hdr_len = vi->hdr_len;
 419         if (vi->mergeable_rx_bufs)
 420                 hdr_padded_len = sizeof(*hdr);
 421         else
 422                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 423
 424         /* hdr_valid means no XDP, so we can copy the vnet header */
 425         if (hdr_valid)
 426                 memcpy(hdr, p, hdr_len);
 427
 428         len -= hdr_len;
 429         offset += hdr_padded_len;
 430         p += hdr_padded_len;
 431
 432         /* Copy all frame if it fits skb->head, otherwise
 433          * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
 434          */
 435         if (len <= skb_tailroom(skb))
 436                 copy = len;
 437         else
 438                 copy = ETH_HLEN + metasize;
 439         skb_put_data(skb, p, copy);
 440
 441         if (metasize) {
 442                 __skb_pull(skb, metasize);
 443                 skb_metadata_set(skb, metasize);
 444         }
 445
 446         len -= copy;
 447         offset += copy;
 448
 449         if (vi->mergeable_rx_bufs) {
 450                 if (len)
 451                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 452                 else
 453                         put_page(page);
 454                 return skb;
 455         }
 456
 457         /*
 458          * Verify that we can indeed put this data into a skb.
 459          * This is here to handle cases when the device erroneously
 460          * tries to receive more than is possible. This is usually
 461          * the case of a broken device.
 462          */
 463         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 464                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 465                 dev_kfree_skb(skb);
 466                 return NULL;
 467         }
 468         BUG_ON(offset >= PAGE_SIZE);
 469         while (len) {
 470                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 471                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 472                                 frag_size, truesize);
 473                 len -= frag_size;
 474                 page = (struct page *)page->private;
 475                 offset = 0;
 476         }
 477
 478         if (page)
 479                 give_pages(rq, page);
 480
 481         return skb;
 482 }
 483
 484 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 485                                    struct send_queue *sq,
 486                                    struct xdp_frame *xdpf)
 487 {
 488         struct virtio_net_hdr_mrg_rxbuf *hdr;
 489         int err;
 490
 491         if (unlikely(xdpf->headroom < vi->hdr_len))
 492                 return -EOVERFLOW;
 493
 494         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 495         xdpf->data -= vi->hdr_len;
 496         /* Zero header and leave csum up to XDP layers */
 497         hdr = xdpf->data;
 498         memset(hdr, 0, vi->hdr_len);
 499         xdpf->len   += vi->hdr_len;
 500
 501         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 502
 503         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
 504                                    GFP_ATOMIC);
 505         if (unlikely(err))
 506                 return -ENOSPC; /* Caller handle free/refcnt */
 507
 508         return 0;
 509 }
 510
 511 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
 512  * the current cpu, so it does not need to be locked.
 513  *
 514  * Here we use marco instead of inline functions because we have to deal with
 515  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
 516  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
 517  * functions to perfectly solve these three problems at the same time.
 518  */
 519 #define virtnet_xdp_get_sq(vi) ({                                       \
 520         struct netdev_queue *txq;                                       \
 521         typeof(vi) v = (vi);                                            \
 522         unsigned int qp;                                                \
 523                                                                         \
 524         if (v->curr_queue_pairs > nr_cpu_ids) {                         \
 525                 qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
 526                 qp += smp_processor_id();                               \
 527                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 528                 __netif_tx_acquire(txq);                                \
 529         } else {                                                        \
 530                 qp = smp_processor_id() % v->curr_queue_pairs;          \
 531                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 532                 __netif_tx_lock(txq, raw_smp_processor_id());           \
 533         }                                                               \
 534         v->sq + qp;                                                     \
 535 })
 536
 537 #define virtnet_xdp_put_sq(vi, q) {                                     \
 538         struct netdev_queue *txq;                                       \
 539         typeof(vi) v = (vi);                                            \
 540                                                                         \
 541         txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
 542         if (v->curr_queue_pairs > nr_cpu_ids)                           \
 543                 __netif_tx_release(txq);                                \
 544         else                                                            \
 545                 __netif_tx_unlock(txq);                                 \
 546 }
 547
 548 static int virtnet_xdp_xmit(struct net_device *dev,
 549                             int n, struct xdp_frame **frames, u32 flags)
 550 {
 551         struct virtnet_info *vi = netdev_priv(dev);
 552         struct receive_queue *rq = vi->rq;
 553         struct bpf_prog *xdp_prog;
 554         struct send_queue *sq;
 555         unsigned int len;
 556         int packets = 0;
 557         int bytes = 0;
 558         int drops = 0;
 559         int kicks = 0;
 560         int ret, err;
 561         void *ptr;
 562         int i;
 563
 564         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 565          * indicate XDP resources have been successfully allocated.
 566          */
 567         xdp_prog = rcu_access_pointer(rq->xdp_prog);
 568         if (!xdp_prog)
 569                 return -ENXIO;
 570
 571         sq = virtnet_xdp_get_sq(vi);
 572
 573         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 574                 ret = -EINVAL;
 575                 drops = n;
 576                 goto out;
 577         }
 578
 579         /* Free up any pending old buffers before queueing new ones. */
 580         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 581                 if (likely(is_xdp_frame(ptr))) {
 582                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 583
 584                         bytes += frame->len;
 585                         xdp_return_frame(frame);
 586                 } else {
 587                         struct sk_buff *skb = ptr;
 588
 589                         bytes += skb->len;
 590                         napi_consume_skb(skb, false);
 591                 }
 592                 packets++;
 593         }
 594
 595         for (i = 0; i < n; i++) {
 596                 struct xdp_frame *xdpf = frames[i];
 597
 598                 err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
 599                 if (err) {
 600                         xdp_return_frame_rx_napi(xdpf);
 601                         drops++;
 602                 }
 603         }
 604         ret = n - drops;
 605
 606         if (flags & XDP_XMIT_FLUSH) {
 607                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
 608                         kicks = 1;
 609         }
 610 out:
 611         u64_stats_update_begin(&sq->stats.syncp);
 612         sq->stats.bytes += bytes;
 613         sq->stats.packets += packets;
 614         sq->stats.xdp_tx += n;
 615         sq->stats.xdp_tx_drops += drops;
 616         sq->stats.kicks += kicks;
 617         u64_stats_update_end(&sq->stats.syncp);
 618
 619         virtnet_xdp_put_sq(vi, sq);
 620         return ret;
 621 }
 622
 623 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 624 {
 625         return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
 626 }
 627
 628 /* We copy the packet for XDP in the following cases:
 629  *
 630  * 1) Packet is scattered across multiple rx buffers.
 631  * 2) Headroom space is insufficient.
 632  *
 633  * This is inefficient but it's a temporary condition that
 634  * we hit right after XDP is enabled and until queue is refilled
 635  * with large buffers with sufficient headroom - so it should affect
 636  * at most queue size packets.
 637  * Afterwards, the conditions to enable
 638  * XDP should preclude the underlying device from sending packets
 639  * across multiple buffers (num_buf > 1), and we make sure buffers
 640  * have enough headroom.
 641  */
 642 static struct page *xdp_linearize_page(struct receive_queue *rq,
 643                                        u16 *num_buf,
 644                                        struct page *p,
 645                                        int offset,
 646                                        int page_off,
 647                                        unsigned int *len)
 648 {
 649         struct page *page = alloc_page(GFP_ATOMIC);
 650
 651         if (!page)
 652                 return NULL;
 653
 654         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 655         page_off += *len;
 656
 657         while (--*num_buf) {
 658                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 659                 unsigned int buflen;
 660                 void *buf;
 661                 int off;
 662
 663                 buf = virtqueue_get_buf(rq->vq, &buflen);
 664                 if (unlikely(!buf))
 665                         goto err_buf;
 666
 667                 p = virt_to_head_page(buf);
 668                 off = buf - page_address(p);
 669
 670                 /* guard against a misconfigured or uncooperative backend that
 671                  * is sending packet larger than the MTU.
 672                  */
 673                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 674                         put_page(p);
 675                         goto err_buf;
 676                 }
 677
 678                 memcpy(page_address(page) + page_off,
 679                        page_address(p) + off, buflen);
 680                 page_off += buflen;
 681                 put_page(p);
 682         }
 683
 684         /* Headroom does not contribute to packet length */
 685         *len = page_off - VIRTIO_XDP_HEADROOM;
 686         return page;
 687 err_buf:
 688         __free_pages(page, 0);
 689         return NULL;
 690 }
 691
 692 static struct sk_buff *receive_small(struct net_device *dev,
 693                                      struct virtnet_info *vi,
 694                                      struct receive_queue *rq,
 695                                      void *buf, void *ctx,
 696                                      unsigned int len,
 697                                      unsigned int *xdp_xmit,
 698                                      struct virtnet_rq_stats *stats)
 699 {
 700         struct sk_buff *skb;
 701         struct bpf_prog *xdp_prog;
 702         unsigned int xdp_headroom = (unsigned long)ctx;
 703         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 704         unsigned int headroom = vi->hdr_len + header_offset;
 705         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 706                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 707         struct page *page = virt_to_head_page(buf);
 708         unsigned int delta = 0;
 709         struct page *xdp_page;
 710         int err;
 711         unsigned int metasize = 0;
 712
 713         len -= vi->hdr_len;
 714         stats->bytes += len;
 715
 716         if (unlikely(len > GOOD_PACKET_LEN)) {
 717                 pr_debug("%s: rx error: len %u exceeds max size %d\n",
 718                          dev->name, len, GOOD_PACKET_LEN);
 719                 dev->stats.rx_length_errors++;
 720                 goto err_len;
 721         }
 722         rcu_read_lock();
 723         xdp_prog = rcu_dereference(rq->xdp_prog);
 724         if (xdp_prog) {
 725                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 726                 struct xdp_frame *xdpf;
 727                 struct xdp_buff xdp;
 728                 void *orig_data;
 729                 u32 act;
 730
 731                 if (unlikely(hdr->hdr.gso_type))
 732                         goto err_xdp;
 733
 734                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 735                         int offset = buf - page_address(page) + header_offset;
 736                         unsigned int tlen = len + vi->hdr_len;
 737                         u16 num_buf = 1;
 738
 739                         xdp_headroom = virtnet_get_headroom(vi);
 740                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 741                         headroom = vi->hdr_len + header_offset;
 742                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 743                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 744                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 745                                                       offset, header_offset,
 746                                                       &tlen);
 747                         if (!xdp_page)
 748                                 goto err_xdp;
 749
 750                         buf = page_address(xdp_page);
 751                         put_page(page);
 752                         page = xdp_page;
 753                 }
 754
 755                 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
 756                 xdp.data = xdp.data_hard_start + xdp_headroom;
 757                 xdp.data_end = xdp.data + len;
 758                 xdp.data_meta = xdp.data;
 759                 xdp.rxq = &rq->xdp_rxq;
 760                 xdp.frame_sz = buflen;
 761                 orig_data = xdp.data;
 762                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 763                 stats->xdp_packets++;
 764
 765                 switch (act) {
 766                 case XDP_PASS:
 767                         /* Recalculate length in case bpf program changed it */
 768                         delta = orig_data - xdp.data;
 769                         len = xdp.data_end - xdp.data;
 770                         metasize = xdp.data - xdp.data_meta;
 771                         break;
 772                 case XDP_TX:
 773                         stats->xdp_tx++;
 774                         xdpf = xdp_convert_buff_to_frame(&xdp);
 775                         if (unlikely(!xdpf))
 776                                 goto err_xdp;
 777                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 778                         if (unlikely(err < 0)) {
 779                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 780                                 goto err_xdp;
 781                         }
 782                         *xdp_xmit |= VIRTIO_XDP_TX;
 783                         rcu_read_unlock();
 784                         goto xdp_xmit;
 785                 case XDP_REDIRECT:
 786                         stats->xdp_redirects++;
 787                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 788                         if (err)
 789                                 goto err_xdp;
 790                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 791                         rcu_read_unlock();
 792                         goto xdp_xmit;
 793                 default:
 794                         bpf_warn_invalid_xdp_action(act);
 795                         fallthrough;
 796                 case XDP_ABORTED:
 797                         trace_xdp_exception(vi->dev, xdp_prog, act);
 798                 case XDP_DROP:
 799                         goto err_xdp;
 800                 }
 801         }
 802         rcu_read_unlock();
 803
 804         skb = build_skb(buf, buflen);
 805         if (!skb) {
 806                 put_page(page);
 807                 goto err;
 808         }
 809         skb_reserve(skb, headroom - delta);
 810         skb_put(skb, len);
 811         if (!xdp_prog) {
 812                 buf += header_offset;
 813                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 814         } /* keep zeroed vnet hdr since XDP is loaded */
 815
 816         if (metasize)
 817                 skb_metadata_set(skb, metasize);
 818
 819 err:
 820         return skb;
 821
 822 err_xdp:
 823         rcu_read_unlock();
 824         stats->xdp_drops++;
 825 err_len:
 826         stats->drops++;
 827         put_page(page);
 828 xdp_xmit:
 829         return NULL;
 830 }
 831
 832 static struct sk_buff *receive_big(struct net_device *dev,
 833                                    struct virtnet_info *vi,
 834                                    struct receive_queue *rq,
 835                                    void *buf,
 836                                    unsigned int len,
 837                                    struct virtnet_rq_stats *stats)
 838 {
 839         struct page *page = buf;
 840         struct sk_buff *skb =
 841                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 842
 843         stats->bytes += len - vi->hdr_len;
 844         if (unlikely(!skb))
 845                 goto err;
 846
 847         return skb;
 848
 849 err:
 850         stats->drops++;
 851         give_pages(rq, page);
 852         return NULL;
 853 }
 854
 855 static struct sk_buff *receive_mergeable(struct net_device *dev,
 856                                          struct virtnet_info *vi,
 857                                          struct receive_queue *rq,
 858                                          void *buf,
 859                                          void *ctx,
 860                                          unsigned int len,
 861                                          unsigned int *xdp_xmit,
 862                                          struct virtnet_rq_stats *stats)
 863 {
 864         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 865         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 866         struct page *page = virt_to_head_page(buf);
 867         int offset = buf - page_address(page);
 868         struct sk_buff *head_skb, *curr_skb;
 869         struct bpf_prog *xdp_prog;
 870         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
 871         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 872         unsigned int metasize = 0;
 873         unsigned int frame_sz;
 874         int err;
 875
 876         head_skb = NULL;
 877         stats->bytes += len - vi->hdr_len;
 878
 879         if (unlikely(len > truesize)) {
 880                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 881                          dev->name, len, (unsigned long)ctx);
 882                 dev->stats.rx_length_errors++;
 883                 goto err_skb;
 884         }
 885         rcu_read_lock();
 886         xdp_prog = rcu_dereference(rq->xdp_prog);
 887         if (xdp_prog) {
 888                 struct xdp_frame *xdpf;
 889                 struct page *xdp_page;
 890                 struct xdp_buff xdp;
 891                 void *data;
 892                 u32 act;
 893
 894                 /* Transient failure which in theory could occur if
 895                  * in-flight packets from before XDP was enabled reach
 896                  * the receive path after XDP is loaded.
 897                  */
 898                 if (unlikely(hdr->hdr.gso_type))
 899                         goto err_xdp;
 900
 901                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 902                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 903                  */
 904                 frame_sz = headroom ? PAGE_SIZE : truesize;
 905
 906                 /* This happens when rx buffer size is underestimated
 907                  * or headroom is not enough because of the buffer
 908                  * was refilled before XDP is set. This should only
 909                  * happen for the first several packets, so we don't
 910                  * care much about its performance.
 911                  */
 912                 if (unlikely(num_buf > 1 ||
 913                              headroom < virtnet_get_headroom(vi))) {
 914                         /* linearize data for XDP */
 915                         xdp_page = xdp_linearize_page(rq, &num_buf,
 916                                                       page, offset,
 917                                                       VIRTIO_XDP_HEADROOM,
 918                                                       &len);
 919                         frame_sz = PAGE_SIZE;
 920
 921                         if (!xdp_page)
 922                                 goto err_xdp;
 923                         offset = VIRTIO_XDP_HEADROOM;
 924                 } else {
 925                         xdp_page = page;
 926                 }
 927
 928                 /* Allow consuming headroom but reserve enough space to push
 929                  * the descriptor on if we get an XDP_TX return code.
 930                  */
 931                 data = page_address(xdp_page) + offset;
 932                 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
 933                 xdp.data = data + vi->hdr_len;
 934                 xdp.data_end = xdp.data + (len - vi->hdr_len);
 935                 xdp.data_meta = xdp.data;
 936                 xdp.rxq = &rq->xdp_rxq;
 937                 xdp.frame_sz = frame_sz - vi->hdr_len;
 938
 939                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 940                 stats->xdp_packets++;
 941
 942                 switch (act) {
 943                 case XDP_PASS:
 944                         metasize = xdp.data - xdp.data_meta;
 945
 946                         /* recalculate offset to account for any header
 947                          * adjustments and minus the metasize to copy the
 948                          * metadata in page_to_skb(). Note other cases do not
 949                          * build an skb and avoid using offset
 950                          */
 951                         offset = xdp.data - page_address(xdp_page) -
 952                                  vi->hdr_len - metasize;
 953
 954                         /* recalculate len if xdp.data, xdp.data_end or
 955                          * xdp.data_meta were adjusted
 956                          */
 957                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 958                         /* We can only create skb based on xdp_page. */
 959                         if (unlikely(xdp_page != page)) {
 960                                 rcu_read_unlock();
 961                                 put_page(page);
 962                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
 963                                                        len, PAGE_SIZE, false,
 964                                                        metasize);
 965                                 return head_skb;
 966                         }
 967                         break;
 968                 case XDP_TX:
 969                         stats->xdp_tx++;
 970                         xdpf = xdp_convert_buff_to_frame(&xdp);
 971                         if (unlikely(!xdpf)) {
 972                                 if (unlikely(xdp_page != page))
 973                                         put_page(xdp_page);
 974                                 goto err_xdp;
 975                         }
 976                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 977                         if (unlikely(err < 0)) {
 978                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 979                                 if (unlikely(xdp_page != page))
 980                                         put_page(xdp_page);
 981                                 goto err_xdp;
 982                         }
 983                         *xdp_xmit |= VIRTIO_XDP_TX;
 984                         if (unlikely(xdp_page != page))
 985                                 put_page(page);
 986                         rcu_read_unlock();
 987                         goto xdp_xmit;
 988                 case XDP_REDIRECT:
 989                         stats->xdp_redirects++;
 990                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 991                         if (err) {
 992                                 if (unlikely(xdp_page != page))
 993                                         put_page(xdp_page);
 994                                 goto err_xdp;
 995                         }
 996                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 997                         if (unlikely(xdp_page != page))
 998                                 put_page(page);
 999                         rcu_read_unlock();
1000                         goto xdp_xmit;
1001                 default:
1002                         bpf_warn_invalid_xdp_action(act);
1003                         fallthrough;
1004                 case XDP_ABORTED:
1005                         trace_xdp_exception(vi->dev, xdp_prog, act);
1006                         fallthrough;
1007                 case XDP_DROP:
1008                         if (unlikely(xdp_page != page))
1009                                 __free_pages(xdp_page, 0);
1010                         goto err_xdp;
1011                 }
1012         }
1013         rcu_read_unlock();
1014
1015         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1016                                metasize);
1017         curr_skb = head_skb;
1018
1019         if (unlikely(!curr_skb))
1020                 goto err_skb;
1021         while (--num_buf) {
1022                 int num_skb_frags;
1023
1024                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1025                 if (unlikely(!buf)) {
1026                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1027                                  dev->name, num_buf,
1028                                  virtio16_to_cpu(vi->vdev,
1029                                                  hdr->num_buffers));
1030                         dev->stats.rx_length_errors++;
1031                         goto err_buf;
1032                 }
1033
1034                 stats->bytes += len;
1035                 page = virt_to_head_page(buf);
1036
1037                 truesize = mergeable_ctx_to_truesize(ctx);
1038                 if (unlikely(len > truesize)) {
1039                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1040                                  dev->name, len, (unsigned long)ctx);
1041                         dev->stats.rx_length_errors++;
1042                         goto err_skb;
1043                 }
1044
1045                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1046                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1047                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1048
1049                         if (unlikely(!nskb))
1050                                 goto err_skb;
1051                         if (curr_skb == head_skb)
1052                                 skb_shinfo(curr_skb)->frag_list = nskb;
1053                         else
1054                                 curr_skb->next = nskb;
1055                         curr_skb = nskb;
1056                         head_skb->truesize += nskb->truesize;
1057                         num_skb_frags = 0;
1058                 }
1059                 if (curr_skb != head_skb) {
1060                         head_skb->data_len += len;
1061                         head_skb->len += len;
1062                         head_skb->truesize += truesize;
1063                 }
1064                 offset = buf - page_address(page);
1065                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1066                         put_page(page);
1067                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1068                                              len, truesize);
1069                 } else {
1070                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1071                                         offset, len, truesize);
1072                 }
1073         }
1074
1075         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1076         return head_skb;
1077
1078 err_xdp:
1079         rcu_read_unlock();
1080         stats->xdp_drops++;
1081 err_skb:
1082         put_page(page);
1083         while (num_buf-- > 1) {
1084                 buf = virtqueue_get_buf(rq->vq, &len);
1085                 if (unlikely(!buf)) {
1086                         pr_debug("%s: rx error: %d buffers missing\n",
1087                                  dev->name, num_buf);
1088                         dev->stats.rx_length_errors++;
1089                         break;
1090                 }
1091                 stats->bytes += len;
1092                 page = virt_to_head_page(buf);
1093                 put_page(page);
1094         }
1095 err_buf:
1096         stats->drops++;
1097         dev_kfree_skb(head_skb);
1098 xdp_xmit:
1099         return NULL;
1100 }
1101
1102 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1103                         void *buf, unsigned int len, void **ctx,
1104                         unsigned int *xdp_xmit,
1105                         struct virtnet_rq_stats *stats)
1106 {
1107         struct net_device *dev = vi->dev;
1108         struct sk_buff *skb;
1109         struct virtio_net_hdr_mrg_rxbuf *hdr;
1110
1111         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1112                 pr_debug("%s: short packet %i\n", dev->name, len);
1113                 dev->stats.rx_length_errors++;
1114                 if (vi->mergeable_rx_bufs) {
1115                         put_page(virt_to_head_page(buf));
1116                 } else if (vi->big_packets) {
1117                         give_pages(rq, buf);
1118                 } else {
1119                         put_page(virt_to_head_page(buf));
1120                 }
1121                 return;
1122         }
1123
1124         if (vi->mergeable_rx_bufs)
1125                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1126                                         stats);
1127         else if (vi->big_packets)
1128                 skb = receive_big(dev, vi, rq, buf, len, stats);
1129         else
1130                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1131
1132         if (unlikely(!skb))
1133                 return;
1134
1135         hdr = skb_vnet_hdr(skb);
1136
1137         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1138                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1139
1140         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1141                                   virtio_is_little_endian(vi->vdev))) {
1142                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1143                                      dev->name, hdr->hdr.gso_type,
1144                                      hdr->hdr.gso_size);
1145                 goto frame_err;
1146         }
1147
1148         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1149         skb->protocol = eth_type_trans(skb, dev);
1150         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1151                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1152
1153         napi_gro_receive(&rq->napi, skb);
1154         return;
1155
1156 frame_err:
1157         dev->stats.rx_frame_errors++;
1158         dev_kfree_skb(skb);
1159 }
1160
1161 /* Unlike mergeable buffers, all buffers are allocated to the
1162  * same size, except for the headroom. For this reason we do
1163  * not need to use  mergeable_len_to_ctx here - it is enough
1164  * to store the headroom as the context ignoring the truesize.
1165  */
1166 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1167                              gfp_t gfp)
1168 {
1169         struct page_frag *alloc_frag = &rq->alloc_frag;
1170         char *buf;
1171         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1172         void *ctx = (void *)(unsigned long)xdp_headroom;
1173         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1174         int err;
1175
1176         len = SKB_DATA_ALIGN(len) +
1177               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1178         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1179                 return -ENOMEM;
1180
1181         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1182         get_page(alloc_frag->page);
1183         alloc_frag->offset += len;
1184         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1185                     vi->hdr_len + GOOD_PACKET_LEN);
1186         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1187         if (err < 0)
1188                 put_page(virt_to_head_page(buf));
1189         return err;
1190 }
1191
1192 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1193                            gfp_t gfp)
1194 {
1195         struct page *first, *list = NULL;
1196         char *p;
1197         int i, err, offset;
1198
1199         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1200
1201         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1202         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1203                 first = get_a_page(rq, gfp);
1204                 if (!first) {
1205                         if (list)
1206                                 give_pages(rq, list);
1207                         return -ENOMEM;
1208                 }
1209                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1210
1211                 /* chain new page in list head to match sg */
1212                 first->private = (unsigned long)list;
1213                 list = first;
1214         }
1215
1216         first = get_a_page(rq, gfp);
1217         if (!first) {
1218                 give_pages(rq, list);
1219                 return -ENOMEM;
1220         }
1221         p = page_address(first);
1222
1223         /* rq->sg[0], rq->sg[1] share the same page */
1224         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1225         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1226
1227         /* rq->sg[1] for data packet, from offset */
1228         offset = sizeof(struct padded_vnet_hdr);
1229         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1230
1231         /* chain first in list head */
1232         first->private = (unsigned long)list;
1233         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1234                                   first, gfp);
1235         if (err < 0)
1236                 give_pages(rq, first);
1237
1238         return err;
1239 }
1240
1241 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1242                                           struct ewma_pkt_len *avg_pkt_len,
1243                                           unsigned int room)
1244 {
1245         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1246         unsigned int len;
1247
1248         if (room)
1249                 return PAGE_SIZE - room;
1250
1251         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1252                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1253
1254         return ALIGN(len, L1_CACHE_BYTES);
1255 }
1256
1257 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1258                                  struct receive_queue *rq, gfp_t gfp)
1259 {
1260         struct page_frag *alloc_frag = &rq->alloc_frag;
1261         unsigned int headroom = virtnet_get_headroom(vi);
1262         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1263         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1264         char *buf;
1265         void *ctx;
1266         int err;
1267         unsigned int len, hole;
1268
1269         /* Extra tailroom is needed to satisfy XDP's assumption. This
1270          * means rx frags coalescing won't work, but consider we've
1271          * disabled GSO for XDP, it won't be a big issue.
1272          */
1273         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1274         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1275                 return -ENOMEM;
1276
1277         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1278         buf += headroom; /* advance address leaving hole at front of pkt */
1279         get_page(alloc_frag->page);
1280         alloc_frag->offset += len + room;
1281         hole = alloc_frag->size - alloc_frag->offset;
1282         if (hole < len + room) {
1283                 /* To avoid internal fragmentation, if there is very likely not
1284                  * enough space for another buffer, add the remaining space to
1285                  * the current buffer.
1286                  */
1287                 len += hole;
1288                 alloc_frag->offset += hole;
1289         }
1290
1291         sg_init_one(rq->sg, buf, len);
1292         ctx = mergeable_len_to_ctx(len, headroom);
1293         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1294         if (err < 0)
1295                 put_page(virt_to_head_page(buf));
1296
1297         return err;
1298 }
1299
1300 /*
1301  * Returns false if we couldn't fill entirely (OOM).
1302  *
1303  * Normally run in the receive path, but can also be run from ndo_open
1304  * before we're receiving packets, or from refill_work which is
1305  * careful to disable receiving (using napi_disable).
1306  */
1307 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1308                           gfp_t gfp)
1309 {
1310         int err;
1311         bool oom;
1312
1313         do {
1314                 if (vi->mergeable_rx_bufs)
1315                         err = add_recvbuf_mergeable(vi, rq, gfp);
1316                 else if (vi->big_packets)
1317                         err = add_recvbuf_big(vi, rq, gfp);
1318                 else
1319                         err = add_recvbuf_small(vi, rq, gfp);
1320
1321                 oom = err == -ENOMEM;
1322                 if (err)
1323                         break;
1324         } while (rq->vq->num_free);
1325         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1326                 unsigned long flags;
1327
1328                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1329                 rq->stats.kicks++;
1330                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1331         }
1332
1333         return !oom;
1334 }
1335
1336 static void skb_recv_done(struct virtqueue *rvq)
1337 {
1338         struct virtnet_info *vi = rvq->vdev->priv;
1339         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1340
1341         virtqueue_napi_schedule(&rq->napi, rvq);
1342 }
1343
1344 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1345 {
1346         napi_enable(napi);
1347
1348         /* If all buffers were filled by other side before we napi_enabled, we
1349          * won't get another interrupt, so process any outstanding packets now.
1350          * Call local_bh_enable after to trigger softIRQ processing.
1351          */
1352         local_bh_disable();
1353         virtqueue_napi_schedule(napi, vq);
1354         local_bh_enable();
1355 }
1356
1357 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1358                                    struct virtqueue *vq,
1359                                    struct napi_struct *napi)
1360 {
1361         if (!napi->weight)
1362                 return;
1363
1364         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1365          * enable the feature if this is likely affine with the transmit path.
1366          */
1367         if (!vi->affinity_hint_set) {
1368                 napi->weight = 0;
1369                 return;
1370         }
1371
1372         return virtnet_napi_enable(vq, napi);
1373 }
1374
1375 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1376 {
1377         if (napi->weight)
1378                 napi_disable(napi);
1379 }
1380
1381 static void refill_work(struct work_struct *work)
1382 {
1383         struct virtnet_info *vi =
1384                 container_of(work, struct virtnet_info, refill.work);
1385         bool still_empty;
1386         int i;
1387
1388         for (i = 0; i < vi->curr_queue_pairs; i++) {
1389                 struct receive_queue *rq = &vi->rq[i];
1390
1391                 napi_disable(&rq->napi);
1392                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1393                 virtnet_napi_enable(rq->vq, &rq->napi);
1394
1395                 /* In theory, this can happen: if we don't get any buffers in
1396                  * we will *never* try to fill again.
1397                  */
1398                 if (still_empty)
1399                         schedule_delayed_work(&vi->refill, HZ/2);
1400         }
1401 }
1402
1403 static int virtnet_receive(struct receive_queue *rq, int budget,
1404                            unsigned int *xdp_xmit)
1405 {
1406         struct virtnet_info *vi = rq->vq->vdev->priv;
1407         struct virtnet_rq_stats stats = {};
1408         unsigned int len;
1409         void *buf;
1410         int i;
1411
1412         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1413                 void *ctx;
1414
1415                 while (stats.packets < budget &&
1416                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1417                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1418                         stats.packets++;
1419                 }
1420         } else {
1421                 while (stats.packets < budget &&
1422                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1423                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1424                         stats.packets++;
1425                 }
1426         }
1427
1428         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1429                 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
1430                         spin_lock(&vi->refill_lock);
1431                         if (vi->refill_enabled)
1432                                 schedule_delayed_work(&vi->refill, 0);
1433                         spin_unlock(&vi->refill_lock);
1434                 }
1435         }
1436
1437         u64_stats_update_begin(&rq->stats.syncp);
1438         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1439                 size_t offset = virtnet_rq_stats_desc[i].offset;
1440                 u64 *item;
1441
1442                 item = (u64 *)((u8 *)&rq->stats + offset);
1443                 *item += *(u64 *)((u8 *)&stats + offset);
1444         }
1445         u64_stats_update_end(&rq->stats.syncp);
1446
1447         return stats.packets;
1448 }
1449
1450 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1451 {
1452         unsigned int len;
1453         unsigned int packets = 0;
1454         unsigned int bytes = 0;
1455         void *ptr;
1456
1457         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1458                 if (likely(!is_xdp_frame(ptr))) {
1459                         struct sk_buff *skb = ptr;
1460
1461                         pr_debug("Sent skb %p\n", skb);
1462
1463                         bytes += skb->len;
1464                         napi_consume_skb(skb, in_napi);
1465                 } else {
1466                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1467
1468                         bytes += frame->len;
1469                         xdp_return_frame(frame);
1470                 }
1471                 packets++;
1472         }
1473
1474         /* Avoid overhead when no packets have been processed
1475          * happens when called speculatively from start_xmit.
1476          */
1477         if (!packets)
1478                 return;
1479
1480         u64_stats_update_begin(&sq->stats.syncp);
1481         sq->stats.bytes += bytes;
1482         sq->stats.packets += packets;
1483         u64_stats_update_end(&sq->stats.syncp);
1484 }
1485
1486 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1487 {
1488         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1489                 return false;
1490         else if (q < vi->curr_queue_pairs)
1491                 return true;
1492         else
1493                 return false;
1494 }
1495
1496 static void virtnet_poll_cleantx(struct receive_queue *rq)
1497 {
1498         struct virtnet_info *vi = rq->vq->vdev->priv;
1499         unsigned int index = vq2rxq(rq->vq);
1500         struct send_queue *sq = &vi->sq[index];
1501         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1502
1503         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1504                 return;
1505
1506         if (__netif_tx_trylock(txq)) {
1507                 free_old_xmit_skbs(sq, true);
1508                 __netif_tx_unlock(txq);
1509         }
1510
1511         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1512                 netif_tx_wake_queue(txq);
1513 }
1514
1515 static int virtnet_poll(struct napi_struct *napi, int budget)
1516 {
1517         struct receive_queue *rq =
1518                 container_of(napi, struct receive_queue, napi);
1519         struct virtnet_info *vi = rq->vq->vdev->priv;
1520         struct send_queue *sq;
1521         unsigned int received;
1522         unsigned int xdp_xmit = 0;
1523
1524         virtnet_poll_cleantx(rq);
1525
1526         received = virtnet_receive(rq, budget, &xdp_xmit);
1527
1528         /* Out of packets? */
1529         if (received < budget)
1530                 virtqueue_napi_complete(napi, rq->vq, received);
1531
1532         if (xdp_xmit & VIRTIO_XDP_REDIR)
1533                 xdp_do_flush();
1534
1535         if (xdp_xmit & VIRTIO_XDP_TX) {
1536                 sq = virtnet_xdp_get_sq(vi);
1537                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1538                         u64_stats_update_begin(&sq->stats.syncp);
1539                         sq->stats.kicks++;
1540                         u64_stats_update_end(&sq->stats.syncp);
1541                 }
1542                 virtnet_xdp_put_sq(vi, sq);
1543         }
1544
1545         return received;
1546 }
1547
1548 static int virtnet_open(struct net_device *dev)
1549 {
1550         struct virtnet_info *vi = netdev_priv(dev);
1551         int i, err;
1552
1553         enable_delayed_refill(vi);
1554
1555         for (i = 0; i < vi->max_queue_pairs; i++) {
1556                 if (i < vi->curr_queue_pairs)
1557                         /* Make sure we have some buffers: if oom use wq. */
1558                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1559                                 schedule_delayed_work(&vi->refill, 0);
1560
1561                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1562                 if (err < 0)
1563                         return err;
1564
1565                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1566                                                  MEM_TYPE_PAGE_SHARED, NULL);
1567                 if (err < 0) {
1568                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1569                         return err;
1570                 }
1571
1572                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1573                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1574         }
1575
1576         return 0;
1577 }
1578
1579 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1580 {
1581         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1582         struct virtnet_info *vi = sq->vq->vdev->priv;
1583         unsigned int index = vq2txq(sq->vq);
1584         struct netdev_queue *txq;
1585         int opaque;
1586         bool done;
1587
1588         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1589                 /* We don't need to enable cb for XDP */
1590                 napi_complete_done(napi, 0);
1591                 return 0;
1592         }
1593
1594         txq = netdev_get_tx_queue(vi->dev, index);
1595         __netif_tx_lock(txq, raw_smp_processor_id());
1596         virtqueue_disable_cb(sq->vq);
1597         free_old_xmit_skbs(sq, true);
1598
1599         opaque = virtqueue_enable_cb_prepare(sq->vq);
1600
1601         done = napi_complete_done(napi, 0);
1602
1603         if (!done)
1604                 virtqueue_disable_cb(sq->vq);
1605
1606         __netif_tx_unlock(txq);
1607
1608         if (done) {
1609                 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1610                         if (napi_schedule_prep(napi)) {
1611                                 __netif_tx_lock(txq, raw_smp_processor_id());
1612                                 virtqueue_disable_cb(sq->vq);
1613                                 __netif_tx_unlock(txq);
1614                                 __napi_schedule(napi);
1615                         }
1616                 }
1617         }
1618
1619         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1620                 netif_tx_wake_queue(txq);
1621
1622         return 0;
1623 }
1624
1625 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1626 {
1627         struct virtio_net_hdr_mrg_rxbuf *hdr;
1628         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1629         struct virtnet_info *vi = sq->vq->vdev->priv;
1630         int num_sg;
1631         unsigned hdr_len = vi->hdr_len;
1632         bool can_push;
1633
1634         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1635
1636         can_push = vi->any_header_sg &&
1637                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1638                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1639         /* Even if we can, don't push here yet as this would skew
1640          * csum_start offset below. */
1641         if (can_push)
1642                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1643         else
1644                 hdr = skb_vnet_hdr(skb);
1645
1646         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1647                                     virtio_is_little_endian(vi->vdev), false,
1648                                     0))
1649                 return -EPROTO;
1650
1651         if (vi->mergeable_rx_bufs)
1652                 hdr->num_buffers = 0;
1653
1654         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1655         if (can_push) {
1656                 __skb_push(skb, hdr_len);
1657                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1658                 if (unlikely(num_sg < 0))
1659                         return num_sg;
1660                 /* Pull header back to avoid skew in tx bytes calculations. */
1661                 __skb_pull(skb, hdr_len);
1662         } else {
1663                 sg_set_buf(sq->sg, hdr, hdr_len);
1664                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1665                 if (unlikely(num_sg < 0))
1666                         return num_sg;
1667                 num_sg++;
1668         }
1669         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1670 }
1671
1672 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1673 {
1674         struct virtnet_info *vi = netdev_priv(dev);
1675         int qnum = skb_get_queue_mapping(skb);
1676         struct send_queue *sq = &vi->sq[qnum];
1677         int err;
1678         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1679         bool kick = !netdev_xmit_more();
1680         bool use_napi = sq->napi.weight;
1681
1682         /* Free up any pending old buffers before queueing new ones. */
1683         free_old_xmit_skbs(sq, false);
1684
1685         if (use_napi && kick)
1686                 virtqueue_enable_cb_delayed(sq->vq);
1687
1688         /* timestamp packet in software */
1689         skb_tx_timestamp(skb);
1690
1691         /* Try to transmit */
1692         err = xmit_skb(sq, skb);
1693
1694         /* This should not happen! */
1695         if (unlikely(err)) {
1696                 dev->stats.tx_fifo_errors++;
1697                 if (net_ratelimit())
1698                         dev_warn(&dev->dev,
1699                                  "Unexpected TXQ (%d) queue failure: %d\n",
1700                                  qnum, err);
1701                 dev->stats.tx_dropped++;
1702                 dev_kfree_skb_any(skb);
1703                 return NETDEV_TX_OK;
1704         }
1705
1706         /* Don't wait up for transmitted skbs to be freed. */
1707         if (!use_napi) {
1708                 skb_orphan(skb);
1709                 nf_reset_ct(skb);
1710         }
1711
1712         /* If running out of space, stop queue to avoid getting packets that we
1713          * are then unable to transmit.
1714          * An alternative would be to force queuing layer to requeue the skb by
1715          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1716          * returned in a normal path of operation: it means that driver is not
1717          * maintaining the TX queue stop/start state properly, and causes
1718          * the stack to do a non-trivial amount of useless work.
1719          * Since most packets only take 1 or 2 ring slots, stopping the queue
1720          * early means 16 slots are typically wasted.
1721          */
1722         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1723                 netif_stop_subqueue(dev, qnum);
1724                 if (!use_napi &&
1725                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1726                         /* More just got used, free them then recheck. */
1727                         free_old_xmit_skbs(sq, false);
1728                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1729                                 netif_start_subqueue(dev, qnum);
1730                                 virtqueue_disable_cb(sq->vq);
1731                         }
1732                 }
1733         }
1734
1735         if (kick || netif_xmit_stopped(txq)) {
1736                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1737                         u64_stats_update_begin(&sq->stats.syncp);
1738                         sq->stats.kicks++;
1739                         u64_stats_update_end(&sq->stats.syncp);
1740                 }
1741         }
1742
1743         return NETDEV_TX_OK;
1744 }
1745
1746 /*
1747  * Send command via the control virtqueue and check status.  Commands
1748  * supported by the hypervisor, as indicated by feature bits, should
1749  * never fail unless improperly formatted.
1750  */
1751 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1752                                  struct scatterlist *out)
1753 {
1754         struct scatterlist *sgs[4], hdr, stat;
1755         unsigned out_num = 0, tmp;
1756
1757         /* Caller should know better */
1758         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1759
1760         vi->ctrl->status = ~0;
1761         vi->ctrl->hdr.class = class;
1762         vi->ctrl->hdr.cmd = cmd;
1763         /* Add header */
1764         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1765         sgs[out_num++] = &hdr;
1766
1767         if (out)
1768                 sgs[out_num++] = out;
1769
1770         /* Add return status. */
1771         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1772         sgs[out_num] = &stat;
1773
1774         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1775         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1776
1777         if (unlikely(!virtqueue_kick(vi->cvq)))
1778                 return vi->ctrl->status == VIRTIO_NET_OK;
1779
1780         /* Spin for a response, the kick causes an ioport write, trapping
1781          * into the hypervisor, so the request should be handled immediately.
1782          */
1783         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1784                !virtqueue_is_broken(vi->cvq))
1785                 cpu_relax();
1786
1787         return vi->ctrl->status == VIRTIO_NET_OK;
1788 }
1789
1790 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1791 {
1792         struct virtnet_info *vi = netdev_priv(dev);
1793         struct virtio_device *vdev = vi->vdev;
1794         int ret;
1795         struct sockaddr *addr;
1796         struct scatterlist sg;
1797
1798         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1799                 return -EOPNOTSUPP;
1800
1801         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1802         if (!addr)
1803                 return -ENOMEM;
1804
1805         ret = eth_prepare_mac_addr_change(dev, addr);
1806         if (ret)
1807                 goto out;
1808
1809         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1810                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1811                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1812                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1813                         dev_warn(&vdev->dev,
1814                                  "Failed to set mac address by vq command.\n");
1815                         ret = -EINVAL;
1816                         goto out;
1817                 }
1818         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1819                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1820                 unsigned int i;
1821
1822                 /* Naturally, this has an atomicity problem. */
1823                 for (i = 0; i < dev->addr_len; i++)
1824                         virtio_cwrite8(vdev,
1825                                        offsetof(struct virtio_net_config, mac) +
1826                                        i, addr->sa_data[i]);
1827         }
1828
1829         eth_commit_mac_addr_change(dev, p);
1830         ret = 0;
1831
1832 out:
1833         kfree(addr);
1834         return ret;
1835 }
1836
1837 static void virtnet_stats(struct net_device *dev,
1838                           struct rtnl_link_stats64 *tot)
1839 {
1840         struct virtnet_info *vi = netdev_priv(dev);
1841         unsigned int start;
1842         int i;
1843
1844         for (i = 0; i < vi->max_queue_pairs; i++) {
1845                 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1846                 struct receive_queue *rq = &vi->rq[i];
1847                 struct send_queue *sq = &vi->sq[i];
1848
1849                 do {
1850                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1851                         tpackets = sq->stats.packets;
1852                         tbytes   = sq->stats.bytes;
1853                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1854
1855                 do {
1856                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1857                         rpackets = rq->stats.packets;
1858                         rbytes   = rq->stats.bytes;
1859                         rdrops   = rq->stats.drops;
1860                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1861
1862                 tot->rx_packets += rpackets;
1863                 tot->tx_packets += tpackets;
1864                 tot->rx_bytes   += rbytes;
1865                 tot->tx_bytes   += tbytes;
1866                 tot->rx_dropped += rdrops;
1867         }
1868
1869         tot->tx_dropped = dev->stats.tx_dropped;
1870         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1871         tot->rx_length_errors = dev->stats.rx_length_errors;
1872         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1873 }
1874
1875 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1876 {
1877         rtnl_lock();
1878         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1879                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1880                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1881         rtnl_unlock();
1882 }
1883
1884 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1885 {
1886         struct scatterlist sg;
1887         struct net_device *dev = vi->dev;
1888
1889         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1890                 return 0;
1891
1892         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1893         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1894
1895         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1896                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1897                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1898                          queue_pairs);
1899                 return -EINVAL;
1900         } else {
1901                 vi->curr_queue_pairs = queue_pairs;
1902                 /* virtnet_open() will refill when device is going to up. */
1903                 if (dev->flags & IFF_UP)
1904                         schedule_delayed_work(&vi->refill, 0);
1905         }
1906
1907         return 0;
1908 }
1909
1910 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1911 {
1912         int err;
1913
1914         rtnl_lock();
1915         err = _virtnet_set_queues(vi, queue_pairs);
1916         rtnl_unlock();
1917         return err;
1918 }
1919
1920 static int virtnet_close(struct net_device *dev)
1921 {
1922         struct virtnet_info *vi = netdev_priv(dev);
1923         int i;
1924
1925         /* Make sure NAPI doesn't schedule refill work */
1926         disable_delayed_refill(vi);
1927         /* Make sure refill_work doesn't re-enable napi! */
1928         cancel_delayed_work_sync(&vi->refill);
1929
1930         for (i = 0; i < vi->max_queue_pairs; i++) {
1931                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1932                 napi_disable(&vi->rq[i].napi);
1933                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1934         }
1935
1936         return 0;
1937 }
1938
1939 static void virtnet_set_rx_mode(struct net_device *dev)
1940 {
1941         struct virtnet_info *vi = netdev_priv(dev);
1942         struct scatterlist sg[2];
1943         struct virtio_net_ctrl_mac *mac_data;
1944         struct netdev_hw_addr *ha;
1945         int uc_count;
1946         int mc_count;
1947         void *buf;
1948         int i;
1949
1950         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1951         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1952                 return;
1953
1954         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1955         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1956
1957         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1958
1959         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1960                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1961                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1962                          vi->ctrl->promisc ? "en" : "dis");
1963
1964         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1965
1966         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1967                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1968                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1969                          vi->ctrl->allmulti ? "en" : "dis");
1970
1971         uc_count = netdev_uc_count(dev);
1972         mc_count = netdev_mc_count(dev);
1973         /* MAC filter - use one buffer for both lists */
1974         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1975                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1976         mac_data = buf;
1977         if (!buf)
1978                 return;
1979
1980         sg_init_table(sg, 2);
1981
1982         /* Store the unicast list and count in the front of the buffer */
1983         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1984         i = 0;
1985         netdev_for_each_uc_addr(ha, dev)
1986                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1987
1988         sg_set_buf(&sg[0], mac_data,
1989                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1990
1991         /* multicast list and count fill the end */
1992         mac_data = (void *)&mac_data->macs[uc_count][0];
1993
1994         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1995         i = 0;
1996         netdev_for_each_mc_addr(ha, dev)
1997                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1998
1999         sg_set_buf(&sg[1], mac_data,
2000                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2001
2002         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2003                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2004                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2005
2006         kfree(buf);
2007 }
2008
2009 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2010                                    __be16 proto, u16 vid)
2011 {
2012         struct virtnet_info *vi = netdev_priv(dev);
2013         struct scatterlist sg;
2014
2015         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2016         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2017
2018         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2019                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2020                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2021         return 0;
2022 }
2023
2024 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2025                                     __be16 proto, u16 vid)
2026 {
2027         struct virtnet_info *vi = netdev_priv(dev);
2028         struct scatterlist sg;
2029
2030         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2031         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2032
2033         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2034                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2035                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2036         return 0;
2037 }
2038
2039 static void virtnet_clean_affinity(struct virtnet_info *vi)
2040 {
2041         int i;
2042
2043         if (vi->affinity_hint_set) {
2044                 for (i = 0; i < vi->max_queue_pairs; i++) {
2045                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
2046                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
2047                 }
2048
2049                 vi->affinity_hint_set = false;
2050         }
2051 }
2052
2053 static void virtnet_set_affinity(struct virtnet_info *vi)
2054 {
2055         cpumask_var_t mask;
2056         int stragglers;
2057         int group_size;
2058         int i, j, cpu;
2059         int num_cpu;
2060         int stride;
2061
2062         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2063                 virtnet_clean_affinity(vi);
2064                 return;
2065         }
2066
2067         num_cpu = num_online_cpus();
2068         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2069         stragglers = num_cpu >= vi->curr_queue_pairs ?
2070                         num_cpu % vi->curr_queue_pairs :
2071                         0;
2072         cpu = cpumask_next(-1, cpu_online_mask);
2073
2074         for (i = 0; i < vi->curr_queue_pairs; i++) {
2075                 group_size = stride + (i < stragglers ? 1 : 0);
2076
2077                 for (j = 0; j < group_size; j++) {
2078                         cpumask_set_cpu(cpu, mask);
2079                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2080                                                 nr_cpu_ids, false);
2081                 }
2082                 virtqueue_set_affinity(vi->rq[i].vq, mask);
2083                 virtqueue_set_affinity(vi->sq[i].vq, mask);
2084                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, false);
2085                 cpumask_clear(mask);
2086         }
2087
2088         vi->affinity_hint_set = true;
2089         free_cpumask_var(mask);
2090 }
2091
2092 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2093 {
2094         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2095                                                    node);
2096         virtnet_set_affinity(vi);
2097         return 0;
2098 }
2099
2100 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2101 {
2102         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2103                                                    node_dead);
2104         virtnet_set_affinity(vi);
2105         return 0;
2106 }
2107
2108 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2109 {
2110         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2111                                                    node);
2112
2113         virtnet_clean_affinity(vi);
2114         return 0;
2115 }
2116
2117 static enum cpuhp_state virtionet_online;
2118
2119 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2120 {
2121         int ret;
2122
2123         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2124         if (ret)
2125                 return ret;
2126         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2127                                                &vi->node_dead);
2128         if (!ret)
2129                 return ret;
2130         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2131         return ret;
2132 }
2133
2134 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2135 {
2136         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2137         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2138                                             &vi->node_dead);
2139 }
2140
2141 static void virtnet_get_ringparam(struct net_device *dev,
2142                                 struct ethtool_ringparam *ring)
2143 {
2144         struct virtnet_info *vi = netdev_priv(dev);
2145
2146         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2147         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2148         ring->rx_pending = ring->rx_max_pending;
2149         ring->tx_pending = ring->tx_max_pending;
2150 }
2151
2152
2153 static void virtnet_get_drvinfo(struct net_device *dev,
2154                                 struct ethtool_drvinfo *info)
2155 {
2156         struct virtnet_info *vi = netdev_priv(dev);
2157         struct virtio_device *vdev = vi->vdev;
2158
2159         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2160         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2161         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2162
2163 }
2164
2165 /* TODO: Eliminate OOO packets during switching */
2166 static int virtnet_set_channels(struct net_device *dev,
2167                                 struct ethtool_channels *channels)
2168 {
2169         struct virtnet_info *vi = netdev_priv(dev);
2170         u16 queue_pairs = channels->combined_count;
2171         int err;
2172
2173         /* We don't support separate rx/tx channels.
2174          * We don't allow setting 'other' channels.
2175          */
2176         if (channels->rx_count || channels->tx_count || channels->other_count)
2177                 return -EINVAL;
2178
2179         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2180                 return -EINVAL;
2181
2182         /* For now we don't support modifying channels while XDP is loaded
2183          * also when XDP is loaded all RX queues have XDP programs so we only
2184          * need to check a single RX queue.
2185          */
2186         if (vi->rq[0].xdp_prog)
2187                 return -EINVAL;
2188
2189         get_online_cpus();
2190         err = _virtnet_set_queues(vi, queue_pairs);
2191         if (err) {
2192                 put_online_cpus();
2193                 goto err;
2194         }
2195         virtnet_set_affinity(vi);
2196         put_online_cpus();
2197
2198         netif_set_real_num_tx_queues(dev, queue_pairs);
2199         netif_set_real_num_rx_queues(dev, queue_pairs);
2200  err:
2201         return err;
2202 }
2203
2204 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2205 {
2206         struct virtnet_info *vi = netdev_priv(dev);
2207         char *p = (char *)data;
2208         unsigned int i, j;
2209
2210         switch (stringset) {
2211         case ETH_SS_STATS:
2212                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2213                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2214                                 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
2215                                          i, virtnet_rq_stats_desc[j].desc);
2216                                 p += ETH_GSTRING_LEN;
2217                         }
2218                 }
2219
2220                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2221                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2222                                 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
2223                                          i, virtnet_sq_stats_desc[j].desc);
2224                                 p += ETH_GSTRING_LEN;
2225                         }
2226                 }
2227                 break;
2228         }
2229 }
2230
2231 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2232 {
2233         struct virtnet_info *vi = netdev_priv(dev);
2234
2235         switch (sset) {
2236         case ETH_SS_STATS:
2237                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2238                                                VIRTNET_SQ_STATS_LEN);
2239         default:
2240                 return -EOPNOTSUPP;
2241         }
2242 }
2243
2244 static void virtnet_get_ethtool_stats(struct net_device *dev,
2245                                       struct ethtool_stats *stats, u64 *data)
2246 {
2247         struct virtnet_info *vi = netdev_priv(dev);
2248         unsigned int idx = 0, start, i, j;
2249         const u8 *stats_base;
2250         size_t offset;
2251
2252         for (i = 0; i < vi->curr_queue_pairs; i++) {
2253                 struct receive_queue *rq = &vi->rq[i];
2254
2255                 stats_base = (u8 *)&rq->stats;
2256                 do {
2257                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2258                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2259                                 offset = virtnet_rq_stats_desc[j].offset;
2260                                 data[idx + j] = *(u64 *)(stats_base + offset);
2261                         }
2262                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2263                 idx += VIRTNET_RQ_STATS_LEN;
2264         }
2265
2266         for (i = 0; i < vi->curr_queue_pairs; i++) {
2267                 struct send_queue *sq = &vi->sq[i];
2268
2269                 stats_base = (u8 *)&sq->stats;
2270                 do {
2271                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2272                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2273                                 offset = virtnet_sq_stats_desc[j].offset;
2274                                 data[idx + j] = *(u64 *)(stats_base + offset);
2275                         }
2276                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2277                 idx += VIRTNET_SQ_STATS_LEN;
2278         }
2279 }
2280
2281 static void virtnet_get_channels(struct net_device *dev,
2282                                  struct ethtool_channels *channels)
2283 {
2284         struct virtnet_info *vi = netdev_priv(dev);
2285
2286         channels->combined_count = vi->curr_queue_pairs;
2287         channels->max_combined = vi->max_queue_pairs;
2288         channels->max_other = 0;
2289         channels->rx_count = 0;
2290         channels->tx_count = 0;
2291         channels->other_count = 0;
2292 }
2293
2294 static int virtnet_set_link_ksettings(struct net_device *dev,
2295                                       const struct ethtool_link_ksettings *cmd)
2296 {
2297         struct virtnet_info *vi = netdev_priv(dev);
2298
2299         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2300                                                   &vi->speed, &vi->duplex);
2301 }
2302
2303 static int virtnet_get_link_ksettings(struct net_device *dev,
2304                                       struct ethtool_link_ksettings *cmd)
2305 {
2306         struct virtnet_info *vi = netdev_priv(dev);
2307
2308         cmd->base.speed = vi->speed;
2309         cmd->base.duplex = vi->duplex;
2310         cmd->base.port = PORT_OTHER;
2311
2312         return 0;
2313 }
2314
2315 static int virtnet_set_coalesce(struct net_device *dev,
2316                                 struct ethtool_coalesce *ec)
2317 {
2318         struct virtnet_info *vi = netdev_priv(dev);
2319         int i, napi_weight;
2320
2321         if (ec->tx_max_coalesced_frames > 1 ||
2322             ec->rx_max_coalesced_frames != 1)
2323                 return -EINVAL;
2324
2325         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2326         if (napi_weight ^ vi->sq[0].napi.weight) {
2327                 if (dev->flags & IFF_UP)
2328                         return -EBUSY;
2329                 for (i = 0; i < vi->max_queue_pairs; i++)
2330                         vi->sq[i].napi.weight = napi_weight;
2331         }
2332
2333         return 0;
2334 }
2335
2336 static int virtnet_get_coalesce(struct net_device *dev,
2337                                 struct ethtool_coalesce *ec)
2338 {
2339         struct ethtool_coalesce ec_default = {
2340                 .cmd = ETHTOOL_GCOALESCE,
2341                 .rx_max_coalesced_frames = 1,
2342         };
2343         struct virtnet_info *vi = netdev_priv(dev);
2344
2345         memcpy(ec, &ec_default, sizeof(ec_default));
2346
2347         if (vi->sq[0].napi.weight)
2348                 ec->tx_max_coalesced_frames = 1;
2349
2350         return 0;
2351 }
2352
2353 static void virtnet_init_settings(struct net_device *dev)
2354 {
2355         struct virtnet_info *vi = netdev_priv(dev);
2356
2357         vi->speed = SPEED_UNKNOWN;
2358         vi->duplex = DUPLEX_UNKNOWN;
2359 }
2360
2361 static void virtnet_update_settings(struct virtnet_info *vi)
2362 {
2363         u32 speed;
2364         u8 duplex;
2365
2366         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2367                 return;
2368
2369         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2370
2371         if (ethtool_validate_speed(speed))
2372                 vi->speed = speed;
2373
2374         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2375
2376         if (ethtool_validate_duplex(duplex))
2377                 vi->duplex = duplex;
2378 }
2379
2380 static const struct ethtool_ops virtnet_ethtool_ops = {
2381         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2382         .get_drvinfo = virtnet_get_drvinfo,
2383         .get_link = ethtool_op_get_link,
2384         .get_ringparam = virtnet_get_ringparam,
2385         .get_strings = virtnet_get_strings,
2386         .get_sset_count = virtnet_get_sset_count,
2387         .get_ethtool_stats = virtnet_get_ethtool_stats,
2388         .set_channels = virtnet_set_channels,
2389         .get_channels = virtnet_get_channels,
2390         .get_ts_info = ethtool_op_get_ts_info,
2391         .get_link_ksettings = virtnet_get_link_ksettings,
2392         .set_link_ksettings = virtnet_set_link_ksettings,
2393         .set_coalesce = virtnet_set_coalesce,
2394         .get_coalesce = virtnet_get_coalesce,
2395 };
2396
2397 static void virtnet_freeze_down(struct virtio_device *vdev)
2398 {
2399         struct virtnet_info *vi = vdev->priv;
2400
2401         /* Make sure no work handler is accessing the device */
2402         flush_work(&vi->config_work);
2403
2404         netif_tx_lock_bh(vi->dev);
2405         netif_device_detach(vi->dev);
2406         netif_tx_unlock_bh(vi->dev);
2407         if (netif_running(vi->dev))
2408                 virtnet_close(vi->dev);
2409 }
2410
2411 static int init_vqs(struct virtnet_info *vi);
2412
2413 static int virtnet_restore_up(struct virtio_device *vdev)
2414 {
2415         struct virtnet_info *vi = vdev->priv;
2416         int err;
2417
2418         err = init_vqs(vi);
2419         if (err)
2420                 return err;
2421
2422         virtio_device_ready(vdev);
2423
2424         enable_delayed_refill(vi);
2425
2426         if (netif_running(vi->dev)) {
2427                 err = virtnet_open(vi->dev);
2428                 if (err)
2429                         return err;
2430         }
2431
2432         netif_tx_lock_bh(vi->dev);
2433         netif_device_attach(vi->dev);
2434         netif_tx_unlock_bh(vi->dev);
2435         return err;
2436 }
2437
2438 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2439 {
2440         struct scatterlist sg;
2441         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2442
2443         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2444
2445         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2446                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2447                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2448                 return -EINVAL;
2449         }
2450
2451         return 0;
2452 }
2453
2454 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2455 {
2456         u64 offloads = 0;
2457
2458         if (!vi->guest_offloads)
2459                 return 0;
2460
2461         return virtnet_set_guest_offloads(vi, offloads);
2462 }
2463
2464 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2465 {
2466         u64 offloads = vi->guest_offloads;
2467
2468         if (!vi->guest_offloads)
2469                 return 0;
2470
2471         return virtnet_set_guest_offloads(vi, offloads);
2472 }
2473
2474 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2475                            struct netlink_ext_ack *extack)
2476 {
2477         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2478         struct virtnet_info *vi = netdev_priv(dev);
2479         struct bpf_prog *old_prog;
2480         u16 xdp_qp = 0, curr_qp;
2481         int i, err;
2482
2483         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2484             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2485                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2486                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2487                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2488                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2489                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
2490                 return -EOPNOTSUPP;
2491         }
2492
2493         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2494                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2495                 return -EINVAL;
2496         }
2497
2498         if (dev->mtu > max_sz) {
2499                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2500                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2501                 return -EINVAL;
2502         }
2503
2504         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2505         if (prog)
2506                 xdp_qp = nr_cpu_ids;
2507
2508         /* XDP requires extra queues for XDP_TX */
2509         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2510                 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",
2511                             curr_qp + xdp_qp, vi->max_queue_pairs);
2512                 xdp_qp = 0;
2513         }
2514
2515         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2516         if (!prog && !old_prog)
2517                 return 0;
2518
2519         if (prog)
2520                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2521
2522         /* Make sure NAPI is not using any XDP TX queues for RX. */
2523         if (netif_running(dev)) {
2524                 for (i = 0; i < vi->max_queue_pairs; i++) {
2525                         napi_disable(&vi->rq[i].napi);
2526                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2527                 }
2528         }
2529
2530         if (!prog) {
2531                 for (i = 0; i < vi->max_queue_pairs; i++) {
2532                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2533                         if (i == 0)
2534                                 virtnet_restore_guest_offloads(vi);
2535                 }
2536                 synchronize_net();
2537         }
2538
2539         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2540         if (err)
2541                 goto err;
2542         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2543         vi->xdp_queue_pairs = xdp_qp;
2544
2545         if (prog) {
2546                 vi->xdp_enabled = true;
2547                 for (i = 0; i < vi->max_queue_pairs; i++) {
2548                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2549                         if (i == 0 && !old_prog)
2550                                 virtnet_clear_guest_offloads(vi);
2551                 }
2552         } else {
2553                 vi->xdp_enabled = false;
2554         }
2555
2556         for (i = 0; i < vi->max_queue_pairs; i++) {
2557                 if (old_prog)
2558                         bpf_prog_put(old_prog);
2559                 if (netif_running(dev)) {
2560                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2561                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2562                                                &vi->sq[i].napi);
2563                 }
2564         }
2565
2566         return 0;
2567
2568 err:
2569         if (!prog) {
2570                 virtnet_clear_guest_offloads(vi);
2571                 for (i = 0; i < vi->max_queue_pairs; i++)
2572                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2573         }
2574
2575         if (netif_running(dev)) {
2576                 for (i = 0; i < vi->max_queue_pairs; i++) {
2577                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2578                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2579                                                &vi->sq[i].napi);
2580                 }
2581         }
2582         if (prog)
2583                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2584         return err;
2585 }
2586
2587 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2588 {
2589         switch (xdp->command) {
2590         case XDP_SETUP_PROG:
2591                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2592         default:
2593                 return -EINVAL;
2594         }
2595 }
2596
2597 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2598                                       size_t len)
2599 {
2600         struct virtnet_info *vi = netdev_priv(dev);
2601         int ret;
2602
2603         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2604                 return -EOPNOTSUPP;
2605
2606         ret = snprintf(buf, len, "sby");
2607         if (ret >= len)
2608                 return -EOPNOTSUPP;
2609
2610         return 0;
2611 }
2612
2613 static int virtnet_set_features(struct net_device *dev,
2614                                 netdev_features_t features)
2615 {
2616         struct virtnet_info *vi = netdev_priv(dev);
2617         u64 offloads;
2618         int err;
2619
2620         if ((dev->features ^ features) & NETIF_F_GRO_HW) {
2621                 if (vi->xdp_enabled)
2622                         return -EBUSY;
2623
2624                 if (features & NETIF_F_GRO_HW)
2625                         offloads = vi->guest_offloads_capable;
2626                 else
2627                         offloads = vi->guest_offloads_capable &
2628                                    ~GUEST_OFFLOAD_GRO_HW_MASK;
2629
2630                 err = virtnet_set_guest_offloads(vi, offloads);
2631                 if (err)
2632                         return err;
2633                 vi->guest_offloads = offloads;
2634         }
2635
2636         return 0;
2637 }
2638
2639 static const struct net_device_ops virtnet_netdev = {
2640         .ndo_open            = virtnet_open,
2641         .ndo_stop            = virtnet_close,
2642         .ndo_start_xmit      = start_xmit,
2643         .ndo_validate_addr   = eth_validate_addr,
2644         .ndo_set_mac_address = virtnet_set_mac_address,
2645         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2646         .ndo_get_stats64     = virtnet_stats,
2647         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2648         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2649         .ndo_bpf                = virtnet_xdp,
2650         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2651         .ndo_features_check     = passthru_features_check,
2652         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2653         .ndo_set_features       = virtnet_set_features,
2654 };
2655
2656 static void virtnet_config_changed_work(struct work_struct *work)
2657 {
2658         struct virtnet_info *vi =
2659                 container_of(work, struct virtnet_info, config_work);
2660         u16 v;
2661
2662         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2663                                  struct virtio_net_config, status, &v) < 0)
2664                 return;
2665
2666         if (v & VIRTIO_NET_S_ANNOUNCE) {
2667                 netdev_notify_peers(vi->dev);
2668                 virtnet_ack_link_announce(vi);
2669         }
2670
2671         /* Ignore unknown (future) status bits */
2672         v &= VIRTIO_NET_S_LINK_UP;
2673
2674         if (vi->status == v)
2675                 return;
2676
2677         vi->status = v;
2678
2679         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2680                 virtnet_update_settings(vi);
2681                 netif_carrier_on(vi->dev);
2682                 netif_tx_wake_all_queues(vi->dev);
2683         } else {
2684                 netif_carrier_off(vi->dev);
2685                 netif_tx_stop_all_queues(vi->dev);
2686         }
2687 }
2688
2689 static void virtnet_config_changed(struct virtio_device *vdev)
2690 {
2691         struct virtnet_info *vi = vdev->priv;
2692
2693         schedule_work(&vi->config_work);
2694 }
2695
2696 static void virtnet_free_queues(struct virtnet_info *vi)
2697 {
2698         int i;
2699
2700         for (i = 0; i < vi->max_queue_pairs; i++) {
2701                 __netif_napi_del(&vi->rq[i].napi);
2702                 __netif_napi_del(&vi->sq[i].napi);
2703         }
2704
2705         /* We called __netif_napi_del(),
2706          * we need to respect an RCU grace period before freeing vi->rq
2707          */
2708         synchronize_net();
2709
2710         kfree(vi->rq);
2711         kfree(vi->sq);
2712         kfree(vi->ctrl);
2713 }
2714
2715 static void _free_receive_bufs(struct virtnet_info *vi)
2716 {
2717         struct bpf_prog *old_prog;
2718         int i;
2719
2720         for (i = 0; i < vi->max_queue_pairs; i++) {
2721                 while (vi->rq[i].pages)
2722                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2723
2724                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2725                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2726                 if (old_prog)
2727                         bpf_prog_put(old_prog);
2728         }
2729 }
2730
2731 static void free_receive_bufs(struct virtnet_info *vi)
2732 {
2733         rtnl_lock();
2734         _free_receive_bufs(vi);
2735         rtnl_unlock();
2736 }
2737
2738 static void free_receive_page_frags(struct virtnet_info *vi)
2739 {
2740         int i;
2741         for (i = 0; i < vi->max_queue_pairs; i++)
2742                 if (vi->rq[i].alloc_frag.page)
2743                         put_page(vi->rq[i].alloc_frag.page);
2744 }
2745
2746 static void free_unused_bufs(struct virtnet_info *vi)
2747 {
2748         void *buf;
2749         int i;
2750
2751         for (i = 0; i < vi->max_queue_pairs; i++) {
2752                 struct virtqueue *vq = vi->sq[i].vq;
2753                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2754                         if (!is_xdp_frame(buf))
2755                                 dev_kfree_skb(buf);
2756                         else
2757                                 xdp_return_frame(ptr_to_xdp(buf));
2758                 }
2759         }
2760
2761         for (i = 0; i < vi->max_queue_pairs; i++) {
2762                 struct virtqueue *vq = vi->rq[i].vq;
2763
2764                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2765                         if (vi->mergeable_rx_bufs) {
2766                                 put_page(virt_to_head_page(buf));
2767                         } else if (vi->big_packets) {
2768                                 give_pages(&vi->rq[i], buf);
2769                         } else {
2770                                 put_page(virt_to_head_page(buf));
2771                         }
2772                 }
2773         }
2774 }
2775
2776 static void virtnet_del_vqs(struct virtnet_info *vi)
2777 {
2778         struct virtio_device *vdev = vi->vdev;
2779
2780         virtnet_clean_affinity(vi);
2781
2782         vdev->config->del_vqs(vdev);
2783
2784         virtnet_free_queues(vi);
2785 }
2786
2787 /* How large should a single buffer be so a queue full of these can fit at
2788  * least one full packet?
2789  * Logic below assumes the mergeable buffer header is used.
2790  */
2791 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2792 {
2793         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2794         unsigned int rq_size = virtqueue_get_vring_size(vq);
2795         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2796         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2797         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2798
2799         return max(max(min_buf_len, hdr_len) - hdr_len,
2800                    (unsigned int)GOOD_PACKET_LEN);
2801 }
2802
2803 static int virtnet_find_vqs(struct virtnet_info *vi)
2804 {
2805         vq_callback_t **callbacks;
2806         struct virtqueue **vqs;
2807         int ret = -ENOMEM;
2808         int i, total_vqs;
2809         const char **names;
2810         bool *ctx;
2811
2812         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2813          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2814          * possible control vq.
2815          */
2816         total_vqs = vi->max_queue_pairs * 2 +
2817                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2818
2819         /* Allocate space for find_vqs parameters */
2820         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2821         if (!vqs)
2822                 goto err_vq;
2823         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2824         if (!callbacks)
2825                 goto err_callback;
2826         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2827         if (!names)
2828                 goto err_names;
2829         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2830                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2831                 if (!ctx)
2832                         goto err_ctx;
2833         } else {
2834                 ctx = NULL;
2835         }
2836
2837         /* Parameters for control virtqueue, if any */
2838         if (vi->has_cvq) {
2839                 callbacks[total_vqs - 1] = NULL;
2840                 names[total_vqs - 1] = "control";
2841         }
2842
2843         /* Allocate/initialize parameters for send/receive virtqueues */
2844         for (i = 0; i < vi->max_queue_pairs; i++) {
2845                 callbacks[rxq2vq(i)] = skb_recv_done;
2846                 callbacks[txq2vq(i)] = skb_xmit_done;
2847                 sprintf(vi->rq[i].name, "input.%d", i);
2848                 sprintf(vi->sq[i].name, "output.%d", i);
2849                 names[rxq2vq(i)] = vi->rq[i].name;
2850                 names[txq2vq(i)] = vi->sq[i].name;
2851                 if (ctx)
2852                         ctx[rxq2vq(i)] = true;
2853         }
2854
2855         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2856                                          names, ctx, NULL);
2857         if (ret)
2858                 goto err_find;
2859
2860         if (vi->has_cvq) {
2861                 vi->cvq = vqs[total_vqs - 1];
2862                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2863                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2864         }
2865
2866         for (i = 0; i < vi->max_queue_pairs; i++) {
2867                 vi->rq[i].vq = vqs[rxq2vq(i)];
2868                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2869                 vi->sq[i].vq = vqs[txq2vq(i)];
2870         }
2871
2872         /* run here: ret == 0. */
2873
2874
2875 err_find:
2876         kfree(ctx);
2877 err_ctx:
2878         kfree(names);
2879 err_names:
2880         kfree(callbacks);
2881 err_callback:
2882         kfree(vqs);
2883 err_vq:
2884         return ret;
2885 }
2886
2887 static int virtnet_alloc_queues(struct virtnet_info *vi)
2888 {
2889         int i;
2890
2891         vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2892         if (!vi->ctrl)
2893                 goto err_ctrl;
2894         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2895         if (!vi->sq)
2896                 goto err_sq;
2897         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2898         if (!vi->rq)
2899                 goto err_rq;
2900
2901         INIT_DELAYED_WORK(&vi->refill, refill_work);
2902         for (i = 0; i < vi->max_queue_pairs; i++) {
2903                 vi->rq[i].pages = NULL;
2904                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2905                                napi_weight);
2906                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2907                                   napi_tx ? napi_weight : 0);
2908
2909                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2910                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2911                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2912
2913                 u64_stats_init(&vi->rq[i].stats.syncp);
2914                 u64_stats_init(&vi->sq[i].stats.syncp);
2915         }
2916
2917         return 0;
2918
2919 err_rq:
2920         kfree(vi->sq);
2921 err_sq:
2922         kfree(vi->ctrl);
2923 err_ctrl:
2924         return -ENOMEM;
2925 }
2926
2927 static int init_vqs(struct virtnet_info *vi)
2928 {
2929         int ret;
2930
2931         /* Allocate send & receive queues */
2932         ret = virtnet_alloc_queues(vi);
2933         if (ret)
2934                 goto err;
2935
2936         ret = virtnet_find_vqs(vi);
2937         if (ret)
2938                 goto err_free;
2939
2940         get_online_cpus();
2941         virtnet_set_affinity(vi);
2942         put_online_cpus();
2943
2944         return 0;
2945
2946 err_free:
2947         virtnet_free_queues(vi);
2948 err:
2949         return ret;
2950 }
2951
2952 #ifdef CONFIG_SYSFS
2953 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2954                 char *buf)
2955 {
2956         struct virtnet_info *vi = netdev_priv(queue->dev);
2957         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2958         unsigned int headroom = virtnet_get_headroom(vi);
2959         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2960         struct ewma_pkt_len *avg;
2961
2962         BUG_ON(queue_index >= vi->max_queue_pairs);
2963         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2964         return sprintf(buf, "%u\n",
2965                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
2966                                        SKB_DATA_ALIGN(headroom + tailroom)));
2967 }
2968
2969 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2970         __ATTR_RO(mergeable_rx_buffer_size);
2971
2972 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2973         &mergeable_rx_buffer_size_attribute.attr,
2974         NULL
2975 };
2976
2977 static const struct attribute_group virtio_net_mrg_rx_group = {
2978         .name = "virtio_net",
2979         .attrs = virtio_net_mrg_rx_attrs
2980 };
2981 #endif
2982
2983 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2984                                     unsigned int fbit,
2985                                     const char *fname, const char *dname)
2986 {
2987         if (!virtio_has_feature(vdev, fbit))
2988                 return false;
2989
2990         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2991                 fname, dname);
2992
2993         return true;
2994 }
2995
2996 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2997         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2998
2999 static bool virtnet_validate_features(struct virtio_device *vdev)
3000 {
3001         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3002             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3003                              "VIRTIO_NET_F_CTRL_VQ") ||
3004              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3005                              "VIRTIO_NET_F_CTRL_VQ") ||
3006              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3007                              "VIRTIO_NET_F_CTRL_VQ") ||
3008              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3009              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3010                              "VIRTIO_NET_F_CTRL_VQ"))) {
3011                 return false;
3012         }
3013
3014         return true;
3015 }
3016
3017 #define MIN_MTU ETH_MIN_MTU
3018 #define MAX_MTU ETH_MAX_MTU
3019
3020 static int virtnet_validate(struct virtio_device *vdev)
3021 {
3022         if (!vdev->config->get) {
3023                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3024                         __func__);
3025                 return -EINVAL;
3026         }
3027
3028         if (!virtnet_validate_features(vdev))
3029                 return -EINVAL;
3030
3031         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3032                 int mtu = virtio_cread16(vdev,
3033                                          offsetof(struct virtio_net_config,
3034                                                   mtu));
3035                 if (mtu < MIN_MTU)
3036                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3037         }
3038
3039         return 0;
3040 }
3041
3042 static int virtnet_probe(struct virtio_device *vdev)
3043 {
3044         int i, err = -ENOMEM;
3045         struct net_device *dev;
3046         struct virtnet_info *vi;
3047         u16 max_queue_pairs;
3048         int mtu;
3049
3050         /* Find if host supports multiqueue virtio_net device */
3051         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3052                                    struct virtio_net_config,
3053                                    max_virtqueue_pairs, &max_queue_pairs);
3054
3055         /* We need at least 2 queue's */
3056         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3057             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3058             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3059                 max_queue_pairs = 1;
3060
3061         /* Allocate ourselves a network device with room for our info */
3062         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3063         if (!dev)
3064                 return -ENOMEM;
3065
3066         /* Set up network device as normal. */
3067         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
3068         dev->netdev_ops = &virtnet_netdev;
3069         dev->features = NETIF_F_HIGHDMA;
3070
3071         dev->ethtool_ops = &virtnet_ethtool_ops;
3072         SET_NETDEV_DEV(dev, &vdev->dev);
3073
3074         /* Do we support "hardware" checksums? */
3075         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3076                 /* This opens up the world of extra features. */
3077                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3078                 if (csum)
3079                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3080
3081                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3082                         dev->hw_features |= NETIF_F_TSO
3083                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3084                 }
3085                 /* Individual feature bits: what can host handle? */
3086                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3087                         dev->hw_features |= NETIF_F_TSO;
3088                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3089                         dev->hw_features |= NETIF_F_TSO6;
3090                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3091                         dev->hw_features |= NETIF_F_TSO_ECN;
3092
3093                 dev->features |= NETIF_F_GSO_ROBUST;
3094
3095                 if (gso)
3096                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3097                 /* (!csum && gso) case will be fixed by register_netdev() */
3098         }
3099         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3100                 dev->features |= NETIF_F_RXCSUM;
3101         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3102             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3103                 dev->features |= NETIF_F_GRO_HW;
3104         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3105                 dev->hw_features |= NETIF_F_GRO_HW;
3106
3107         dev->vlan_features = dev->features;
3108
3109         /* MTU range: 68 - 65535 */
3110         dev->min_mtu = MIN_MTU;
3111         dev->max_mtu = MAX_MTU;
3112
3113         /* Configuration may specify what MAC to use.  Otherwise random. */
3114         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3115                 virtio_cread_bytes(vdev,
3116                                    offsetof(struct virtio_net_config, mac),
3117                                    dev->dev_addr, dev->addr_len);
3118         else
3119                 eth_hw_addr_random(dev);
3120
3121         /* Set up our device-specific information */
3122         vi = netdev_priv(dev);
3123         vi->dev = dev;
3124         vi->vdev = vdev;
3125         vdev->priv = vi;
3126
3127         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3128         spin_lock_init(&vi->refill_lock);
3129
3130         /* If we can receive ANY GSO packets, we must allocate large ones. */
3131         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3132             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3133             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3134             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3135                 vi->big_packets = true;
3136
3137         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3138                 vi->mergeable_rx_bufs = true;
3139
3140         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3141             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3142                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3143         else
3144                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3145
3146         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3147             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3148                 vi->any_header_sg = true;
3149
3150         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3151                 vi->has_cvq = true;
3152
3153         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3154                 mtu = virtio_cread16(vdev,
3155                                      offsetof(struct virtio_net_config,
3156                                               mtu));
3157                 if (mtu < dev->min_mtu) {
3158                         /* Should never trigger: MTU was previously validated
3159                          * in virtnet_validate.
3160                          */
3161                         dev_err(&vdev->dev,
3162                                 "device MTU appears to have changed it is now %d < %d",
3163                                 mtu, dev->min_mtu);
3164                         err = -EINVAL;
3165                         goto free;
3166                 }
3167
3168                 dev->mtu = mtu;
3169                 dev->max_mtu = mtu;
3170
3171                 /* TODO: size buffers correctly in this case. */
3172                 if (dev->mtu > ETH_DATA_LEN)
3173                         vi->big_packets = true;
3174         }
3175
3176         if (vi->any_header_sg)
3177                 dev->needed_headroom = vi->hdr_len;
3178
3179         /* Enable multiqueue by default */
3180         if (num_online_cpus() >= max_queue_pairs)
3181                 vi->curr_queue_pairs = max_queue_pairs;
3182         else
3183                 vi->curr_queue_pairs = num_online_cpus();
3184         vi->max_queue_pairs = max_queue_pairs;
3185
3186         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3187         err = init_vqs(vi);
3188         if (err)
3189                 goto free;
3190
3191 #ifdef CONFIG_SYSFS
3192         if (vi->mergeable_rx_bufs)
3193                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3194 #endif
3195         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3196         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3197
3198         virtnet_init_settings(dev);
3199
3200         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3201                 vi->failover = net_failover_create(vi->dev);
3202                 if (IS_ERR(vi->failover)) {
3203                         err = PTR_ERR(vi->failover);
3204                         goto free_vqs;
3205                 }
3206         }
3207
3208         /* serialize netdev register + virtio_device_ready() with ndo_open() */
3209         rtnl_lock();
3210
3211         err = register_netdevice(dev);
3212         if (err) {
3213                 pr_debug("virtio_net: registering device failed\n");
3214                 rtnl_unlock();
3215                 goto free_failover;
3216         }
3217
3218         virtio_device_ready(vdev);
3219
3220         rtnl_unlock();
3221
3222         err = virtnet_cpu_notif_add(vi);
3223         if (err) {
3224                 pr_debug("virtio_net: registering cpu notifier failed\n");
3225                 goto free_unregister_netdev;
3226         }
3227
3228         virtnet_set_queues(vi, vi->curr_queue_pairs);
3229
3230         /* Assume link up if device can't report link status,
3231            otherwise get link status from config. */
3232         netif_carrier_off(dev);
3233         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3234                 schedule_work(&vi->config_work);
3235         } else {
3236                 vi->status = VIRTIO_NET_S_LINK_UP;
3237                 virtnet_update_settings(vi);
3238                 netif_carrier_on(dev);
3239         }
3240
3241         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3242                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3243                         set_bit(guest_offloads[i], &vi->guest_offloads);
3244         vi->guest_offloads_capable = vi->guest_offloads;
3245
3246         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3247                  dev->name, max_queue_pairs);
3248
3249         return 0;
3250
3251 free_unregister_netdev:
3252         vi->vdev->config->reset(vdev);
3253
3254         unregister_netdev(dev);
3255 free_failover:
3256         net_failover_destroy(vi->failover);
3257 free_vqs:
3258         cancel_delayed_work_sync(&vi->refill);
3259         free_receive_page_frags(vi);
3260         virtnet_del_vqs(vi);
3261 free:
3262         free_netdev(dev);
3263         return err;
3264 }
3265
3266 static void remove_vq_common(struct virtnet_info *vi)
3267 {
3268         vi->vdev->config->reset(vi->vdev);
3269
3270         /* Free unused buffers in both send and recv, if any. */
3271         free_unused_bufs(vi);
3272
3273         free_receive_bufs(vi);
3274
3275         free_receive_page_frags(vi);
3276
3277         virtnet_del_vqs(vi);
3278 }
3279
3280 static void virtnet_remove(struct virtio_device *vdev)
3281 {
3282         struct virtnet_info *vi = vdev->priv;
3283
3284         virtnet_cpu_notif_remove(vi);
3285
3286         /* Make sure no work handler is accessing the device. */
3287         flush_work(&vi->config_work);
3288
3289         unregister_netdev(vi->dev);
3290
3291         net_failover_destroy(vi->failover);
3292
3293         remove_vq_common(vi);
3294
3295         free_netdev(vi->dev);
3296 }
3297
3298 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3299 {
3300         struct virtnet_info *vi = vdev->priv;
3301
3302         virtnet_cpu_notif_remove(vi);
3303         virtnet_freeze_down(vdev);
3304         remove_vq_common(vi);
3305
3306         return 0;
3307 }
3308
3309 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3310 {
3311         struct virtnet_info *vi = vdev->priv;
3312         int err;
3313
3314         err = virtnet_restore_up(vdev);
3315         if (err)
3316                 return err;
3317         virtnet_set_queues(vi, vi->curr_queue_pairs);
3318
3319         err = virtnet_cpu_notif_add(vi);
3320         if (err) {
3321                 virtnet_freeze_down(vdev);
3322                 remove_vq_common(vi);
3323                 return err;
3324         }
3325
3326         return 0;
3327 }
3328
3329 static struct virtio_device_id id_table[] = {
3330         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3331         { 0 },
3332 };
3333
3334 #define VIRTNET_FEATURES \
3335         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3336         VIRTIO_NET_F_MAC, \
3337         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3338         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3339         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3340         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3341         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3342         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3343         VIRTIO_NET_F_CTRL_MAC_ADDR, \
3344         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3345         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3346
3347 static unsigned int features[] = {
3348         VIRTNET_FEATURES,
3349 };
3350
3351 static unsigned int features_legacy[] = {
3352         VIRTNET_FEATURES,
3353         VIRTIO_NET_F_GSO,
3354         VIRTIO_F_ANY_LAYOUT,
3355 };
3356
3357 static struct virtio_driver virtio_net_driver = {
3358         .feature_table = features,
3359         .feature_table_size = ARRAY_SIZE(features),
3360         .feature_table_legacy = features_legacy,
3361         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3362         .driver.name =  KBUILD_MODNAME,
3363         .driver.owner = THIS_MODULE,
3364         .id_table =     id_table,
3365         .validate =     virtnet_validate,
3366         .probe =        virtnet_probe,
3367         .remove =       virtnet_remove,
3368         .config_changed = virtnet_config_changed,
3369 #ifdef CONFIG_PM_SLEEP
3370         .freeze =       virtnet_freeze,
3371         .restore =      virtnet_restore,
3372 #endif
3373 };
3374
3375 static __init int virtio_net_driver_init(void)
3376 {
3377         int ret;
3378
3379         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3380                                       virtnet_cpu_online,
3381                                       virtnet_cpu_down_prep);
3382         if (ret < 0)
3383                 goto out;
3384         virtionet_online = ret;
3385         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3386                                       NULL, virtnet_cpu_dead);
3387         if (ret)
3388                 goto err_dead;
3389
3390         ret = register_virtio_driver(&virtio_net_driver);
3391         if (ret)
3392                 goto err_virtio;
3393         return 0;
3394 err_virtio:
3395         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3396 err_dead:
3397         cpuhp_remove_multi_state(virtionet_online);
3398 out:
3399         return ret;
3400 }
3401 module_init(virtio_net_driver_init);
3402
3403 static __exit void virtio_net_driver_exit(void)
3404 {
3405         unregister_virtio_driver(&virtio_net_driver);
3406         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3407         cpuhp_remove_multi_state(virtionet_online);
3408 }
3409 module_exit(virtio_net_driver_exit);
3410
3411 MODULE_DEVICE_TABLE(virtio, id_table);
3412 MODULE_DESCRIPTION("Virtio network driver");
3413 MODULE_LICENSE("GPL");