drivers/net/virtio_net.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* A network driver using virtio.
   3  *
   4  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   5  */
   6 //#define DEBUG
   7 #include <linux/netdevice.h>
   8 #include <linux/etherdevice.h>
   9 #include <linux/ethtool.h>
  10 #include <linux/module.h>
  11 #include <linux/virtio.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/bpf.h>
  14 #include <linux/bpf_trace.h>
  15 #include <linux/scatterlist.h>
  16 #include <linux/if_vlan.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/average.h>
  20 #include <linux/filter.h>
  21 #include <linux/kernel.h>
  22 #include <net/route.h>
  23 #include <net/xdp.h>
  24 #include <net/net_failover.h>
  25
  26 static int napi_weight = NAPI_POLL_WEIGHT;
  27 module_param(napi_weight, int, 0444);
  28
  29 static bool csum = true, gso = true, napi_tx = true;
  30 module_param(csum, bool, 0444);
  31 module_param(gso, bool, 0444);
  32 module_param(napi_tx, bool, 0644);
  33
  34 /* FIXME: MTU in config. */
  35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  36 #define GOOD_COPY_LEN   128
  37
  38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  39
  40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  41 #define VIRTIO_XDP_HEADROOM 256
  42
  43 /* Separating two types of XDP xmit */
  44 #define VIRTIO_XDP_TX           BIT(0)
  45 #define VIRTIO_XDP_REDIR        BIT(1)
  46
  47 #define VIRTIO_XDP_FLAG BIT(0)
  48
  49 /* RX packet size EWMA. The average packet size is used to determine the packet
  50  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  51  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  52  * term, transient changes in packet size.
  53  */
  54 DECLARE_EWMA(pkt_len, 0, 64)
  55
  56 #define VIRTNET_DRIVER_VERSION "1.0.0"
  57
  58 static const unsigned long guest_offloads[] = {
  59         VIRTIO_NET_F_GUEST_TSO4,
  60         VIRTIO_NET_F_GUEST_TSO6,
  61         VIRTIO_NET_F_GUEST_ECN,
  62         VIRTIO_NET_F_GUEST_UFO,
  63         VIRTIO_NET_F_GUEST_CSUM
  64 };
  65
  66 #define GUEST_OFFLOAD_LRO_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
  67                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
  68                                 (1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
  69                                 (1ULL << VIRTIO_NET_F_GUEST_UFO))
  70
  71 struct virtnet_stat_desc {
  72         char desc[ETH_GSTRING_LEN];
  73         size_t offset;
  74 };
  75
  76 struct virtnet_sq_stats {
  77         struct u64_stats_sync syncp;
  78         u64 packets;
  79         u64 bytes;
  80         u64 xdp_tx;
  81         u64 xdp_tx_drops;
  82         u64 kicks;
  83 };
  84
  85 struct virtnet_rq_stats {
  86         struct u64_stats_sync syncp;
  87         u64 packets;
  88         u64 bytes;
  89         u64 drops;
  90         u64 xdp_packets;
  91         u64 xdp_tx;
  92         u64 xdp_redirects;
  93         u64 xdp_drops;
  94         u64 kicks;
  95 };
  96
  97 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
  98 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
  99
 100 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
 101         { "packets",            VIRTNET_SQ_STAT(packets) },
 102         { "bytes",              VIRTNET_SQ_STAT(bytes) },
 103         { "xdp_tx",             VIRTNET_SQ_STAT(xdp_tx) },
 104         { "xdp_tx_drops",       VIRTNET_SQ_STAT(xdp_tx_drops) },
 105         { "kicks",              VIRTNET_SQ_STAT(kicks) },
 106 };
 107
 108 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
 109         { "packets",            VIRTNET_RQ_STAT(packets) },
 110         { "bytes",              VIRTNET_RQ_STAT(bytes) },
 111         { "drops",              VIRTNET_RQ_STAT(drops) },
 112         { "xdp_packets",        VIRTNET_RQ_STAT(xdp_packets) },
 113         { "xdp_tx",             VIRTNET_RQ_STAT(xdp_tx) },
 114         { "xdp_redirects",      VIRTNET_RQ_STAT(xdp_redirects) },
 115         { "xdp_drops",          VIRTNET_RQ_STAT(xdp_drops) },
 116         { "kicks",              VIRTNET_RQ_STAT(kicks) },
 117 };
 118
 119 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 120 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 121
 122 /* Internal representation of a send virtqueue */
 123 struct send_queue {
 124         /* Virtqueue associated with this send _queue */
 125         struct virtqueue *vq;
 126
 127         /* TX: fragments + linear part + virtio header */
 128         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 129
 130         /* Name of the send queue: output.$index */
 131         char name[40];
 132
 133         struct virtnet_sq_stats stats;
 134
 135         struct napi_struct napi;
 136 };
 137
 138 /* Internal representation of a receive virtqueue */
 139 struct receive_queue {
 140         /* Virtqueue associated with this receive_queue */
 141         struct virtqueue *vq;
 142
 143         struct napi_struct napi;
 144
 145         struct bpf_prog __rcu *xdp_prog;
 146
 147         struct virtnet_rq_stats stats;
 148
 149         /* Chain pages by the private ptr. */
 150         struct page *pages;
 151
 152         /* Average packet length for mergeable receive buffers. */
 153         struct ewma_pkt_len mrg_avg_pkt_len;
 154
 155         /* Page frag for packet buffer allocation. */
 156         struct page_frag alloc_frag;
 157
 158         /* RX: fragments + linear part + virtio header */
 159         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 160
 161         /* Min single buffer size for mergeable buffers case. */
 162         unsigned int min_buf_len;
 163
 164         /* Name of this receive queue: input.$index */
 165         char name[40];
 166
 167         struct xdp_rxq_info xdp_rxq;
 168 };
 169
 170 /* Control VQ buffers: protected by the rtnl lock */
 171 struct control_buf {
 172         struct virtio_net_ctrl_hdr hdr;
 173         virtio_net_ctrl_ack status;
 174         struct virtio_net_ctrl_mq mq;
 175         u8 promisc;
 176         u8 allmulti;
 177         __virtio16 vid;
 178         __virtio64 offloads;
 179 };
 180
 181 struct virtnet_info {
 182         struct virtio_device *vdev;
 183         struct virtqueue *cvq;
 184         struct net_device *dev;
 185         struct send_queue *sq;
 186         struct receive_queue *rq;
 187         unsigned int status;
 188
 189         /* Max # of queue pairs supported by the device */
 190         u16 max_queue_pairs;
 191
 192         /* # of queue pairs currently used by the driver */
 193         u16 curr_queue_pairs;
 194
 195         /* # of XDP queue pairs currently used by the driver */
 196         u16 xdp_queue_pairs;
 197
 198         /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
 199         bool xdp_enabled;
 200
 201         /* I like... big packets and I cannot lie! */
 202         bool big_packets;
 203
 204         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 205         bool mergeable_rx_bufs;
 206
 207         /* Has control virtqueue */
 208         bool has_cvq;
 209
 210         /* Host can handle any s/g split between our header and packet data */
 211         bool any_header_sg;
 212
 213         /* Packet virtio header size */
 214         u8 hdr_len;
 215
 216         /* Work struct for refilling if we run low on memory. */
 217         struct delayed_work refill;
 218
 219         /* Work struct for config space updates */
 220         struct work_struct config_work;
 221
 222         /* Does the affinity hint is set for virtqueues? */
 223         bool affinity_hint_set;
 224
 225         /* CPU hotplug instances for online & dead */
 226         struct hlist_node node;
 227         struct hlist_node node_dead;
 228
 229         struct control_buf *ctrl;
 230
 231         /* Ethtool settings */
 232         u8 duplex;
 233         u32 speed;
 234
 235         unsigned long guest_offloads;
 236         unsigned long guest_offloads_capable;
 237
 238         /* failover when STANDBY feature enabled */
 239         struct failover *failover;
 240 };
 241
 242 struct padded_vnet_hdr {
 243         struct virtio_net_hdr_mrg_rxbuf hdr;
 244         /*
 245          * hdr is in a separate sg buffer, and data sg buffer shares same page
 246          * with this header sg. This padding makes next sg 16 byte aligned
 247          * after the header.
 248          */
 249         char padding[4];
 250 };
 251
 252 static bool is_xdp_frame(void *ptr)
 253 {
 254         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 255 }
 256
 257 static void *xdp_to_ptr(struct xdp_frame *ptr)
 258 {
 259         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 260 }
 261
 262 static struct xdp_frame *ptr_to_xdp(void *ptr)
 263 {
 264         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 265 }
 266
 267 /* Converting between virtqueue no. and kernel tx/rx queue no.
 268  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 269  */
 270 static int vq2txq(struct virtqueue *vq)
 271 {
 272         return (vq->index - 1) / 2;
 273 }
 274
 275 static int txq2vq(int txq)
 276 {
 277         return txq * 2 + 1;
 278 }
 279
 280 static int vq2rxq(struct virtqueue *vq)
 281 {
 282         return vq->index / 2;
 283 }
 284
 285 static int rxq2vq(int rxq)
 286 {
 287         return rxq * 2;
 288 }
 289
 290 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 291 {
 292         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 293 }
 294
 295 /*
 296  * private is used to chain pages for big packets, put the whole
 297  * most recent used list in the beginning for reuse
 298  */
 299 static void give_pages(struct receive_queue *rq, struct page *page)
 300 {
 301         struct page *end;
 302
 303         /* Find end of list, sew whole thing into vi->rq.pages. */
 304         for (end = page; end->private; end = (struct page *)end->private);
 305         end->private = (unsigned long)rq->pages;
 306         rq->pages = page;
 307 }
 308
 309 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 310 {
 311         struct page *p = rq->pages;
 312
 313         if (p) {
 314                 rq->pages = (struct page *)p->private;
 315                 /* clear private here, it is used to chain pages */
 316                 p->private = 0;
 317         } else
 318                 p = alloc_page(gfp_mask);
 319         return p;
 320 }
 321
 322 static void virtqueue_napi_schedule(struct napi_struct *napi,
 323                                     struct virtqueue *vq)
 324 {
 325         if (napi_schedule_prep(napi)) {
 326                 virtqueue_disable_cb(vq);
 327                 __napi_schedule(napi);
 328         }
 329 }
 330
 331 static void virtqueue_napi_complete(struct napi_struct *napi,
 332                                     struct virtqueue *vq, int processed)
 333 {
 334         int opaque;
 335
 336         opaque = virtqueue_enable_cb_prepare(vq);
 337         if (napi_complete_done(napi, processed)) {
 338                 if (unlikely(virtqueue_poll(vq, opaque)))
 339                         virtqueue_napi_schedule(napi, vq);
 340         } else {
 341                 virtqueue_disable_cb(vq);
 342         }
 343 }
 344
 345 static void skb_xmit_done(struct virtqueue *vq)
 346 {
 347         struct virtnet_info *vi = vq->vdev->priv;
 348         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 349
 350         /* Suppress further interrupts. */
 351         virtqueue_disable_cb(vq);
 352
 353         if (napi->weight)
 354                 virtqueue_napi_schedule(napi, vq);
 355         else
 356                 /* We were probably waiting for more output buffers. */
 357                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 358 }
 359
 360 #define MRG_CTX_HEADER_SHIFT 22
 361 static void *mergeable_len_to_ctx(unsigned int truesize,
 362                                   unsigned int headroom)
 363 {
 364         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 365 }
 366
 367 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 368 {
 369         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 370 }
 371
 372 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 373 {
 374         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 375 }
 376
 377 /* Called from bottom half context */
 378 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 379                                    struct receive_queue *rq,
 380                                    struct page *page, unsigned int offset,
 381                                    unsigned int len, unsigned int truesize,
 382                                    bool hdr_valid, unsigned int metasize,
 383                                    bool whole_page)
 384 {
 385         struct sk_buff *skb;
 386         struct virtio_net_hdr_mrg_rxbuf *hdr;
 387         unsigned int copy, hdr_len, hdr_padded_len;
 388         struct page *page_to_free = NULL;
 389         int tailroom, shinfo_size;
 390         char *p, *hdr_p, *buf;
 391
 392         p = page_address(page) + offset;
 393         hdr_p = p;
 394
 395         hdr_len = vi->hdr_len;
 396         if (vi->mergeable_rx_bufs)
 397                 hdr_padded_len = sizeof(*hdr);
 398         else
 399                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 400
 401         /* If whole_page, there is an offset between the beginning of the
 402          * data and the allocated space, otherwise the data and the allocated
 403          * space are aligned.
 404          *
 405          * Buffers with headroom use PAGE_SIZE as alloc size, see
 406          * add_recvbuf_mergeable() + get_mergeable_buf_len()
 407          */
 408         if (whole_page) {
 409                 /* Buffers with whole_page use PAGE_SIZE as alloc size,
 410                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 411                  */
 412                 truesize = PAGE_SIZE;
 413
 414                 /* page maybe head page, so we should get the buf by p, not the
 415                  * page
 416                  */
 417                 tailroom = truesize - len - offset_in_page(p);
 418                 buf = (char *)((unsigned long)p & PAGE_MASK);
 419         } else {
 420                 tailroom = truesize - len;
 421                 buf = p;
 422         }
 423
 424         len -= hdr_len;
 425         offset += hdr_padded_len;
 426         p += hdr_padded_len;
 427
 428         shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 429
 430         /* copy small packet so we can reuse these pages */
 431         if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
 432                 skb = build_skb(buf, truesize);
 433                 if (unlikely(!skb))
 434                         return NULL;
 435
 436                 skb_reserve(skb, p - buf);
 437                 skb_put(skb, len);
 438                 goto ok;
 439         }
 440
 441         /* copy small packet so we can reuse these pages for small data */
 442         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 443         if (unlikely(!skb))
 444                 return NULL;
 445
 446         /* Copy all frame if it fits skb->head, otherwise
 447          * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
 448          */
 449         if (len <= skb_tailroom(skb))
 450                 copy = len;
 451         else
 452                 copy = ETH_HLEN + metasize;
 453         skb_put_data(skb, p, copy);
 454
 455         len -= copy;
 456         offset += copy;
 457
 458         if (vi->mergeable_rx_bufs) {
 459                 if (len)
 460                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 461                 else
 462                         page_to_free = page;
 463                 goto ok;
 464         }
 465
 466         /*
 467          * Verify that we can indeed put this data into a skb.
 468          * This is here to handle cases when the device erroneously
 469          * tries to receive more than is possible. This is usually
 470          * the case of a broken device.
 471          */
 472         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 473                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 474                 dev_kfree_skb(skb);
 475                 return NULL;
 476         }
 477         BUG_ON(offset >= PAGE_SIZE);
 478         while (len) {
 479                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 480                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 481                                 frag_size, truesize);
 482                 len -= frag_size;
 483                 page = (struct page *)page->private;
 484                 offset = 0;
 485         }
 486
 487         if (page)
 488                 give_pages(rq, page);
 489
 490 ok:
 491         /* hdr_valid means no XDP, so we can copy the vnet header */
 492         if (hdr_valid) {
 493                 hdr = skb_vnet_hdr(skb);
 494                 memcpy(hdr, hdr_p, hdr_len);
 495         }
 496         if (page_to_free)
 497                 put_page(page_to_free);
 498
 499         if (metasize) {
 500                 __skb_pull(skb, metasize);
 501                 skb_metadata_set(skb, metasize);
 502         }
 503
 504         return skb;
 505 }
 506
 507 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 508                                    struct send_queue *sq,
 509                                    struct xdp_frame *xdpf)
 510 {
 511         struct virtio_net_hdr_mrg_rxbuf *hdr;
 512         int err;
 513
 514         if (unlikely(xdpf->headroom < vi->hdr_len))
 515                 return -EOVERFLOW;
 516
 517         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 518         xdpf->data -= vi->hdr_len;
 519         /* Zero header and leave csum up to XDP layers */
 520         hdr = xdpf->data;
 521         memset(hdr, 0, vi->hdr_len);
 522         xdpf->len   += vi->hdr_len;
 523
 524         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 525
 526         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
 527                                    GFP_ATOMIC);
 528         if (unlikely(err))
 529                 return -ENOSPC; /* Caller handle free/refcnt */
 530
 531         return 0;
 532 }
 533
 534 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
 535  * the current cpu, so it does not need to be locked.
 536  *
 537  * Here we use marco instead of inline functions because we have to deal with
 538  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
 539  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
 540  * functions to perfectly solve these three problems at the same time.
 541  */
 542 #define virtnet_xdp_get_sq(vi) ({                                       \
 543         struct netdev_queue *txq;                                       \
 544         typeof(vi) v = (vi);                                            \
 545         unsigned int qp;                                                \
 546                                                                         \
 547         if (v->curr_queue_pairs > nr_cpu_ids) {                         \
 548                 qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
 549                 qp += smp_processor_id();                               \
 550                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 551                 __netif_tx_acquire(txq);                                \
 552         } else {                                                        \
 553                 qp = smp_processor_id() % v->curr_queue_pairs;          \
 554                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 555                 __netif_tx_lock(txq, raw_smp_processor_id());           \
 556         }                                                               \
 557         v->sq + qp;                                                     \
 558 })
 559
 560 #define virtnet_xdp_put_sq(vi, q) {                                     \
 561         struct netdev_queue *txq;                                       \
 562         typeof(vi) v = (vi);                                            \
 563                                                                         \
 564         txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
 565         if (v->curr_queue_pairs > nr_cpu_ids)                           \
 566                 __netif_tx_release(txq);                                \
 567         else                                                            \
 568                 __netif_tx_unlock(txq);                                 \
 569 }
 570
 571 static int virtnet_xdp_xmit(struct net_device *dev,
 572                             int n, struct xdp_frame **frames, u32 flags)
 573 {
 574         struct virtnet_info *vi = netdev_priv(dev);
 575         struct receive_queue *rq = vi->rq;
 576         struct bpf_prog *xdp_prog;
 577         struct send_queue *sq;
 578         unsigned int len;
 579         int packets = 0;
 580         int bytes = 0;
 581         int nxmit = 0;
 582         int kicks = 0;
 583         void *ptr;
 584         int ret;
 585         int i;
 586
 587         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 588          * indicate XDP resources have been successfully allocated.
 589          */
 590         xdp_prog = rcu_access_pointer(rq->xdp_prog);
 591         if (!xdp_prog)
 592                 return -ENXIO;
 593
 594         sq = virtnet_xdp_get_sq(vi);
 595
 596         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 597                 ret = -EINVAL;
 598                 goto out;
 599         }
 600
 601         /* Free up any pending old buffers before queueing new ones. */
 602         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 603                 if (likely(is_xdp_frame(ptr))) {
 604                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 605
 606                         bytes += frame->len;
 607                         xdp_return_frame(frame);
 608                 } else {
 609                         struct sk_buff *skb = ptr;
 610
 611                         bytes += skb->len;
 612                         napi_consume_skb(skb, false);
 613                 }
 614                 packets++;
 615         }
 616
 617         for (i = 0; i < n; i++) {
 618                 struct xdp_frame *xdpf = frames[i];
 619
 620                 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
 621                         break;
 622                 nxmit++;
 623         }
 624         ret = nxmit;
 625
 626         if (flags & XDP_XMIT_FLUSH) {
 627                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
 628                         kicks = 1;
 629         }
 630 out:
 631         u64_stats_update_begin(&sq->stats.syncp);
 632         sq->stats.bytes += bytes;
 633         sq->stats.packets += packets;
 634         sq->stats.xdp_tx += n;
 635         sq->stats.xdp_tx_drops += n - nxmit;
 636         sq->stats.kicks += kicks;
 637         u64_stats_update_end(&sq->stats.syncp);
 638
 639         virtnet_xdp_put_sq(vi, sq);
 640         return ret;
 641 }
 642
 643 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 644 {
 645         return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
 646 }
 647
 648 /* We copy the packet for XDP in the following cases:
 649  *
 650  * 1) Packet is scattered across multiple rx buffers.
 651  * 2) Headroom space is insufficient.
 652  *
 653  * This is inefficient but it's a temporary condition that
 654  * we hit right after XDP is enabled and until queue is refilled
 655  * with large buffers with sufficient headroom - so it should affect
 656  * at most queue size packets.
 657  * Afterwards, the conditions to enable
 658  * XDP should preclude the underlying device from sending packets
 659  * across multiple buffers (num_buf > 1), and we make sure buffers
 660  * have enough headroom.
 661  */
 662 static struct page *xdp_linearize_page(struct receive_queue *rq,
 663                                        u16 *num_buf,
 664                                        struct page *p,
 665                                        int offset,
 666                                        int page_off,
 667                                        unsigned int *len)
 668 {
 669         struct page *page = alloc_page(GFP_ATOMIC);
 670
 671         if (!page)
 672                 return NULL;
 673
 674         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 675         page_off += *len;
 676
 677         while (--*num_buf) {
 678                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 679                 unsigned int buflen;
 680                 void *buf;
 681                 int off;
 682
 683                 buf = virtqueue_get_buf(rq->vq, &buflen);
 684                 if (unlikely(!buf))
 685                         goto err_buf;
 686
 687                 p = virt_to_head_page(buf);
 688                 off = buf - page_address(p);
 689
 690                 /* guard against a misconfigured or uncooperative backend that
 691                  * is sending packet larger than the MTU.
 692                  */
 693                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 694                         put_page(p);
 695                         goto err_buf;
 696                 }
 697
 698                 memcpy(page_address(page) + page_off,
 699                        page_address(p) + off, buflen);
 700                 page_off += buflen;
 701                 put_page(p);
 702         }
 703
 704         /* Headroom does not contribute to packet length */
 705         *len = page_off - VIRTIO_XDP_HEADROOM;
 706         return page;
 707 err_buf:
 708         __free_pages(page, 0);
 709         return NULL;
 710 }
 711
 712 static struct sk_buff *receive_small(struct net_device *dev,
 713                                      struct virtnet_info *vi,
 714                                      struct receive_queue *rq,
 715                                      void *buf, void *ctx,
 716                                      unsigned int len,
 717                                      unsigned int *xdp_xmit,
 718                                      struct virtnet_rq_stats *stats)
 719 {
 720         struct sk_buff *skb;
 721         struct bpf_prog *xdp_prog;
 722         unsigned int xdp_headroom = (unsigned long)ctx;
 723         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 724         unsigned int headroom = vi->hdr_len + header_offset;
 725         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 726                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 727         struct page *page = virt_to_head_page(buf);
 728         unsigned int delta = 0;
 729         struct page *xdp_page;
 730         int err;
 731         unsigned int metasize = 0;
 732
 733         len -= vi->hdr_len;
 734         stats->bytes += len;
 735
 736         if (unlikely(len > GOOD_PACKET_LEN)) {
 737                 pr_debug("%s: rx error: len %u exceeds max size %d\n",
 738                          dev->name, len, GOOD_PACKET_LEN);
 739                 dev->stats.rx_length_errors++;
 740                 goto err_len;
 741         }
 742         rcu_read_lock();
 743         xdp_prog = rcu_dereference(rq->xdp_prog);
 744         if (xdp_prog) {
 745                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 746                 struct xdp_frame *xdpf;
 747                 struct xdp_buff xdp;
 748                 void *orig_data;
 749                 u32 act;
 750
 751                 if (unlikely(hdr->hdr.gso_type))
 752                         goto err_xdp;
 753
 754                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 755                         int offset = buf - page_address(page) + header_offset;
 756                         unsigned int tlen = len + vi->hdr_len;
 757                         u16 num_buf = 1;
 758
 759                         xdp_headroom = virtnet_get_headroom(vi);
 760                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 761                         headroom = vi->hdr_len + header_offset;
 762                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 763                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 764                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 765                                                       offset, header_offset,
 766                                                       &tlen);
 767                         if (!xdp_page)
 768                                 goto err_xdp;
 769
 770                         buf = page_address(xdp_page);
 771                         put_page(page);
 772                         page = xdp_page;
 773                 }
 774
 775                 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
 776                 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
 777                                  xdp_headroom, len, true);
 778                 orig_data = xdp.data;
 779                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 780                 stats->xdp_packets++;
 781
 782                 switch (act) {
 783                 case XDP_PASS:
 784                         /* Recalculate length in case bpf program changed it */
 785                         delta = orig_data - xdp.data;
 786                         len = xdp.data_end - xdp.data;
 787                         metasize = xdp.data - xdp.data_meta;
 788                         break;
 789                 case XDP_TX:
 790                         stats->xdp_tx++;
 791                         xdpf = xdp_convert_buff_to_frame(&xdp);
 792                         if (unlikely(!xdpf))
 793                                 goto err_xdp;
 794                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 795                         if (unlikely(!err)) {
 796                                 xdp_return_frame_rx_napi(xdpf);
 797                         } else if (unlikely(err < 0)) {
 798                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 799                                 goto err_xdp;
 800                         }
 801                         *xdp_xmit |= VIRTIO_XDP_TX;
 802                         rcu_read_unlock();
 803                         goto xdp_xmit;
 804                 case XDP_REDIRECT:
 805                         stats->xdp_redirects++;
 806                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 807                         if (err)
 808                                 goto err_xdp;
 809                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 810                         rcu_read_unlock();
 811                         goto xdp_xmit;
 812                 default:
 813                         bpf_warn_invalid_xdp_action(act);
 814                         fallthrough;
 815                 case XDP_ABORTED:
 816                         trace_xdp_exception(vi->dev, xdp_prog, act);
 817                         goto err_xdp;
 818                 case XDP_DROP:
 819                         goto err_xdp;
 820                 }
 821         }
 822         rcu_read_unlock();
 823
 824         skb = build_skb(buf, buflen);
 825         if (!skb) {
 826                 put_page(page);
 827                 goto err;
 828         }
 829         skb_reserve(skb, headroom - delta);
 830         skb_put(skb, len);
 831         if (!xdp_prog) {
 832                 buf += header_offset;
 833                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 834         } /* keep zeroed vnet hdr since XDP is loaded */
 835
 836         if (metasize)
 837                 skb_metadata_set(skb, metasize);
 838
 839 err:
 840         return skb;
 841
 842 err_xdp:
 843         rcu_read_unlock();
 844         stats->xdp_drops++;
 845 err_len:
 846         stats->drops++;
 847         put_page(page);
 848 xdp_xmit:
 849         return NULL;
 850 }
 851
 852 static struct sk_buff *receive_big(struct net_device *dev,
 853                                    struct virtnet_info *vi,
 854                                    struct receive_queue *rq,
 855                                    void *buf,
 856                                    unsigned int len,
 857                                    struct virtnet_rq_stats *stats)
 858 {
 859         struct page *page = buf;
 860         struct sk_buff *skb =
 861                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
 862
 863         stats->bytes += len - vi->hdr_len;
 864         if (unlikely(!skb))
 865                 goto err;
 866
 867         return skb;
 868
 869 err:
 870         stats->drops++;
 871         give_pages(rq, page);
 872         return NULL;
 873 }
 874
 875 static struct sk_buff *receive_mergeable(struct net_device *dev,
 876                                          struct virtnet_info *vi,
 877                                          struct receive_queue *rq,
 878                                          void *buf,
 879                                          void *ctx,
 880                                          unsigned int len,
 881                                          unsigned int *xdp_xmit,
 882                                          struct virtnet_rq_stats *stats)
 883 {
 884         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 885         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 886         struct page *page = virt_to_head_page(buf);
 887         int offset = buf - page_address(page);
 888         struct sk_buff *head_skb, *curr_skb;
 889         struct bpf_prog *xdp_prog;
 890         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
 891         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 892         unsigned int metasize = 0;
 893         unsigned int frame_sz;
 894         int err;
 895
 896         head_skb = NULL;
 897         stats->bytes += len - vi->hdr_len;
 898
 899         if (unlikely(len > truesize)) {
 900                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 901                          dev->name, len, (unsigned long)ctx);
 902                 dev->stats.rx_length_errors++;
 903                 goto err_skb;
 904         }
 905         rcu_read_lock();
 906         xdp_prog = rcu_dereference(rq->xdp_prog);
 907         if (xdp_prog) {
 908                 struct xdp_frame *xdpf;
 909                 struct page *xdp_page;
 910                 struct xdp_buff xdp;
 911                 void *data;
 912                 u32 act;
 913
 914                 /* Transient failure which in theory could occur if
 915                  * in-flight packets from before XDP was enabled reach
 916                  * the receive path after XDP is loaded.
 917                  */
 918                 if (unlikely(hdr->hdr.gso_type))
 919                         goto err_xdp;
 920
 921                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 922                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 923                  */
 924                 frame_sz = headroom ? PAGE_SIZE : truesize;
 925
 926                 /* This happens when rx buffer size is underestimated
 927                  * or headroom is not enough because of the buffer
 928                  * was refilled before XDP is set. This should only
 929                  * happen for the first several packets, so we don't
 930                  * care much about its performance.
 931                  */
 932                 if (unlikely(num_buf > 1 ||
 933                              headroom < virtnet_get_headroom(vi))) {
 934                         /* linearize data for XDP */
 935                         xdp_page = xdp_linearize_page(rq, &num_buf,
 936                                                       page, offset,
 937                                                       VIRTIO_XDP_HEADROOM,
 938                                                       &len);
 939                         frame_sz = PAGE_SIZE;
 940
 941                         if (!xdp_page)
 942                                 goto err_xdp;
 943                         offset = VIRTIO_XDP_HEADROOM;
 944                 } else {
 945                         xdp_page = page;
 946                 }
 947
 948                 /* Allow consuming headroom but reserve enough space to push
 949                  * the descriptor on if we get an XDP_TX return code.
 950                  */
 951                 data = page_address(xdp_page) + offset;
 952                 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
 953                 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
 954                                  VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
 955
 956                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 957                 stats->xdp_packets++;
 958
 959                 switch (act) {
 960                 case XDP_PASS:
 961                         metasize = xdp.data - xdp.data_meta;
 962
 963                         /* recalculate offset to account for any header
 964                          * adjustments and minus the metasize to copy the
 965                          * metadata in page_to_skb(). Note other cases do not
 966                          * build an skb and avoid using offset
 967                          */
 968                         offset = xdp.data - page_address(xdp_page) -
 969                                  vi->hdr_len - metasize;
 970
 971                         /* recalculate len if xdp.data, xdp.data_end or
 972                          * xdp.data_meta were adjusted
 973                          */
 974                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 975                         /* We can only create skb based on xdp_page. */
 976                         if (unlikely(xdp_page != page)) {
 977                                 rcu_read_unlock();
 978                                 put_page(page);
 979                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
 980                                                        len, PAGE_SIZE, false,
 981                                                        metasize, true);
 982                                 return head_skb;
 983                         }
 984                         break;
 985                 case XDP_TX:
 986                         stats->xdp_tx++;
 987                         xdpf = xdp_convert_buff_to_frame(&xdp);
 988                         if (unlikely(!xdpf))
 989                                 goto err_xdp;
 990                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 991                         if (unlikely(!err)) {
 992                                 xdp_return_frame_rx_napi(xdpf);
 993                         } else if (unlikely(err < 0)) {
 994                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 995                                 if (unlikely(xdp_page != page))
 996                                         put_page(xdp_page);
 997                                 goto err_xdp;
 998                         }
 999                         *xdp_xmit |= VIRTIO_XDP_TX;
1000                         if (unlikely(xdp_page != page))
1001                                 put_page(page);
1002                         rcu_read_unlock();
1003                         goto xdp_xmit;
1004                 case XDP_REDIRECT:
1005                         stats->xdp_redirects++;
1006                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
1007                         if (err) {
1008                                 if (unlikely(xdp_page != page))
1009                                         put_page(xdp_page);
1010                                 goto err_xdp;
1011                         }
1012                         *xdp_xmit |= VIRTIO_XDP_REDIR;
1013                         if (unlikely(xdp_page != page))
1014                                 put_page(page);
1015                         rcu_read_unlock();
1016                         goto xdp_xmit;
1017                 default:
1018                         bpf_warn_invalid_xdp_action(act);
1019                         fallthrough;
1020                 case XDP_ABORTED:
1021                         trace_xdp_exception(vi->dev, xdp_prog, act);
1022                         fallthrough;
1023                 case XDP_DROP:
1024                         if (unlikely(xdp_page != page))
1025                                 __free_pages(xdp_page, 0);
1026                         goto err_xdp;
1027                 }
1028         }
1029         rcu_read_unlock();
1030
1031         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1032                                metasize, !!headroom);
1033         curr_skb = head_skb;
1034
1035         if (unlikely(!curr_skb))
1036                 goto err_skb;
1037         while (--num_buf) {
1038                 int num_skb_frags;
1039
1040                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1041                 if (unlikely(!buf)) {
1042                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1043                                  dev->name, num_buf,
1044                                  virtio16_to_cpu(vi->vdev,
1045                                                  hdr->num_buffers));
1046                         dev->stats.rx_length_errors++;
1047                         goto err_buf;
1048                 }
1049
1050                 stats->bytes += len;
1051                 page = virt_to_head_page(buf);
1052
1053                 truesize = mergeable_ctx_to_truesize(ctx);
1054                 if (unlikely(len > truesize)) {
1055                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1056                                  dev->name, len, (unsigned long)ctx);
1057                         dev->stats.rx_length_errors++;
1058                         goto err_skb;
1059                 }
1060
1061                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1062                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1063                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1064
1065                         if (unlikely(!nskb))
1066                                 goto err_skb;
1067                         if (curr_skb == head_skb)
1068                                 skb_shinfo(curr_skb)->frag_list = nskb;
1069                         else
1070                                 curr_skb->next = nskb;
1071                         curr_skb = nskb;
1072                         head_skb->truesize += nskb->truesize;
1073                         num_skb_frags = 0;
1074                 }
1075                 if (curr_skb != head_skb) {
1076                         head_skb->data_len += len;
1077                         head_skb->len += len;
1078                         head_skb->truesize += truesize;
1079                 }
1080                 offset = buf - page_address(page);
1081                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1082                         put_page(page);
1083                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1084                                              len, truesize);
1085                 } else {
1086                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1087                                         offset, len, truesize);
1088                 }
1089         }
1090
1091         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1092         return head_skb;
1093
1094 err_xdp:
1095         rcu_read_unlock();
1096         stats->xdp_drops++;
1097 err_skb:
1098         put_page(page);
1099         while (num_buf-- > 1) {
1100                 buf = virtqueue_get_buf(rq->vq, &len);
1101                 if (unlikely(!buf)) {
1102                         pr_debug("%s: rx error: %d buffers missing\n",
1103                                  dev->name, num_buf);
1104                         dev->stats.rx_length_errors++;
1105                         break;
1106                 }
1107                 stats->bytes += len;
1108                 page = virt_to_head_page(buf);
1109                 put_page(page);
1110         }
1111 err_buf:
1112         stats->drops++;
1113         dev_kfree_skb(head_skb);
1114 xdp_xmit:
1115         return NULL;
1116 }
1117
1118 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1119                         void *buf, unsigned int len, void **ctx,
1120                         unsigned int *xdp_xmit,
1121                         struct virtnet_rq_stats *stats)
1122 {
1123         struct net_device *dev = vi->dev;
1124         struct sk_buff *skb;
1125         struct virtio_net_hdr_mrg_rxbuf *hdr;
1126
1127         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1128                 pr_debug("%s: short packet %i\n", dev->name, len);
1129                 dev->stats.rx_length_errors++;
1130                 if (vi->mergeable_rx_bufs) {
1131                         put_page(virt_to_head_page(buf));
1132                 } else if (vi->big_packets) {
1133                         give_pages(rq, buf);
1134                 } else {
1135                         put_page(virt_to_head_page(buf));
1136                 }
1137                 return;
1138         }
1139
1140         if (vi->mergeable_rx_bufs)
1141                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1142                                         stats);
1143         else if (vi->big_packets)
1144                 skb = receive_big(dev, vi, rq, buf, len, stats);
1145         else
1146                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1147
1148         if (unlikely(!skb))
1149                 return;
1150
1151         hdr = skb_vnet_hdr(skb);
1152
1153         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1154                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1155
1156         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1157                                   virtio_is_little_endian(vi->vdev))) {
1158                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1159                                      dev->name, hdr->hdr.gso_type,
1160                                      hdr->hdr.gso_size);
1161                 goto frame_err;
1162         }
1163
1164         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1165         skb->protocol = eth_type_trans(skb, dev);
1166         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1167                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1168
1169         napi_gro_receive(&rq->napi, skb);
1170         return;
1171
1172 frame_err:
1173         dev->stats.rx_frame_errors++;
1174         dev_kfree_skb(skb);
1175 }
1176
1177 /* Unlike mergeable buffers, all buffers are allocated to the
1178  * same size, except for the headroom. For this reason we do
1179  * not need to use  mergeable_len_to_ctx here - it is enough
1180  * to store the headroom as the context ignoring the truesize.
1181  */
1182 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1183                              gfp_t gfp)
1184 {
1185         struct page_frag *alloc_frag = &rq->alloc_frag;
1186         char *buf;
1187         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1188         void *ctx = (void *)(unsigned long)xdp_headroom;
1189         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1190         int err;
1191
1192         len = SKB_DATA_ALIGN(len) +
1193               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1194         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1195                 return -ENOMEM;
1196
1197         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1198         get_page(alloc_frag->page);
1199         alloc_frag->offset += len;
1200         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1201                     vi->hdr_len + GOOD_PACKET_LEN);
1202         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1203         if (err < 0)
1204                 put_page(virt_to_head_page(buf));
1205         return err;
1206 }
1207
1208 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1209                            gfp_t gfp)
1210 {
1211         struct page *first, *list = NULL;
1212         char *p;
1213         int i, err, offset;
1214
1215         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1216
1217         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1218         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1219                 first = get_a_page(rq, gfp);
1220                 if (!first) {
1221                         if (list)
1222                                 give_pages(rq, list);
1223                         return -ENOMEM;
1224                 }
1225                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1226
1227                 /* chain new page in list head to match sg */
1228                 first->private = (unsigned long)list;
1229                 list = first;
1230         }
1231
1232         first = get_a_page(rq, gfp);
1233         if (!first) {
1234                 give_pages(rq, list);
1235                 return -ENOMEM;
1236         }
1237         p = page_address(first);
1238
1239         /* rq->sg[0], rq->sg[1] share the same page */
1240         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1241         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1242
1243         /* rq->sg[1] for data packet, from offset */
1244         offset = sizeof(struct padded_vnet_hdr);
1245         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1246
1247         /* chain first in list head */
1248         first->private = (unsigned long)list;
1249         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1250                                   first, gfp);
1251         if (err < 0)
1252                 give_pages(rq, first);
1253
1254         return err;
1255 }
1256
1257 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1258                                           struct ewma_pkt_len *avg_pkt_len,
1259                                           unsigned int room)
1260 {
1261         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1262         unsigned int len;
1263
1264         if (room)
1265                 return PAGE_SIZE - room;
1266
1267         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1268                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1269
1270         return ALIGN(len, L1_CACHE_BYTES);
1271 }
1272
1273 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1274                                  struct receive_queue *rq, gfp_t gfp)
1275 {
1276         struct page_frag *alloc_frag = &rq->alloc_frag;
1277         unsigned int headroom = virtnet_get_headroom(vi);
1278         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1279         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1280         char *buf;
1281         void *ctx;
1282         int err;
1283         unsigned int len, hole;
1284
1285         /* Extra tailroom is needed to satisfy XDP's assumption. This
1286          * means rx frags coalescing won't work, but consider we've
1287          * disabled GSO for XDP, it won't be a big issue.
1288          */
1289         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1290         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1291                 return -ENOMEM;
1292
1293         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1294         buf += headroom; /* advance address leaving hole at front of pkt */
1295         get_page(alloc_frag->page);
1296         alloc_frag->offset += len + room;
1297         hole = alloc_frag->size - alloc_frag->offset;
1298         if (hole < len + room) {
1299                 /* To avoid internal fragmentation, if there is very likely not
1300                  * enough space for another buffer, add the remaining space to
1301                  * the current buffer.
1302                  */
1303                 len += hole;
1304                 alloc_frag->offset += hole;
1305         }
1306
1307         sg_init_one(rq->sg, buf, len);
1308         ctx = mergeable_len_to_ctx(len, headroom);
1309         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1310         if (err < 0)
1311                 put_page(virt_to_head_page(buf));
1312
1313         return err;
1314 }
1315
1316 /*
1317  * Returns false if we couldn't fill entirely (OOM).
1318  *
1319  * Normally run in the receive path, but can also be run from ndo_open
1320  * before we're receiving packets, or from refill_work which is
1321  * careful to disable receiving (using napi_disable).
1322  */
1323 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1324                           gfp_t gfp)
1325 {
1326         int err;
1327         bool oom;
1328
1329         do {
1330                 if (vi->mergeable_rx_bufs)
1331                         err = add_recvbuf_mergeable(vi, rq, gfp);
1332                 else if (vi->big_packets)
1333                         err = add_recvbuf_big(vi, rq, gfp);
1334                 else
1335                         err = add_recvbuf_small(vi, rq, gfp);
1336
1337                 oom = err == -ENOMEM;
1338                 if (err)
1339                         break;
1340         } while (rq->vq->num_free);
1341         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1342                 unsigned long flags;
1343
1344                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1345                 rq->stats.kicks++;
1346                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1347         }
1348
1349         return !oom;
1350 }
1351
1352 static void skb_recv_done(struct virtqueue *rvq)
1353 {
1354         struct virtnet_info *vi = rvq->vdev->priv;
1355         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1356
1357         virtqueue_napi_schedule(&rq->napi, rvq);
1358 }
1359
1360 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1361 {
1362         napi_enable(napi);
1363
1364         /* If all buffers were filled by other side before we napi_enabled, we
1365          * won't get another interrupt, so process any outstanding packets now.
1366          * Call local_bh_enable after to trigger softIRQ processing.
1367          */
1368         local_bh_disable();
1369         virtqueue_napi_schedule(napi, vq);
1370         local_bh_enable();
1371 }
1372
1373 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1374                                    struct virtqueue *vq,
1375                                    struct napi_struct *napi)
1376 {
1377         if (!napi->weight)
1378                 return;
1379
1380         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1381          * enable the feature if this is likely affine with the transmit path.
1382          */
1383         if (!vi->affinity_hint_set) {
1384                 napi->weight = 0;
1385                 return;
1386         }
1387
1388         return virtnet_napi_enable(vq, napi);
1389 }
1390
1391 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1392 {
1393         if (napi->weight)
1394                 napi_disable(napi);
1395 }
1396
1397 static void refill_work(struct work_struct *work)
1398 {
1399         struct virtnet_info *vi =
1400                 container_of(work, struct virtnet_info, refill.work);
1401         bool still_empty;
1402         int i;
1403
1404         for (i = 0; i < vi->curr_queue_pairs; i++) {
1405                 struct receive_queue *rq = &vi->rq[i];
1406
1407                 napi_disable(&rq->napi);
1408                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1409                 virtnet_napi_enable(rq->vq, &rq->napi);
1410
1411                 /* In theory, this can happen: if we don't get any buffers in
1412                  * we will *never* try to fill again.
1413                  */
1414                 if (still_empty)
1415                         schedule_delayed_work(&vi->refill, HZ/2);
1416         }
1417 }
1418
1419 static int virtnet_receive(struct receive_queue *rq, int budget,
1420                            unsigned int *xdp_xmit)
1421 {
1422         struct virtnet_info *vi = rq->vq->vdev->priv;
1423         struct virtnet_rq_stats stats = {};
1424         unsigned int len;
1425         void *buf;
1426         int i;
1427
1428         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1429                 void *ctx;
1430
1431                 while (stats.packets < budget &&
1432                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1433                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1434                         stats.packets++;
1435                 }
1436         } else {
1437                 while (stats.packets < budget &&
1438                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1439                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1440                         stats.packets++;
1441                 }
1442         }
1443
1444         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1445                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1446                         schedule_delayed_work(&vi->refill, 0);
1447         }
1448
1449         u64_stats_update_begin(&rq->stats.syncp);
1450         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1451                 size_t offset = virtnet_rq_stats_desc[i].offset;
1452                 u64 *item;
1453
1454                 item = (u64 *)((u8 *)&rq->stats + offset);
1455                 *item += *(u64 *)((u8 *)&stats + offset);
1456         }
1457         u64_stats_update_end(&rq->stats.syncp);
1458
1459         return stats.packets;
1460 }
1461
1462 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1463 {
1464         unsigned int len;
1465         unsigned int packets = 0;
1466         unsigned int bytes = 0;
1467         void *ptr;
1468
1469         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1470                 if (likely(!is_xdp_frame(ptr))) {
1471                         struct sk_buff *skb = ptr;
1472
1473                         pr_debug("Sent skb %p\n", skb);
1474
1475                         bytes += skb->len;
1476                         napi_consume_skb(skb, in_napi);
1477                 } else {
1478                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1479
1480                         bytes += frame->len;
1481                         xdp_return_frame(frame);
1482                 }
1483                 packets++;
1484         }
1485
1486         /* Avoid overhead when no packets have been processed
1487          * happens when called speculatively from start_xmit.
1488          */
1489         if (!packets)
1490                 return;
1491
1492         u64_stats_update_begin(&sq->stats.syncp);
1493         sq->stats.bytes += bytes;
1494         sq->stats.packets += packets;
1495         u64_stats_update_end(&sq->stats.syncp);
1496 }
1497
1498 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1499 {
1500         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1501                 return false;
1502         else if (q < vi->curr_queue_pairs)
1503                 return true;
1504         else
1505                 return false;
1506 }
1507
1508 static void virtnet_poll_cleantx(struct receive_queue *rq)
1509 {
1510         struct virtnet_info *vi = rq->vq->vdev->priv;
1511         unsigned int index = vq2rxq(rq->vq);
1512         struct send_queue *sq = &vi->sq[index];
1513         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1514
1515         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1516                 return;
1517
1518         if (__netif_tx_trylock(txq)) {
1519                 free_old_xmit_skbs(sq, true);
1520                 __netif_tx_unlock(txq);
1521         }
1522
1523         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1524                 netif_tx_wake_queue(txq);
1525 }
1526
1527 static int virtnet_poll(struct napi_struct *napi, int budget)
1528 {
1529         struct receive_queue *rq =
1530                 container_of(napi, struct receive_queue, napi);
1531         struct virtnet_info *vi = rq->vq->vdev->priv;
1532         struct send_queue *sq;
1533         unsigned int received;
1534         unsigned int xdp_xmit = 0;
1535
1536         virtnet_poll_cleantx(rq);
1537
1538         received = virtnet_receive(rq, budget, &xdp_xmit);
1539
1540         /* Out of packets? */
1541         if (received < budget)
1542                 virtqueue_napi_complete(napi, rq->vq, received);
1543
1544         if (xdp_xmit & VIRTIO_XDP_REDIR)
1545                 xdp_do_flush();
1546
1547         if (xdp_xmit & VIRTIO_XDP_TX) {
1548                 sq = virtnet_xdp_get_sq(vi);
1549                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1550                         u64_stats_update_begin(&sq->stats.syncp);
1551                         sq->stats.kicks++;
1552                         u64_stats_update_end(&sq->stats.syncp);
1553                 }
1554                 virtnet_xdp_put_sq(vi, sq);
1555         }
1556
1557         return received;
1558 }
1559
1560 static int virtnet_open(struct net_device *dev)
1561 {
1562         struct virtnet_info *vi = netdev_priv(dev);
1563         int i, err;
1564
1565         for (i = 0; i < vi->max_queue_pairs; i++) {
1566                 if (i < vi->curr_queue_pairs)
1567                         /* Make sure we have some buffers: if oom use wq. */
1568                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1569                                 schedule_delayed_work(&vi->refill, 0);
1570
1571                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1572                 if (err < 0)
1573                         return err;
1574
1575                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1576                                                  MEM_TYPE_PAGE_SHARED, NULL);
1577                 if (err < 0) {
1578                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1579                         return err;
1580                 }
1581
1582                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1583                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1584         }
1585
1586         return 0;
1587 }
1588
1589 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1590 {
1591         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1592         struct virtnet_info *vi = sq->vq->vdev->priv;
1593         unsigned int index = vq2txq(sq->vq);
1594         struct netdev_queue *txq;
1595
1596         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1597                 /* We don't need to enable cb for XDP */
1598                 napi_complete_done(napi, 0);
1599                 return 0;
1600         }
1601
1602         txq = netdev_get_tx_queue(vi->dev, index);
1603         __netif_tx_lock(txq, raw_smp_processor_id());
1604         free_old_xmit_skbs(sq, true);
1605         __netif_tx_unlock(txq);
1606
1607         virtqueue_napi_complete(napi, sq->vq, 0);
1608
1609         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1610                 netif_tx_wake_queue(txq);
1611
1612         return 0;
1613 }
1614
1615 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1616 {
1617         struct virtio_net_hdr_mrg_rxbuf *hdr;
1618         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1619         struct virtnet_info *vi = sq->vq->vdev->priv;
1620         int num_sg;
1621         unsigned hdr_len = vi->hdr_len;
1622         bool can_push;
1623
1624         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1625
1626         can_push = vi->any_header_sg &&
1627                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1628                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1629         /* Even if we can, don't push here yet as this would skew
1630          * csum_start offset below. */
1631         if (can_push)
1632                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1633         else
1634                 hdr = skb_vnet_hdr(skb);
1635
1636         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1637                                     virtio_is_little_endian(vi->vdev), false,
1638                                     0))
1639                 return -EPROTO;
1640
1641         if (vi->mergeable_rx_bufs)
1642                 hdr->num_buffers = 0;
1643
1644         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1645         if (can_push) {
1646                 __skb_push(skb, hdr_len);
1647                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1648                 if (unlikely(num_sg < 0))
1649                         return num_sg;
1650                 /* Pull header back to avoid skew in tx bytes calculations. */
1651                 __skb_pull(skb, hdr_len);
1652         } else {
1653                 sg_set_buf(sq->sg, hdr, hdr_len);
1654                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1655                 if (unlikely(num_sg < 0))
1656                         return num_sg;
1657                 num_sg++;
1658         }
1659         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1660 }
1661
1662 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1663 {
1664         struct virtnet_info *vi = netdev_priv(dev);
1665         int qnum = skb_get_queue_mapping(skb);
1666         struct send_queue *sq = &vi->sq[qnum];
1667         int err;
1668         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1669         bool kick = !netdev_xmit_more();
1670         bool use_napi = sq->napi.weight;
1671
1672         /* Free up any pending old buffers before queueing new ones. */
1673         free_old_xmit_skbs(sq, false);
1674
1675         if (use_napi && kick)
1676                 virtqueue_enable_cb_delayed(sq->vq);
1677
1678         /* timestamp packet in software */
1679         skb_tx_timestamp(skb);
1680
1681         /* Try to transmit */
1682         err = xmit_skb(sq, skb);
1683
1684         /* This should not happen! */
1685         if (unlikely(err)) {
1686                 dev->stats.tx_fifo_errors++;
1687                 if (net_ratelimit())
1688                         dev_warn(&dev->dev,
1689                                  "Unexpected TXQ (%d) queue failure: %d\n",
1690                                  qnum, err);
1691                 dev->stats.tx_dropped++;
1692                 dev_kfree_skb_any(skb);
1693                 return NETDEV_TX_OK;
1694         }
1695
1696         /* Don't wait up for transmitted skbs to be freed. */
1697         if (!use_napi) {
1698                 skb_orphan(skb);
1699                 nf_reset_ct(skb);
1700         }
1701
1702         /* If running out of space, stop queue to avoid getting packets that we
1703          * are then unable to transmit.
1704          * An alternative would be to force queuing layer to requeue the skb by
1705          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1706          * returned in a normal path of operation: it means that driver is not
1707          * maintaining the TX queue stop/start state properly, and causes
1708          * the stack to do a non-trivial amount of useless work.
1709          * Since most packets only take 1 or 2 ring slots, stopping the queue
1710          * early means 16 slots are typically wasted.
1711          */
1712         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1713                 netif_stop_subqueue(dev, qnum);
1714                 if (!use_napi &&
1715                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1716                         /* More just got used, free them then recheck. */
1717                         free_old_xmit_skbs(sq, false);
1718                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1719                                 netif_start_subqueue(dev, qnum);
1720                                 virtqueue_disable_cb(sq->vq);
1721                         }
1722                 }
1723         }
1724
1725         if (kick || netif_xmit_stopped(txq)) {
1726                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1727                         u64_stats_update_begin(&sq->stats.syncp);
1728                         sq->stats.kicks++;
1729                         u64_stats_update_end(&sq->stats.syncp);
1730                 }
1731         }
1732
1733         return NETDEV_TX_OK;
1734 }
1735
1736 /*
1737  * Send command via the control virtqueue and check status.  Commands
1738  * supported by the hypervisor, as indicated by feature bits, should
1739  * never fail unless improperly formatted.
1740  */
1741 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1742                                  struct scatterlist *out)
1743 {
1744         struct scatterlist *sgs[4], hdr, stat;
1745         unsigned out_num = 0, tmp;
1746
1747         /* Caller should know better */
1748         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1749
1750         vi->ctrl->status = ~0;
1751         vi->ctrl->hdr.class = class;
1752         vi->ctrl->hdr.cmd = cmd;
1753         /* Add header */
1754         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1755         sgs[out_num++] = &hdr;
1756
1757         if (out)
1758                 sgs[out_num++] = out;
1759
1760         /* Add return status. */
1761         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1762         sgs[out_num] = &stat;
1763
1764         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1765         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1766
1767         if (unlikely(!virtqueue_kick(vi->cvq)))
1768                 return vi->ctrl->status == VIRTIO_NET_OK;
1769
1770         /* Spin for a response, the kick causes an ioport write, trapping
1771          * into the hypervisor, so the request should be handled immediately.
1772          */
1773         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1774                !virtqueue_is_broken(vi->cvq))
1775                 cpu_relax();
1776
1777         return vi->ctrl->status == VIRTIO_NET_OK;
1778 }
1779
1780 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1781 {
1782         struct virtnet_info *vi = netdev_priv(dev);
1783         struct virtio_device *vdev = vi->vdev;
1784         int ret;
1785         struct sockaddr *addr;
1786         struct scatterlist sg;
1787
1788         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1789                 return -EOPNOTSUPP;
1790
1791         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1792         if (!addr)
1793                 return -ENOMEM;
1794
1795         ret = eth_prepare_mac_addr_change(dev, addr);
1796         if (ret)
1797                 goto out;
1798
1799         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1800                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1801                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1802                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1803                         dev_warn(&vdev->dev,
1804                                  "Failed to set mac address by vq command.\n");
1805                         ret = -EINVAL;
1806                         goto out;
1807                 }
1808         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1809                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1810                 unsigned int i;
1811
1812                 /* Naturally, this has an atomicity problem. */
1813                 for (i = 0; i < dev->addr_len; i++)
1814                         virtio_cwrite8(vdev,
1815                                        offsetof(struct virtio_net_config, mac) +
1816                                        i, addr->sa_data[i]);
1817         }
1818
1819         eth_commit_mac_addr_change(dev, p);
1820         ret = 0;
1821
1822 out:
1823         kfree(addr);
1824         return ret;
1825 }
1826
1827 static void virtnet_stats(struct net_device *dev,
1828                           struct rtnl_link_stats64 *tot)
1829 {
1830         struct virtnet_info *vi = netdev_priv(dev);
1831         unsigned int start;
1832         int i;
1833
1834         for (i = 0; i < vi->max_queue_pairs; i++) {
1835                 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1836                 struct receive_queue *rq = &vi->rq[i];
1837                 struct send_queue *sq = &vi->sq[i];
1838
1839                 do {
1840                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1841                         tpackets = sq->stats.packets;
1842                         tbytes   = sq->stats.bytes;
1843                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1844
1845                 do {
1846                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1847                         rpackets = rq->stats.packets;
1848                         rbytes   = rq->stats.bytes;
1849                         rdrops   = rq->stats.drops;
1850                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1851
1852                 tot->rx_packets += rpackets;
1853                 tot->tx_packets += tpackets;
1854                 tot->rx_bytes   += rbytes;
1855                 tot->tx_bytes   += tbytes;
1856                 tot->rx_dropped += rdrops;
1857         }
1858
1859         tot->tx_dropped = dev->stats.tx_dropped;
1860         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1861         tot->rx_length_errors = dev->stats.rx_length_errors;
1862         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1863 }
1864
1865 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1866 {
1867         rtnl_lock();
1868         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1869                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1870                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1871         rtnl_unlock();
1872 }
1873
1874 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1875 {
1876         struct scatterlist sg;
1877         struct net_device *dev = vi->dev;
1878
1879         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1880                 return 0;
1881
1882         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1883         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1884
1885         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1886                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1887                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1888                          queue_pairs);
1889                 return -EINVAL;
1890         } else {
1891                 vi->curr_queue_pairs = queue_pairs;
1892                 /* virtnet_open() will refill when device is going to up. */
1893                 if (dev->flags & IFF_UP)
1894                         schedule_delayed_work(&vi->refill, 0);
1895         }
1896
1897         return 0;
1898 }
1899
1900 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1901 {
1902         int err;
1903
1904         rtnl_lock();
1905         err = _virtnet_set_queues(vi, queue_pairs);
1906         rtnl_unlock();
1907         return err;
1908 }
1909
1910 static int virtnet_close(struct net_device *dev)
1911 {
1912         struct virtnet_info *vi = netdev_priv(dev);
1913         int i;
1914
1915         /* Make sure refill_work doesn't re-enable napi! */
1916         cancel_delayed_work_sync(&vi->refill);
1917
1918         for (i = 0; i < vi->max_queue_pairs; i++) {
1919                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1920                 napi_disable(&vi->rq[i].napi);
1921                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1922         }
1923
1924         return 0;
1925 }
1926
1927 static void virtnet_set_rx_mode(struct net_device *dev)
1928 {
1929         struct virtnet_info *vi = netdev_priv(dev);
1930         struct scatterlist sg[2];
1931         struct virtio_net_ctrl_mac *mac_data;
1932         struct netdev_hw_addr *ha;
1933         int uc_count;
1934         int mc_count;
1935         void *buf;
1936         int i;
1937
1938         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1939         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1940                 return;
1941
1942         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1943         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1944
1945         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1946
1947         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1948                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1949                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1950                          vi->ctrl->promisc ? "en" : "dis");
1951
1952         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1953
1954         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1955                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1956                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1957                          vi->ctrl->allmulti ? "en" : "dis");
1958
1959         uc_count = netdev_uc_count(dev);
1960         mc_count = netdev_mc_count(dev);
1961         /* MAC filter - use one buffer for both lists */
1962         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1963                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1964         mac_data = buf;
1965         if (!buf)
1966                 return;
1967
1968         sg_init_table(sg, 2);
1969
1970         /* Store the unicast list and count in the front of the buffer */
1971         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1972         i = 0;
1973         netdev_for_each_uc_addr(ha, dev)
1974                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1975
1976         sg_set_buf(&sg[0], mac_data,
1977                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1978
1979         /* multicast list and count fill the end */
1980         mac_data = (void *)&mac_data->macs[uc_count][0];
1981
1982         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1983         i = 0;
1984         netdev_for_each_mc_addr(ha, dev)
1985                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1986
1987         sg_set_buf(&sg[1], mac_data,
1988                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1989
1990         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1991                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1992                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1993
1994         kfree(buf);
1995 }
1996
1997 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1998                                    __be16 proto, u16 vid)
1999 {
2000         struct virtnet_info *vi = netdev_priv(dev);
2001         struct scatterlist sg;
2002
2003         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2004         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2005
2006         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2007                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2008                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2009         return 0;
2010 }
2011
2012 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2013                                     __be16 proto, u16 vid)
2014 {
2015         struct virtnet_info *vi = netdev_priv(dev);
2016         struct scatterlist sg;
2017
2018         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2019         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2020
2021         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2022                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2023                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2024         return 0;
2025 }
2026
2027 static void virtnet_clean_affinity(struct virtnet_info *vi)
2028 {
2029         int i;
2030
2031         if (vi->affinity_hint_set) {
2032                 for (i = 0; i < vi->max_queue_pairs; i++) {
2033                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
2034                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
2035                 }
2036
2037                 vi->affinity_hint_set = false;
2038         }
2039 }
2040
2041 static void virtnet_set_affinity(struct virtnet_info *vi)
2042 {
2043         cpumask_var_t mask;
2044         int stragglers;
2045         int group_size;
2046         int i, j, cpu;
2047         int num_cpu;
2048         int stride;
2049
2050         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2051                 virtnet_clean_affinity(vi);
2052                 return;
2053         }
2054
2055         num_cpu = num_online_cpus();
2056         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2057         stragglers = num_cpu >= vi->curr_queue_pairs ?
2058                         num_cpu % vi->curr_queue_pairs :
2059                         0;
2060         cpu = cpumask_next(-1, cpu_online_mask);
2061
2062         for (i = 0; i < vi->curr_queue_pairs; i++) {
2063                 group_size = stride + (i < stragglers ? 1 : 0);
2064
2065                 for (j = 0; j < group_size; j++) {
2066                         cpumask_set_cpu(cpu, mask);
2067                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2068                                                 nr_cpu_ids, false);
2069                 }
2070                 virtqueue_set_affinity(vi->rq[i].vq, mask);
2071                 virtqueue_set_affinity(vi->sq[i].vq, mask);
2072                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2073                 cpumask_clear(mask);
2074         }
2075
2076         vi->affinity_hint_set = true;
2077         free_cpumask_var(mask);
2078 }
2079
2080 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2081 {
2082         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2083                                                    node);
2084         virtnet_set_affinity(vi);
2085         return 0;
2086 }
2087
2088 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2089 {
2090         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2091                                                    node_dead);
2092         virtnet_set_affinity(vi);
2093         return 0;
2094 }
2095
2096 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2097 {
2098         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2099                                                    node);
2100
2101         virtnet_clean_affinity(vi);
2102         return 0;
2103 }
2104
2105 static enum cpuhp_state virtionet_online;
2106
2107 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2108 {
2109         int ret;
2110
2111         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2112         if (ret)
2113                 return ret;
2114         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2115                                                &vi->node_dead);
2116         if (!ret)
2117                 return ret;
2118         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2119         return ret;
2120 }
2121
2122 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2123 {
2124         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2125         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2126                                             &vi->node_dead);
2127 }
2128
2129 static void virtnet_get_ringparam(struct net_device *dev,
2130                                 struct ethtool_ringparam *ring)
2131 {
2132         struct virtnet_info *vi = netdev_priv(dev);
2133
2134         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2135         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2136         ring->rx_pending = ring->rx_max_pending;
2137         ring->tx_pending = ring->tx_max_pending;
2138 }
2139
2140
2141 static void virtnet_get_drvinfo(struct net_device *dev,
2142                                 struct ethtool_drvinfo *info)
2143 {
2144         struct virtnet_info *vi = netdev_priv(dev);
2145         struct virtio_device *vdev = vi->vdev;
2146
2147         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2148         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2149         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2150
2151 }
2152
2153 /* TODO: Eliminate OOO packets during switching */
2154 static int virtnet_set_channels(struct net_device *dev,
2155                                 struct ethtool_channels *channels)
2156 {
2157         struct virtnet_info *vi = netdev_priv(dev);
2158         u16 queue_pairs = channels->combined_count;
2159         int err;
2160
2161         /* We don't support separate rx/tx channels.
2162          * We don't allow setting 'other' channels.
2163          */
2164         if (channels->rx_count || channels->tx_count || channels->other_count)
2165                 return -EINVAL;
2166
2167         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2168                 return -EINVAL;
2169
2170         /* For now we don't support modifying channels while XDP is loaded
2171          * also when XDP is loaded all RX queues have XDP programs so we only
2172          * need to check a single RX queue.
2173          */
2174         if (vi->rq[0].xdp_prog)
2175                 return -EINVAL;
2176
2177         get_online_cpus();
2178         err = _virtnet_set_queues(vi, queue_pairs);
2179         if (err) {
2180                 put_online_cpus();
2181                 goto err;
2182         }
2183         virtnet_set_affinity(vi);
2184         put_online_cpus();
2185
2186         netif_set_real_num_tx_queues(dev, queue_pairs);
2187         netif_set_real_num_rx_queues(dev, queue_pairs);
2188  err:
2189         return err;
2190 }
2191
2192 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2193 {
2194         struct virtnet_info *vi = netdev_priv(dev);
2195         unsigned int i, j;
2196         u8 *p = data;
2197
2198         switch (stringset) {
2199         case ETH_SS_STATS:
2200                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2201                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2202                                 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2203                                                 virtnet_rq_stats_desc[j].desc);
2204                 }
2205
2206                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2207                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2208                                 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2209                                                 virtnet_sq_stats_desc[j].desc);
2210                 }
2211                 break;
2212         }
2213 }
2214
2215 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2216 {
2217         struct virtnet_info *vi = netdev_priv(dev);
2218
2219         switch (sset) {
2220         case ETH_SS_STATS:
2221                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2222                                                VIRTNET_SQ_STATS_LEN);
2223         default:
2224                 return -EOPNOTSUPP;
2225         }
2226 }
2227
2228 static void virtnet_get_ethtool_stats(struct net_device *dev,
2229                                       struct ethtool_stats *stats, u64 *data)
2230 {
2231         struct virtnet_info *vi = netdev_priv(dev);
2232         unsigned int idx = 0, start, i, j;
2233         const u8 *stats_base;
2234         size_t offset;
2235
2236         for (i = 0; i < vi->curr_queue_pairs; i++) {
2237                 struct receive_queue *rq = &vi->rq[i];
2238
2239                 stats_base = (u8 *)&rq->stats;
2240                 do {
2241                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2242                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2243                                 offset = virtnet_rq_stats_desc[j].offset;
2244                                 data[idx + j] = *(u64 *)(stats_base + offset);
2245                         }
2246                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2247                 idx += VIRTNET_RQ_STATS_LEN;
2248         }
2249
2250         for (i = 0; i < vi->curr_queue_pairs; i++) {
2251                 struct send_queue *sq = &vi->sq[i];
2252
2253                 stats_base = (u8 *)&sq->stats;
2254                 do {
2255                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2256                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2257                                 offset = virtnet_sq_stats_desc[j].offset;
2258                                 data[idx + j] = *(u64 *)(stats_base + offset);
2259                         }
2260                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2261                 idx += VIRTNET_SQ_STATS_LEN;
2262         }
2263 }
2264
2265 static void virtnet_get_channels(struct net_device *dev,
2266                                  struct ethtool_channels *channels)
2267 {
2268         struct virtnet_info *vi = netdev_priv(dev);
2269
2270         channels->combined_count = vi->curr_queue_pairs;
2271         channels->max_combined = vi->max_queue_pairs;
2272         channels->max_other = 0;
2273         channels->rx_count = 0;
2274         channels->tx_count = 0;
2275         channels->other_count = 0;
2276 }
2277
2278 static int virtnet_set_link_ksettings(struct net_device *dev,
2279                                       const struct ethtool_link_ksettings *cmd)
2280 {
2281         struct virtnet_info *vi = netdev_priv(dev);
2282
2283         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2284                                                   &vi->speed, &vi->duplex);
2285 }
2286
2287 static int virtnet_get_link_ksettings(struct net_device *dev,
2288                                       struct ethtool_link_ksettings *cmd)
2289 {
2290         struct virtnet_info *vi = netdev_priv(dev);
2291
2292         cmd->base.speed = vi->speed;
2293         cmd->base.duplex = vi->duplex;
2294         cmd->base.port = PORT_OTHER;
2295
2296         return 0;
2297 }
2298
2299 static int virtnet_set_coalesce(struct net_device *dev,
2300                                 struct ethtool_coalesce *ec)
2301 {
2302         struct virtnet_info *vi = netdev_priv(dev);
2303         int i, napi_weight;
2304
2305         if (ec->tx_max_coalesced_frames > 1 ||
2306             ec->rx_max_coalesced_frames != 1)
2307                 return -EINVAL;
2308
2309         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2310         if (napi_weight ^ vi->sq[0].napi.weight) {
2311                 if (dev->flags & IFF_UP)
2312                         return -EBUSY;
2313                 for (i = 0; i < vi->max_queue_pairs; i++)
2314                         vi->sq[i].napi.weight = napi_weight;
2315         }
2316
2317         return 0;
2318 }
2319
2320 static int virtnet_get_coalesce(struct net_device *dev,
2321                                 struct ethtool_coalesce *ec)
2322 {
2323         struct ethtool_coalesce ec_default = {
2324                 .cmd = ETHTOOL_GCOALESCE,
2325                 .rx_max_coalesced_frames = 1,
2326         };
2327         struct virtnet_info *vi = netdev_priv(dev);
2328
2329         memcpy(ec, &ec_default, sizeof(ec_default));
2330
2331         if (vi->sq[0].napi.weight)
2332                 ec->tx_max_coalesced_frames = 1;
2333
2334         return 0;
2335 }
2336
2337 static void virtnet_init_settings(struct net_device *dev)
2338 {
2339         struct virtnet_info *vi = netdev_priv(dev);
2340
2341         vi->speed = SPEED_UNKNOWN;
2342         vi->duplex = DUPLEX_UNKNOWN;
2343 }
2344
2345 static void virtnet_update_settings(struct virtnet_info *vi)
2346 {
2347         u32 speed;
2348         u8 duplex;
2349
2350         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2351                 return;
2352
2353         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2354
2355         if (ethtool_validate_speed(speed))
2356                 vi->speed = speed;
2357
2358         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2359
2360         if (ethtool_validate_duplex(duplex))
2361                 vi->duplex = duplex;
2362 }
2363
2364 static const struct ethtool_ops virtnet_ethtool_ops = {
2365         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2366         .get_drvinfo = virtnet_get_drvinfo,
2367         .get_link = ethtool_op_get_link,
2368         .get_ringparam = virtnet_get_ringparam,
2369         .get_strings = virtnet_get_strings,
2370         .get_sset_count = virtnet_get_sset_count,
2371         .get_ethtool_stats = virtnet_get_ethtool_stats,
2372         .set_channels = virtnet_set_channels,
2373         .get_channels = virtnet_get_channels,
2374         .get_ts_info = ethtool_op_get_ts_info,
2375         .get_link_ksettings = virtnet_get_link_ksettings,
2376         .set_link_ksettings = virtnet_set_link_ksettings,
2377         .set_coalesce = virtnet_set_coalesce,
2378         .get_coalesce = virtnet_get_coalesce,
2379 };
2380
2381 static void virtnet_freeze_down(struct virtio_device *vdev)
2382 {
2383         struct virtnet_info *vi = vdev->priv;
2384         int i;
2385
2386         /* Make sure no work handler is accessing the device */
2387         flush_work(&vi->config_work);
2388
2389         netif_tx_lock_bh(vi->dev);
2390         netif_device_detach(vi->dev);
2391         netif_tx_unlock_bh(vi->dev);
2392         cancel_delayed_work_sync(&vi->refill);
2393
2394         if (netif_running(vi->dev)) {
2395                 for (i = 0; i < vi->max_queue_pairs; i++) {
2396                         napi_disable(&vi->rq[i].napi);
2397                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2398                 }
2399         }
2400 }
2401
2402 static int init_vqs(struct virtnet_info *vi);
2403
2404 static int virtnet_restore_up(struct virtio_device *vdev)
2405 {
2406         struct virtnet_info *vi = vdev->priv;
2407         int err, i;
2408
2409         err = init_vqs(vi);
2410         if (err)
2411                 return err;
2412
2413         virtio_device_ready(vdev);
2414
2415         if (netif_running(vi->dev)) {
2416                 for (i = 0; i < vi->curr_queue_pairs; i++)
2417                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2418                                 schedule_delayed_work(&vi->refill, 0);
2419
2420                 for (i = 0; i < vi->max_queue_pairs; i++) {
2421                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2422                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2423                                                &vi->sq[i].napi);
2424                 }
2425         }
2426
2427         netif_tx_lock_bh(vi->dev);
2428         netif_device_attach(vi->dev);
2429         netif_tx_unlock_bh(vi->dev);
2430         return err;
2431 }
2432
2433 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2434 {
2435         struct scatterlist sg;
2436         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2437
2438         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2439
2440         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2441                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2442                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2443                 return -EINVAL;
2444         }
2445
2446         return 0;
2447 }
2448
2449 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2450 {
2451         u64 offloads = 0;
2452
2453         if (!vi->guest_offloads)
2454                 return 0;
2455
2456         return virtnet_set_guest_offloads(vi, offloads);
2457 }
2458
2459 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2460 {
2461         u64 offloads = vi->guest_offloads;
2462
2463         if (!vi->guest_offloads)
2464                 return 0;
2465
2466         return virtnet_set_guest_offloads(vi, offloads);
2467 }
2468
2469 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2470                            struct netlink_ext_ack *extack)
2471 {
2472         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2473         struct virtnet_info *vi = netdev_priv(dev);
2474         struct bpf_prog *old_prog;
2475         u16 xdp_qp = 0, curr_qp;
2476         int i, err;
2477
2478         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2479             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2480                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2481                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2482                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2483                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2484                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2485                 return -EOPNOTSUPP;
2486         }
2487
2488         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2489                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2490                 return -EINVAL;
2491         }
2492
2493         if (dev->mtu > max_sz) {
2494                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2495                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2496                 return -EINVAL;
2497         }
2498
2499         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2500         if (prog)
2501                 xdp_qp = nr_cpu_ids;
2502
2503         /* XDP requires extra queues for XDP_TX */
2504         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2505                 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",
2506                             curr_qp + xdp_qp, vi->max_queue_pairs);
2507                 xdp_qp = 0;
2508         }
2509
2510         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2511         if (!prog && !old_prog)
2512                 return 0;
2513
2514         if (prog)
2515                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2516
2517         /* Make sure NAPI is not using any XDP TX queues for RX. */
2518         if (netif_running(dev)) {
2519                 for (i = 0; i < vi->max_queue_pairs; i++) {
2520                         napi_disable(&vi->rq[i].napi);
2521                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2522                 }
2523         }
2524
2525         if (!prog) {
2526                 for (i = 0; i < vi->max_queue_pairs; i++) {
2527                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2528                         if (i == 0)
2529                                 virtnet_restore_guest_offloads(vi);
2530                 }
2531                 synchronize_net();
2532         }
2533
2534         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2535         if (err)
2536                 goto err;
2537         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2538         vi->xdp_queue_pairs = xdp_qp;
2539
2540         if (prog) {
2541                 vi->xdp_enabled = true;
2542                 for (i = 0; i < vi->max_queue_pairs; i++) {
2543                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2544                         if (i == 0 && !old_prog)
2545                                 virtnet_clear_guest_offloads(vi);
2546                 }
2547         } else {
2548                 vi->xdp_enabled = false;
2549         }
2550
2551         for (i = 0; i < vi->max_queue_pairs; i++) {
2552                 if (old_prog)
2553                         bpf_prog_put(old_prog);
2554                 if (netif_running(dev)) {
2555                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2556                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2557                                                &vi->sq[i].napi);
2558                 }
2559         }
2560
2561         return 0;
2562
2563 err:
2564         if (!prog) {
2565                 virtnet_clear_guest_offloads(vi);
2566                 for (i = 0; i < vi->max_queue_pairs; i++)
2567                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2568         }
2569
2570         if (netif_running(dev)) {
2571                 for (i = 0; i < vi->max_queue_pairs; i++) {
2572                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2573                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2574                                                &vi->sq[i].napi);
2575                 }
2576         }
2577         if (prog)
2578                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2579         return err;
2580 }
2581
2582 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2583 {
2584         switch (xdp->command) {
2585         case XDP_SETUP_PROG:
2586                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2587         default:
2588                 return -EINVAL;
2589         }
2590 }
2591
2592 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2593                                       size_t len)
2594 {
2595         struct virtnet_info *vi = netdev_priv(dev);
2596         int ret;
2597
2598         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2599                 return -EOPNOTSUPP;
2600
2601         ret = snprintf(buf, len, "sby");
2602         if (ret >= len)
2603                 return -EOPNOTSUPP;
2604
2605         return 0;
2606 }
2607
2608 static int virtnet_set_features(struct net_device *dev,
2609                                 netdev_features_t features)
2610 {
2611         struct virtnet_info *vi = netdev_priv(dev);
2612         u64 offloads;
2613         int err;
2614
2615         if ((dev->features ^ features) & NETIF_F_LRO) {
2616                 if (vi->xdp_enabled)
2617                         return -EBUSY;
2618
2619                 if (features & NETIF_F_LRO)
2620                         offloads = vi->guest_offloads_capable;
2621                 else
2622                         offloads = vi->guest_offloads_capable &
2623                                    ~GUEST_OFFLOAD_LRO_MASK;
2624
2625                 err = virtnet_set_guest_offloads(vi, offloads);
2626                 if (err)
2627                         return err;
2628                 vi->guest_offloads = offloads;
2629         }
2630
2631         return 0;
2632 }
2633
2634 static const struct net_device_ops virtnet_netdev = {
2635         .ndo_open            = virtnet_open,
2636         .ndo_stop            = virtnet_close,
2637         .ndo_start_xmit      = start_xmit,
2638         .ndo_validate_addr   = eth_validate_addr,
2639         .ndo_set_mac_address = virtnet_set_mac_address,
2640         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2641         .ndo_get_stats64     = virtnet_stats,
2642         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2643         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2644         .ndo_bpf                = virtnet_xdp,
2645         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2646         .ndo_features_check     = passthru_features_check,
2647         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2648         .ndo_set_features       = virtnet_set_features,
2649 };
2650
2651 static void virtnet_config_changed_work(struct work_struct *work)
2652 {
2653         struct virtnet_info *vi =
2654                 container_of(work, struct virtnet_info, config_work);
2655         u16 v;
2656
2657         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2658                                  struct virtio_net_config, status, &v) < 0)
2659                 return;
2660
2661         if (v & VIRTIO_NET_S_ANNOUNCE) {
2662                 netdev_notify_peers(vi->dev);
2663                 virtnet_ack_link_announce(vi);
2664         }
2665
2666         /* Ignore unknown (future) status bits */
2667         v &= VIRTIO_NET_S_LINK_UP;
2668
2669         if (vi->status == v)
2670                 return;
2671
2672         vi->status = v;
2673
2674         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2675                 virtnet_update_settings(vi);
2676                 netif_carrier_on(vi->dev);
2677                 netif_tx_wake_all_queues(vi->dev);
2678         } else {
2679                 netif_carrier_off(vi->dev);
2680                 netif_tx_stop_all_queues(vi->dev);
2681         }
2682 }
2683
2684 static void virtnet_config_changed(struct virtio_device *vdev)
2685 {
2686         struct virtnet_info *vi = vdev->priv;
2687
2688         schedule_work(&vi->config_work);
2689 }
2690
2691 static void virtnet_free_queues(struct virtnet_info *vi)
2692 {
2693         int i;
2694
2695         for (i = 0; i < vi->max_queue_pairs; i++) {
2696                 __netif_napi_del(&vi->rq[i].napi);
2697                 __netif_napi_del(&vi->sq[i].napi);
2698         }
2699
2700         /* We called __netif_napi_del(),
2701          * we need to respect an RCU grace period before freeing vi->rq
2702          */
2703         synchronize_net();
2704
2705         kfree(vi->rq);
2706         kfree(vi->sq);
2707         kfree(vi->ctrl);
2708 }
2709
2710 static void _free_receive_bufs(struct virtnet_info *vi)
2711 {
2712         struct bpf_prog *old_prog;
2713         int i;
2714
2715         for (i = 0; i < vi->max_queue_pairs; i++) {
2716                 while (vi->rq[i].pages)
2717                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2718
2719                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2720                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2721                 if (old_prog)
2722                         bpf_prog_put(old_prog);
2723         }
2724 }
2725
2726 static void free_receive_bufs(struct virtnet_info *vi)
2727 {
2728         rtnl_lock();
2729         _free_receive_bufs(vi);
2730         rtnl_unlock();
2731 }
2732
2733 static void free_receive_page_frags(struct virtnet_info *vi)
2734 {
2735         int i;
2736         for (i = 0; i < vi->max_queue_pairs; i++)
2737                 if (vi->rq[i].alloc_frag.page)
2738                         put_page(vi->rq[i].alloc_frag.page);
2739 }
2740
2741 static void free_unused_bufs(struct virtnet_info *vi)
2742 {
2743         void *buf;
2744         int i;
2745
2746         for (i = 0; i < vi->max_queue_pairs; i++) {
2747                 struct virtqueue *vq = vi->sq[i].vq;
2748                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2749                         if (!is_xdp_frame(buf))
2750                                 dev_kfree_skb(buf);
2751                         else
2752                                 xdp_return_frame(ptr_to_xdp(buf));
2753                 }
2754         }
2755
2756         for (i = 0; i < vi->max_queue_pairs; i++) {
2757                 struct virtqueue *vq = vi->rq[i].vq;
2758
2759                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2760                         if (vi->mergeable_rx_bufs) {
2761                                 put_page(virt_to_head_page(buf));
2762                         } else if (vi->big_packets) {
2763                                 give_pages(&vi->rq[i], buf);
2764                         } else {
2765                                 put_page(virt_to_head_page(buf));
2766                         }
2767                 }
2768         }
2769 }
2770
2771 static void virtnet_del_vqs(struct virtnet_info *vi)
2772 {
2773         struct virtio_device *vdev = vi->vdev;
2774
2775         virtnet_clean_affinity(vi);
2776
2777         vdev->config->del_vqs(vdev);
2778
2779         virtnet_free_queues(vi);
2780 }
2781
2782 /* How large should a single buffer be so a queue full of these can fit at
2783  * least one full packet?
2784  * Logic below assumes the mergeable buffer header is used.
2785  */
2786 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2787 {
2788         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2789         unsigned int rq_size = virtqueue_get_vring_size(vq);
2790         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2791         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2792         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2793
2794         return max(max(min_buf_len, hdr_len) - hdr_len,
2795                    (unsigned int)GOOD_PACKET_LEN);
2796 }
2797
2798 static int virtnet_find_vqs(struct virtnet_info *vi)
2799 {
2800         vq_callback_t **callbacks;
2801         struct virtqueue **vqs;
2802         int ret = -ENOMEM;
2803         int i, total_vqs;
2804         const char **names;
2805         bool *ctx;
2806
2807         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2808          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2809          * possible control vq.
2810          */
2811         total_vqs = vi->max_queue_pairs * 2 +
2812                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2813
2814         /* Allocate space for find_vqs parameters */
2815         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2816         if (!vqs)
2817                 goto err_vq;
2818         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2819         if (!callbacks)
2820                 goto err_callback;
2821         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2822         if (!names)
2823                 goto err_names;
2824         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2825                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2826                 if (!ctx)
2827                         goto err_ctx;
2828         } else {
2829                 ctx = NULL;
2830         }
2831
2832         /* Parameters for control virtqueue, if any */
2833         if (vi->has_cvq) {
2834                 callbacks[total_vqs - 1] = NULL;
2835                 names[total_vqs - 1] = "control";
2836         }
2837
2838         /* Allocate/initialize parameters for send/receive virtqueues */
2839         for (i = 0; i < vi->max_queue_pairs; i++) {
2840                 callbacks[rxq2vq(i)] = skb_recv_done;
2841                 callbacks[txq2vq(i)] = skb_xmit_done;
2842                 sprintf(vi->rq[i].name, "input.%d", i);
2843                 sprintf(vi->sq[i].name, "output.%d", i);
2844                 names[rxq2vq(i)] = vi->rq[i].name;
2845                 names[txq2vq(i)] = vi->sq[i].name;
2846                 if (ctx)
2847                         ctx[rxq2vq(i)] = true;
2848         }
2849
2850         ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
2851                                   names, ctx, NULL);
2852         if (ret)
2853                 goto err_find;
2854
2855         if (vi->has_cvq) {
2856                 vi->cvq = vqs[total_vqs - 1];
2857                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2858                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2859         }
2860
2861         for (i = 0; i < vi->max_queue_pairs; i++) {
2862                 vi->rq[i].vq = vqs[rxq2vq(i)];
2863                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2864                 vi->sq[i].vq = vqs[txq2vq(i)];
2865         }
2866
2867         /* run here: ret == 0. */
2868
2869
2870 err_find:
2871         kfree(ctx);
2872 err_ctx:
2873         kfree(names);
2874 err_names:
2875         kfree(callbacks);
2876 err_callback:
2877         kfree(vqs);
2878 err_vq:
2879         return ret;
2880 }
2881
2882 static int virtnet_alloc_queues(struct virtnet_info *vi)
2883 {
2884         int i;
2885
2886         if (vi->has_cvq) {
2887                 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2888                 if (!vi->ctrl)
2889                         goto err_ctrl;
2890         } else {
2891                 vi->ctrl = NULL;
2892         }
2893         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2894         if (!vi->sq)
2895                 goto err_sq;
2896         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2897         if (!vi->rq)
2898                 goto err_rq;
2899
2900         INIT_DELAYED_WORK(&vi->refill, refill_work);
2901         for (i = 0; i < vi->max_queue_pairs; i++) {
2902                 vi->rq[i].pages = NULL;
2903                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2904                                napi_weight);
2905                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2906                                   napi_tx ? napi_weight : 0);
2907
2908                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2909                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2910                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2911
2912                 u64_stats_init(&vi->rq[i].stats.syncp);
2913                 u64_stats_init(&vi->sq[i].stats.syncp);
2914         }
2915
2916         return 0;
2917
2918 err_rq:
2919         kfree(vi->sq);
2920 err_sq:
2921         kfree(vi->ctrl);
2922 err_ctrl:
2923         return -ENOMEM;
2924 }
2925
2926 static int init_vqs(struct virtnet_info *vi)
2927 {
2928         int ret;
2929
2930         /* Allocate send & receive queues */
2931         ret = virtnet_alloc_queues(vi);
2932         if (ret)
2933                 goto err;
2934
2935         ret = virtnet_find_vqs(vi);
2936         if (ret)
2937                 goto err_free;
2938
2939         get_online_cpus();
2940         virtnet_set_affinity(vi);
2941         put_online_cpus();
2942
2943         return 0;
2944
2945 err_free:
2946         virtnet_free_queues(vi);
2947 err:
2948         return ret;
2949 }
2950
2951 #ifdef CONFIG_SYSFS
2952 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2953                 char *buf)
2954 {
2955         struct virtnet_info *vi = netdev_priv(queue->dev);
2956         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2957         unsigned int headroom = virtnet_get_headroom(vi);
2958         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2959         struct ewma_pkt_len *avg;
2960
2961         BUG_ON(queue_index >= vi->max_queue_pairs);
2962         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2963         return sprintf(buf, "%u\n",
2964                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
2965                                        SKB_DATA_ALIGN(headroom + tailroom)));
2966 }
2967
2968 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2969         __ATTR_RO(mergeable_rx_buffer_size);
2970
2971 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2972         &mergeable_rx_buffer_size_attribute.attr,
2973         NULL
2974 };
2975
2976 static const struct attribute_group virtio_net_mrg_rx_group = {
2977         .name = "virtio_net",
2978         .attrs = virtio_net_mrg_rx_attrs
2979 };
2980 #endif
2981
2982 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2983                                     unsigned int fbit,
2984                                     const char *fname, const char *dname)
2985 {
2986         if (!virtio_has_feature(vdev, fbit))
2987                 return false;
2988
2989         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2990                 fname, dname);
2991
2992         return true;
2993 }
2994
2995 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2996         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2997
2998 static bool virtnet_validate_features(struct virtio_device *vdev)
2999 {
3000         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3001             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3002                              "VIRTIO_NET_F_CTRL_VQ") ||
3003              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3004                              "VIRTIO_NET_F_CTRL_VQ") ||
3005              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3006                              "VIRTIO_NET_F_CTRL_VQ") ||
3007              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3008              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3009                              "VIRTIO_NET_F_CTRL_VQ"))) {
3010                 return false;
3011         }
3012
3013         return true;
3014 }
3015
3016 #define MIN_MTU ETH_MIN_MTU
3017 #define MAX_MTU ETH_MAX_MTU
3018
3019 static int virtnet_validate(struct virtio_device *vdev)
3020 {
3021         if (!vdev->config->get) {
3022                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3023                         __func__);
3024                 return -EINVAL;
3025         }
3026
3027         if (!virtnet_validate_features(vdev))
3028                 return -EINVAL;
3029
3030         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3031                 int mtu = virtio_cread16(vdev,
3032                                          offsetof(struct virtio_net_config,
3033                                                   mtu));
3034                 if (mtu < MIN_MTU)
3035                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3036         }
3037
3038         return 0;
3039 }
3040
3041 static int virtnet_probe(struct virtio_device *vdev)
3042 {
3043         int i, err = -ENOMEM;
3044         struct net_device *dev;
3045         struct virtnet_info *vi;
3046         u16 max_queue_pairs;
3047         int mtu;
3048
3049         /* Find if host supports multiqueue virtio_net device */
3050         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3051                                    struct virtio_net_config,
3052                                    max_virtqueue_pairs, &max_queue_pairs);
3053
3054         /* We need at least 2 queue's */
3055         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3056             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3057             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3058                 max_queue_pairs = 1;
3059
3060         /* Allocate ourselves a network device with room for our info */
3061         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3062         if (!dev)
3063                 return -ENOMEM;
3064
3065         /* Set up network device as normal. */
3066         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3067                            IFF_TX_SKB_NO_LINEAR;
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_LRO;
3104         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3105                 dev->hw_features |= NETIF_F_LRO;
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
3129         /* If we can receive ANY GSO packets, we must allocate large ones. */
3130         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3131             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3132             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3133             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3134                 vi->big_packets = true;
3135
3136         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3137                 vi->mergeable_rx_bufs = true;
3138
3139         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3140             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3141                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3142         else
3143                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3144
3145         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3146             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3147                 vi->any_header_sg = true;
3148
3149         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3150                 vi->has_cvq = true;
3151
3152         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3153                 mtu = virtio_cread16(vdev,
3154                                      offsetof(struct virtio_net_config,
3155                                               mtu));
3156                 if (mtu < dev->min_mtu) {
3157                         /* Should never trigger: MTU was previously validated
3158                          * in virtnet_validate.
3159                          */
3160                         dev_err(&vdev->dev,
3161                                 "device MTU appears to have changed it is now %d < %d",
3162                                 mtu, dev->min_mtu);
3163                         err = -EINVAL;
3164                         goto free;
3165                 }
3166
3167                 dev->mtu = mtu;
3168                 dev->max_mtu = mtu;
3169
3170                 /* TODO: size buffers correctly in this case. */
3171                 if (dev->mtu > ETH_DATA_LEN)
3172                         vi->big_packets = true;
3173         }
3174
3175         if (vi->any_header_sg)
3176                 dev->needed_headroom = vi->hdr_len;
3177
3178         /* Enable multiqueue by default */
3179         if (num_online_cpus() >= max_queue_pairs)
3180                 vi->curr_queue_pairs = max_queue_pairs;
3181         else
3182                 vi->curr_queue_pairs = num_online_cpus();
3183         vi->max_queue_pairs = max_queue_pairs;
3184
3185         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3186         err = init_vqs(vi);
3187         if (err)
3188                 goto free;
3189
3190 #ifdef CONFIG_SYSFS
3191         if (vi->mergeable_rx_bufs)
3192                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3193 #endif
3194         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3195         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3196
3197         virtnet_init_settings(dev);
3198
3199         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3200                 vi->failover = net_failover_create(vi->dev);
3201                 if (IS_ERR(vi->failover)) {
3202                         err = PTR_ERR(vi->failover);
3203                         goto free_vqs;
3204                 }
3205         }
3206
3207         err = register_netdev(dev);
3208         if (err) {
3209                 pr_debug("virtio_net: registering device failed\n");
3210                 goto free_failover;
3211         }
3212
3213         virtio_device_ready(vdev);
3214
3215         err = virtnet_cpu_notif_add(vi);
3216         if (err) {
3217                 pr_debug("virtio_net: registering cpu notifier failed\n");
3218                 goto free_unregister_netdev;
3219         }
3220
3221         virtnet_set_queues(vi, vi->curr_queue_pairs);
3222
3223         /* Assume link up if device can't report link status,
3224            otherwise get link status from config. */
3225         netif_carrier_off(dev);
3226         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3227                 schedule_work(&vi->config_work);
3228         } else {
3229                 vi->status = VIRTIO_NET_S_LINK_UP;
3230                 virtnet_update_settings(vi);
3231                 netif_carrier_on(dev);
3232         }
3233
3234         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3235                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3236                         set_bit(guest_offloads[i], &vi->guest_offloads);
3237         vi->guest_offloads_capable = vi->guest_offloads;
3238
3239         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3240                  dev->name, max_queue_pairs);
3241
3242         return 0;
3243
3244 free_unregister_netdev:
3245         vi->vdev->config->reset(vdev);
3246
3247         unregister_netdev(dev);
3248 free_failover:
3249         net_failover_destroy(vi->failover);
3250 free_vqs:
3251         cancel_delayed_work_sync(&vi->refill);
3252         free_receive_page_frags(vi);
3253         virtnet_del_vqs(vi);
3254 free:
3255         free_netdev(dev);
3256         return err;
3257 }
3258
3259 static void remove_vq_common(struct virtnet_info *vi)
3260 {
3261         vi->vdev->config->reset(vi->vdev);
3262
3263         /* Free unused buffers in both send and recv, if any. */
3264         free_unused_bufs(vi);
3265
3266         free_receive_bufs(vi);
3267
3268         free_receive_page_frags(vi);
3269
3270         virtnet_del_vqs(vi);
3271 }
3272
3273 static void virtnet_remove(struct virtio_device *vdev)
3274 {
3275         struct virtnet_info *vi = vdev->priv;
3276
3277         virtnet_cpu_notif_remove(vi);
3278
3279         /* Make sure no work handler is accessing the device. */
3280         flush_work(&vi->config_work);
3281
3282         unregister_netdev(vi->dev);
3283
3284         net_failover_destroy(vi->failover);
3285
3286         remove_vq_common(vi);
3287
3288         free_netdev(vi->dev);
3289 }
3290
3291 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3292 {
3293         struct virtnet_info *vi = vdev->priv;
3294
3295         virtnet_cpu_notif_remove(vi);
3296         virtnet_freeze_down(vdev);
3297         remove_vq_common(vi);
3298
3299         return 0;
3300 }
3301
3302 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3303 {
3304         struct virtnet_info *vi = vdev->priv;
3305         int err;
3306
3307         err = virtnet_restore_up(vdev);
3308         if (err)
3309                 return err;
3310         virtnet_set_queues(vi, vi->curr_queue_pairs);
3311
3312         err = virtnet_cpu_notif_add(vi);
3313         if (err)
3314                 return err;
3315
3316         return 0;
3317 }
3318
3319 static struct virtio_device_id id_table[] = {
3320         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3321         { 0 },
3322 };
3323
3324 #define VIRTNET_FEATURES \
3325         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3326         VIRTIO_NET_F_MAC, \
3327         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3328         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3329         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3330         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3331         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3332         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3333         VIRTIO_NET_F_CTRL_MAC_ADDR, \
3334         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3335         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3336
3337 static unsigned int features[] = {
3338         VIRTNET_FEATURES,
3339 };
3340
3341 static unsigned int features_legacy[] = {
3342         VIRTNET_FEATURES,
3343         VIRTIO_NET_F_GSO,
3344         VIRTIO_F_ANY_LAYOUT,
3345 };
3346
3347 static struct virtio_driver virtio_net_driver = {
3348         .feature_table = features,
3349         .feature_table_size = ARRAY_SIZE(features),
3350         .feature_table_legacy = features_legacy,
3351         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3352         .driver.name =  KBUILD_MODNAME,
3353         .driver.owner = THIS_MODULE,
3354         .id_table =     id_table,
3355         .validate =     virtnet_validate,
3356         .probe =        virtnet_probe,
3357         .remove =       virtnet_remove,
3358         .config_changed = virtnet_config_changed,
3359 #ifdef CONFIG_PM_SLEEP
3360         .freeze =       virtnet_freeze,
3361         .restore =      virtnet_restore,
3362 #endif
3363 };
3364
3365 static __init int virtio_net_driver_init(void)
3366 {
3367         int ret;
3368
3369         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3370                                       virtnet_cpu_online,
3371                                       virtnet_cpu_down_prep);
3372         if (ret < 0)
3373                 goto out;
3374         virtionet_online = ret;
3375         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3376                                       NULL, virtnet_cpu_dead);
3377         if (ret)
3378                 goto err_dead;
3379
3380         ret = register_virtio_driver(&virtio_net_driver);
3381         if (ret)
3382                 goto err_virtio;
3383         return 0;
3384 err_virtio:
3385         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3386 err_dead:
3387         cpuhp_remove_multi_state(virtionet_online);
3388 out:
3389         return ret;
3390 }
3391 module_init(virtio_net_driver_init);
3392
3393 static __exit void virtio_net_driver_exit(void)
3394 {
3395         unregister_virtio_driver(&virtio_net_driver);
3396         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3397         cpuhp_remove_multi_state(virtionet_online);
3398 }
3399 module_exit(virtio_net_driver_exit);
3400
3401 MODULE_DEVICE_TABLE(virtio, id_table);
3402 MODULE_DESCRIPTION("Virtio network driver");
3403 MODULE_LICENSE("GPL");