drivers/net/virtio_net.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* A network driver using virtio.
   3  *
   4  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   5  */
   6 //#define DEBUG
   7 #include <linux/netdevice.h>
   8 #include <linux/etherdevice.h>
   9 #include <linux/ethtool.h>
  10 #include <linux/module.h>
  11 #include <linux/virtio.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/bpf.h>
  14 #include <linux/bpf_trace.h>
  15 #include <linux/scatterlist.h>
  16 #include <linux/if_vlan.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/average.h>
  20 #include <linux/filter.h>
  21 #include <linux/kernel.h>
  22 #include <net/route.h>
  23 #include <net/xdp.h>
  24 #include <net/net_failover.h>
  25
  26 static int napi_weight = NAPI_POLL_WEIGHT;
  27 module_param(napi_weight, int, 0444);
  28
  29 static bool csum = true, gso = true, napi_tx = true;
  30 module_param(csum, bool, 0444);
  31 module_param(gso, bool, 0444);
  32 module_param(napi_tx, bool, 0644);
  33
  34 /* FIXME: MTU in config. */
  35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  36 #define GOOD_COPY_LEN   128
  37
  38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  39
  40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  41 #define VIRTIO_XDP_HEADROOM 256
  42
  43 /* Separating two types of XDP xmit */
  44 #define VIRTIO_XDP_TX           BIT(0)
  45 #define VIRTIO_XDP_REDIR        BIT(1)
  46
  47 #define VIRTIO_XDP_FLAG BIT(0)
  48
  49 /* RX packet size EWMA. The average packet size is used to determine the packet
  50  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  51  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  52  * term, transient changes in packet size.
  53  */
  54 DECLARE_EWMA(pkt_len, 0, 64)
  55
  56 #define VIRTNET_DRIVER_VERSION "1.0.0"
  57
  58 static const unsigned long guest_offloads[] = {
  59         VIRTIO_NET_F_GUEST_TSO4,
  60         VIRTIO_NET_F_GUEST_TSO6,
  61         VIRTIO_NET_F_GUEST_ECN,
  62         VIRTIO_NET_F_GUEST_UFO,
  63         VIRTIO_NET_F_GUEST_CSUM
  64 };
  65
  66 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
  67                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
  68                                 (1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
  69                                 (1ULL << VIRTIO_NET_F_GUEST_UFO))
  70
  71 struct virtnet_stat_desc {
  72         char desc[ETH_GSTRING_LEN];
  73         size_t offset;
  74 };
  75
  76 struct virtnet_sq_stats {
  77         struct u64_stats_sync syncp;
  78         u64 packets;
  79         u64 bytes;
  80         u64 xdp_tx;
  81         u64 xdp_tx_drops;
  82         u64 kicks;
  83 };
  84
  85 struct virtnet_rq_stats {
  86         struct u64_stats_sync syncp;
  87         u64 packets;
  88         u64 bytes;
  89         u64 drops;
  90         u64 xdp_packets;
  91         u64 xdp_tx;
  92         u64 xdp_redirects;
  93         u64 xdp_drops;
  94         u64 kicks;
  95 };
  96
  97 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
  98 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
  99
 100 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
 101         { "packets",            VIRTNET_SQ_STAT(packets) },
 102         { "bytes",              VIRTNET_SQ_STAT(bytes) },
 103         { "xdp_tx",             VIRTNET_SQ_STAT(xdp_tx) },
 104         { "xdp_tx_drops",       VIRTNET_SQ_STAT(xdp_tx_drops) },
 105         { "kicks",              VIRTNET_SQ_STAT(kicks) },
 106 };
 107
 108 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
 109         { "packets",            VIRTNET_RQ_STAT(packets) },
 110         { "bytes",              VIRTNET_RQ_STAT(bytes) },
 111         { "drops",              VIRTNET_RQ_STAT(drops) },
 112         { "xdp_packets",        VIRTNET_RQ_STAT(xdp_packets) },
 113         { "xdp_tx",             VIRTNET_RQ_STAT(xdp_tx) },
 114         { "xdp_redirects",      VIRTNET_RQ_STAT(xdp_redirects) },
 115         { "xdp_drops",          VIRTNET_RQ_STAT(xdp_drops) },
 116         { "kicks",              VIRTNET_RQ_STAT(kicks) },
 117 };
 118
 119 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 120 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 121
 122 /* Internal representation of a send virtqueue */
 123 struct send_queue {
 124         /* Virtqueue associated with this send _queue */
 125         struct virtqueue *vq;
 126
 127         /* TX: fragments + linear part + virtio header */
 128         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 129
 130         /* Name of the send queue: output.$index */
 131         char name[40];
 132
 133         struct virtnet_sq_stats stats;
 134
 135         struct napi_struct napi;
 136 };
 137
 138 /* Internal representation of a receive virtqueue */
 139 struct receive_queue {
 140         /* Virtqueue associated with this receive_queue */
 141         struct virtqueue *vq;
 142
 143         struct napi_struct napi;
 144
 145         struct bpf_prog __rcu *xdp_prog;
 146
 147         struct virtnet_rq_stats stats;
 148
 149         /* Chain pages by the private ptr. */
 150         struct page *pages;
 151
 152         /* Average packet length for mergeable receive buffers. */
 153         struct ewma_pkt_len mrg_avg_pkt_len;
 154
 155         /* Page frag for packet buffer allocation. */
 156         struct page_frag alloc_frag;
 157
 158         /* RX: fragments + linear part + virtio header */
 159         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 160
 161         /* Min single buffer size for mergeable buffers case. */
 162         unsigned int min_buf_len;
 163
 164         /* Name of this receive queue: input.$index */
 165         char name[40];
 166
 167         struct xdp_rxq_info xdp_rxq;
 168 };
 169
 170 /* Control VQ buffers: protected by the rtnl lock */
 171 struct control_buf {
 172         struct virtio_net_ctrl_hdr hdr;
 173         virtio_net_ctrl_ack status;
 174         struct virtio_net_ctrl_mq mq;
 175         u8 promisc;
 176         u8 allmulti;
 177         __virtio16 vid;
 178         __virtio64 offloads;
 179 };
 180
 181 struct virtnet_info {
 182         struct virtio_device *vdev;
 183         struct virtqueue *cvq;
 184         struct net_device *dev;
 185         struct send_queue *sq;
 186         struct receive_queue *rq;
 187         unsigned int status;
 188
 189         /* Max # of queue pairs supported by the device */
 190         u16 max_queue_pairs;
 191
 192         /* # of queue pairs currently used by the driver */
 193         u16 curr_queue_pairs;
 194
 195         /* # of XDP queue pairs currently used by the driver */
 196         u16 xdp_queue_pairs;
 197
 198         /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
 199         bool xdp_enabled;
 200
 201         /* I like... big packets and I cannot lie! */
 202         bool big_packets;
 203
 204         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 205         bool mergeable_rx_bufs;
 206
 207         /* Has control virtqueue */
 208         bool has_cvq;
 209
 210         /* Host can handle any s/g split between our header and packet data */
 211         bool any_header_sg;
 212
 213         /* Packet virtio header size */
 214         u8 hdr_len;
 215
 216         /* Work struct for refilling if we run low on memory. */
 217         struct delayed_work refill;
 218
 219         /* Work struct for config space updates */
 220         struct work_struct config_work;
 221
 222         /* Does the affinity hint is set for virtqueues? */
 223         bool affinity_hint_set;
 224
 225         /* CPU hotplug instances for online & dead */
 226         struct hlist_node node;
 227         struct hlist_node node_dead;
 228
 229         struct control_buf *ctrl;
 230
 231         /* Ethtool settings */
 232         u8 duplex;
 233         u32 speed;
 234
 235         unsigned long guest_offloads;
 236         unsigned long guest_offloads_capable;
 237
 238         /* failover when STANDBY feature enabled */
 239         struct failover *failover;
 240 };
 241
 242 struct padded_vnet_hdr {
 243         struct virtio_net_hdr_mrg_rxbuf hdr;
 244         /*
 245          * hdr is in a separate sg buffer, and data sg buffer shares same page
 246          * with this header sg. This padding makes next sg 16 byte aligned
 247          * after the header.
 248          */
 249         char padding[4];
 250 };
 251
 252 static bool is_xdp_frame(void *ptr)
 253 {
 254         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 255 }
 256
 257 static void *xdp_to_ptr(struct xdp_frame *ptr)
 258 {
 259         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 260 }
 261
 262 static struct xdp_frame *ptr_to_xdp(void *ptr)
 263 {
 264         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 265 }
 266
 267 /* Converting between virtqueue no. and kernel tx/rx queue no.
 268  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 269  */
 270 static int vq2txq(struct virtqueue *vq)
 271 {
 272         return (vq->index - 1) / 2;
 273 }
 274
 275 static int txq2vq(int txq)
 276 {
 277         return txq * 2 + 1;
 278 }
 279
 280 static int vq2rxq(struct virtqueue *vq)
 281 {
 282         return vq->index / 2;
 283 }
 284
 285 static int rxq2vq(int rxq)
 286 {
 287         return rxq * 2;
 288 }
 289
 290 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 291 {
 292         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 293 }
 294
 295 /*
 296  * private is used to chain pages for big packets, put the whole
 297  * most recent used list in the beginning for reuse
 298  */
 299 static void give_pages(struct receive_queue *rq, struct page *page)
 300 {
 301         struct page *end;
 302
 303         /* Find end of list, sew whole thing into vi->rq.pages. */
 304         for (end = page; end->private; end = (struct page *)end->private);
 305         end->private = (unsigned long)rq->pages;
 306         rq->pages = page;
 307 }
 308
 309 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 310 {
 311         struct page *p = rq->pages;
 312
 313         if (p) {
 314                 rq->pages = (struct page *)p->private;
 315                 /* clear private here, it is used to chain pages */
 316                 p->private = 0;
 317         } else
 318                 p = alloc_page(gfp_mask);
 319         return p;
 320 }
 321
 322 static void virtqueue_napi_schedule(struct napi_struct *napi,
 323                                     struct virtqueue *vq)
 324 {
 325         if (napi_schedule_prep(napi)) {
 326                 virtqueue_disable_cb(vq);
 327                 __napi_schedule(napi);
 328         }
 329 }
 330
 331 static void virtqueue_napi_complete(struct napi_struct *napi,
 332                                     struct virtqueue *vq, int processed)
 333 {
 334         int opaque;
 335
 336         opaque = virtqueue_enable_cb_prepare(vq);
 337         if (napi_complete_done(napi, processed)) {
 338                 if (unlikely(virtqueue_poll(vq, opaque)))
 339                         virtqueue_napi_schedule(napi, vq);
 340         } else {
 341                 virtqueue_disable_cb(vq);
 342         }
 343 }
 344
 345 static void skb_xmit_done(struct virtqueue *vq)
 346 {
 347         struct virtnet_info *vi = vq->vdev->priv;
 348         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 349
 350         /* Suppress further interrupts. */
 351         virtqueue_disable_cb(vq);
 352
 353         if (napi->weight)
 354                 virtqueue_napi_schedule(napi, vq);
 355         else
 356                 /* We were probably waiting for more output buffers. */
 357                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 358 }
 359
 360 #define MRG_CTX_HEADER_SHIFT 22
 361 static void *mergeable_len_to_ctx(unsigned int truesize,
 362                                   unsigned int headroom)
 363 {
 364         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 365 }
 366
 367 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 368 {
 369         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 370 }
 371
 372 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 373 {
 374         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 375 }
 376
 377 /* Called from bottom half context */
 378 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 379                                    struct receive_queue *rq,
 380                                    struct page *page, unsigned int offset,
 381                                    unsigned int len, unsigned int truesize,
 382                                    bool hdr_valid, unsigned int metasize,
 383                                    unsigned int headroom)
 384 {
 385         struct sk_buff *skb;
 386         struct virtio_net_hdr_mrg_rxbuf *hdr;
 387         unsigned int copy, hdr_len, hdr_padded_len;
 388         struct page *page_to_free = NULL;
 389         int tailroom, shinfo_size;
 390         char *p, *hdr_p, *buf;
 391
 392         p = page_address(page) + offset;
 393         hdr_p = p;
 394
 395         hdr_len = vi->hdr_len;
 396         if (vi->mergeable_rx_bufs)
 397                 hdr_padded_len = sizeof(*hdr);
 398         else
 399                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 400
 401         /* If headroom is not 0, there is an offset between the beginning of the
 402          * data and the allocated space, otherwise the data and the allocated
 403          * space are aligned.
 404          *
 405          * Buffers with headroom use PAGE_SIZE as alloc size, see
 406          * add_recvbuf_mergeable() + get_mergeable_buf_len()
 407          */
 408         truesize = headroom ? PAGE_SIZE : truesize;
 409         tailroom = truesize - len - headroom;
 410         buf = p - headroom;
 411
 412         len -= hdr_len;
 413         offset += hdr_padded_len;
 414         p += hdr_padded_len;
 415
 416         shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 417
 418         /* copy small packet so we can reuse these pages */
 419         if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
 420                 skb = build_skb(buf, truesize);
 421                 if (unlikely(!skb))
 422                         return NULL;
 423
 424                 skb_reserve(skb, p - buf);
 425                 skb_put(skb, len);
 426
 427                 page = (struct page *)page->private;
 428                 if (page)
 429                         give_pages(rq, page);
 430                 goto ok;
 431         }
 432
 433         /* copy small packet so we can reuse these pages for small data */
 434         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 435         if (unlikely(!skb))
 436                 return NULL;
 437
 438         /* Copy all frame if it fits skb->head, otherwise
 439          * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
 440          */
 441         if (len <= skb_tailroom(skb))
 442                 copy = len;
 443         else
 444                 copy = ETH_HLEN + metasize;
 445         skb_put_data(skb, p, copy);
 446
 447         len -= copy;
 448         offset += copy;
 449
 450         if (vi->mergeable_rx_bufs) {
 451                 if (len)
 452                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 453                 else
 454                         page_to_free = page;
 455                 goto ok;
 456         }
 457
 458         /*
 459          * Verify that we can indeed put this data into a skb.
 460          * This is here to handle cases when the device erroneously
 461          * tries to receive more than is possible. This is usually
 462          * the case of a broken device.
 463          */
 464         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 465                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 466                 dev_kfree_skb(skb);
 467                 return NULL;
 468         }
 469         BUG_ON(offset >= PAGE_SIZE);
 470         while (len) {
 471                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 472                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 473                                 frag_size, truesize);
 474                 len -= frag_size;
 475                 page = (struct page *)page->private;
 476                 offset = 0;
 477         }
 478
 479         if (page)
 480                 give_pages(rq, page);
 481
 482 ok:
 483         /* hdr_valid means no XDP, so we can copy the vnet header */
 484         if (hdr_valid) {
 485                 hdr = skb_vnet_hdr(skb);
 486                 memcpy(hdr, hdr_p, hdr_len);
 487         }
 488         if (page_to_free)
 489                 put_page(page_to_free);
 490
 491         if (metasize) {
 492                 __skb_pull(skb, metasize);
 493                 skb_metadata_set(skb, metasize);
 494         }
 495
 496         return skb;
 497 }
 498
 499 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 500                                    struct send_queue *sq,
 501                                    struct xdp_frame *xdpf)
 502 {
 503         struct virtio_net_hdr_mrg_rxbuf *hdr;
 504         int err;
 505
 506         if (unlikely(xdpf->headroom < vi->hdr_len))
 507                 return -EOVERFLOW;
 508
 509         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 510         xdpf->data -= vi->hdr_len;
 511         /* Zero header and leave csum up to XDP layers */
 512         hdr = xdpf->data;
 513         memset(hdr, 0, vi->hdr_len);
 514         xdpf->len   += vi->hdr_len;
 515
 516         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 517
 518         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
 519                                    GFP_ATOMIC);
 520         if (unlikely(err))
 521                 return -ENOSPC; /* Caller handle free/refcnt */
 522
 523         return 0;
 524 }
 525
 526 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
 527  * the current cpu, so it does not need to be locked.
 528  *
 529  * Here we use marco instead of inline functions because we have to deal with
 530  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
 531  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
 532  * functions to perfectly solve these three problems at the same time.
 533  */
 534 #define virtnet_xdp_get_sq(vi) ({                                       \
 535         int cpu = smp_processor_id();                                   \
 536         struct netdev_queue *txq;                                       \
 537         typeof(vi) v = (vi);                                            \
 538         unsigned int qp;                                                \
 539                                                                         \
 540         if (v->curr_queue_pairs > nr_cpu_ids) {                         \
 541                 qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
 542                 qp += cpu;                                              \
 543                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 544                 __netif_tx_acquire(txq);                                \
 545         } else {                                                        \
 546                 qp = cpu % v->curr_queue_pairs;                         \
 547                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 548                 __netif_tx_lock(txq, cpu);                              \
 549         }                                                               \
 550         v->sq + qp;                                                     \
 551 })
 552
 553 #define virtnet_xdp_put_sq(vi, q) {                                     \
 554         struct netdev_queue *txq;                                       \
 555         typeof(vi) v = (vi);                                            \
 556                                                                         \
 557         txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
 558         if (v->curr_queue_pairs > nr_cpu_ids)                           \
 559                 __netif_tx_release(txq);                                \
 560         else                                                            \
 561                 __netif_tx_unlock(txq);                                 \
 562 }
 563
 564 static int virtnet_xdp_xmit(struct net_device *dev,
 565                             int n, struct xdp_frame **frames, u32 flags)
 566 {
 567         struct virtnet_info *vi = netdev_priv(dev);
 568         struct receive_queue *rq = vi->rq;
 569         struct bpf_prog *xdp_prog;
 570         struct send_queue *sq;
 571         unsigned int len;
 572         int packets = 0;
 573         int bytes = 0;
 574         int nxmit = 0;
 575         int kicks = 0;
 576         void *ptr;
 577         int ret;
 578         int i;
 579
 580         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 581          * indicate XDP resources have been successfully allocated.
 582          */
 583         xdp_prog = rcu_access_pointer(rq->xdp_prog);
 584         if (!xdp_prog)
 585                 return -ENXIO;
 586
 587         sq = virtnet_xdp_get_sq(vi);
 588
 589         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 590                 ret = -EINVAL;
 591                 goto out;
 592         }
 593
 594         /* Free up any pending old buffers before queueing new ones. */
 595         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 596                 if (likely(is_xdp_frame(ptr))) {
 597                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 598
 599                         bytes += frame->len;
 600                         xdp_return_frame(frame);
 601                 } else {
 602                         struct sk_buff *skb = ptr;
 603
 604                         bytes += skb->len;
 605                         napi_consume_skb(skb, false);
 606                 }
 607                 packets++;
 608         }
 609
 610         for (i = 0; i < n; i++) {
 611                 struct xdp_frame *xdpf = frames[i];
 612
 613                 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
 614                         break;
 615                 nxmit++;
 616         }
 617         ret = nxmit;
 618
 619         if (flags & XDP_XMIT_FLUSH) {
 620                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
 621                         kicks = 1;
 622         }
 623 out:
 624         u64_stats_update_begin(&sq->stats.syncp);
 625         sq->stats.bytes += bytes;
 626         sq->stats.packets += packets;
 627         sq->stats.xdp_tx += n;
 628         sq->stats.xdp_tx_drops += n - nxmit;
 629         sq->stats.kicks += kicks;
 630         u64_stats_update_end(&sq->stats.syncp);
 631
 632         virtnet_xdp_put_sq(vi, sq);
 633         return ret;
 634 }
 635
 636 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 637 {
 638         return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
 639 }
 640
 641 /* We copy the packet for XDP in the following cases:
 642  *
 643  * 1) Packet is scattered across multiple rx buffers.
 644  * 2) Headroom space is insufficient.
 645  *
 646  * This is inefficient but it's a temporary condition that
 647  * we hit right after XDP is enabled and until queue is refilled
 648  * with large buffers with sufficient headroom - so it should affect
 649  * at most queue size packets.
 650  * Afterwards, the conditions to enable
 651  * XDP should preclude the underlying device from sending packets
 652  * across multiple buffers (num_buf > 1), and we make sure buffers
 653  * have enough headroom.
 654  */
 655 static struct page *xdp_linearize_page(struct receive_queue *rq,
 656                                        u16 *num_buf,
 657                                        struct page *p,
 658                                        int offset,
 659                                        int page_off,
 660                                        unsigned int *len)
 661 {
 662         struct page *page = alloc_page(GFP_ATOMIC);
 663
 664         if (!page)
 665                 return NULL;
 666
 667         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 668         page_off += *len;
 669
 670         while (--*num_buf) {
 671                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 672                 unsigned int buflen;
 673                 void *buf;
 674                 int off;
 675
 676                 buf = virtqueue_get_buf(rq->vq, &buflen);
 677                 if (unlikely(!buf))
 678                         goto err_buf;
 679
 680                 p = virt_to_head_page(buf);
 681                 off = buf - page_address(p);
 682
 683                 /* guard against a misconfigured or uncooperative backend that
 684                  * is sending packet larger than the MTU.
 685                  */
 686                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 687                         put_page(p);
 688                         goto err_buf;
 689                 }
 690
 691                 memcpy(page_address(page) + page_off,
 692                        page_address(p) + off, buflen);
 693                 page_off += buflen;
 694                 put_page(p);
 695         }
 696
 697         /* Headroom does not contribute to packet length */
 698         *len = page_off - VIRTIO_XDP_HEADROOM;
 699         return page;
 700 err_buf:
 701         __free_pages(page, 0);
 702         return NULL;
 703 }
 704
 705 static struct sk_buff *receive_small(struct net_device *dev,
 706                                      struct virtnet_info *vi,
 707                                      struct receive_queue *rq,
 708                                      void *buf, void *ctx,
 709                                      unsigned int len,
 710                                      unsigned int *xdp_xmit,
 711                                      struct virtnet_rq_stats *stats)
 712 {
 713         struct sk_buff *skb;
 714         struct bpf_prog *xdp_prog;
 715         unsigned int xdp_headroom = (unsigned long)ctx;
 716         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 717         unsigned int headroom = vi->hdr_len + header_offset;
 718         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 719                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 720         struct page *page = virt_to_head_page(buf);
 721         unsigned int delta = 0;
 722         struct page *xdp_page;
 723         int err;
 724         unsigned int metasize = 0;
 725
 726         len -= vi->hdr_len;
 727         stats->bytes += len;
 728
 729         if (unlikely(len > GOOD_PACKET_LEN)) {
 730                 pr_debug("%s: rx error: len %u exceeds max size %d\n",
 731                          dev->name, len, GOOD_PACKET_LEN);
 732                 dev->stats.rx_length_errors++;
 733                 goto err_len;
 734         }
 735         rcu_read_lock();
 736         xdp_prog = rcu_dereference(rq->xdp_prog);
 737         if (xdp_prog) {
 738                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 739                 struct xdp_frame *xdpf;
 740                 struct xdp_buff xdp;
 741                 void *orig_data;
 742                 u32 act;
 743
 744                 if (unlikely(hdr->hdr.gso_type))
 745                         goto err_xdp;
 746
 747                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 748                         int offset = buf - page_address(page) + header_offset;
 749                         unsigned int tlen = len + vi->hdr_len;
 750                         u16 num_buf = 1;
 751
 752                         xdp_headroom = virtnet_get_headroom(vi);
 753                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 754                         headroom = vi->hdr_len + header_offset;
 755                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 756                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 757                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 758                                                       offset, header_offset,
 759                                                       &tlen);
 760                         if (!xdp_page)
 761                                 goto err_xdp;
 762
 763                         buf = page_address(xdp_page);
 764                         put_page(page);
 765                         page = xdp_page;
 766                 }
 767
 768                 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
 769                 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
 770                                  xdp_headroom, len, true);
 771                 orig_data = xdp.data;
 772                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 773                 stats->xdp_packets++;
 774
 775                 switch (act) {
 776                 case XDP_PASS:
 777                         /* Recalculate length in case bpf program changed it */
 778                         delta = orig_data - xdp.data;
 779                         len = xdp.data_end - xdp.data;
 780                         metasize = xdp.data - xdp.data_meta;
 781                         break;
 782                 case XDP_TX:
 783                         stats->xdp_tx++;
 784                         xdpf = xdp_convert_buff_to_frame(&xdp);
 785                         if (unlikely(!xdpf))
 786                                 goto err_xdp;
 787                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 788                         if (unlikely(!err)) {
 789                                 xdp_return_frame_rx_napi(xdpf);
 790                         } else if (unlikely(err < 0)) {
 791                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 792                                 goto err_xdp;
 793                         }
 794                         *xdp_xmit |= VIRTIO_XDP_TX;
 795                         rcu_read_unlock();
 796                         goto xdp_xmit;
 797                 case XDP_REDIRECT:
 798                         stats->xdp_redirects++;
 799                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 800                         if (err)
 801                                 goto err_xdp;
 802                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 803                         rcu_read_unlock();
 804                         goto xdp_xmit;
 805                 default:
 806                         bpf_warn_invalid_xdp_action(act);
 807                         fallthrough;
 808                 case XDP_ABORTED:
 809                         trace_xdp_exception(vi->dev, xdp_prog, act);
 810                         goto err_xdp;
 811                 case XDP_DROP:
 812                         goto err_xdp;
 813                 }
 814         }
 815         rcu_read_unlock();
 816
 817         skb = build_skb(buf, buflen);
 818         if (!skb) {
 819                 put_page(page);
 820                 goto err;
 821         }
 822         skb_reserve(skb, headroom - delta);
 823         skb_put(skb, len);
 824         if (!xdp_prog) {
 825                 buf += header_offset;
 826                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 827         } /* keep zeroed vnet hdr since XDP is loaded */
 828
 829         if (metasize)
 830                 skb_metadata_set(skb, metasize);
 831
 832 err:
 833         return skb;
 834
 835 err_xdp:
 836         rcu_read_unlock();
 837         stats->xdp_drops++;
 838 err_len:
 839         stats->drops++;
 840         put_page(page);
 841 xdp_xmit:
 842         return NULL;
 843 }
 844
 845 static struct sk_buff *receive_big(struct net_device *dev,
 846                                    struct virtnet_info *vi,
 847                                    struct receive_queue *rq,
 848                                    void *buf,
 849                                    unsigned int len,
 850                                    struct virtnet_rq_stats *stats)
 851 {
 852         struct page *page = buf;
 853         struct sk_buff *skb =
 854                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
 855
 856         stats->bytes += len - vi->hdr_len;
 857         if (unlikely(!skb))
 858                 goto err;
 859
 860         return skb;
 861
 862 err:
 863         stats->drops++;
 864         give_pages(rq, page);
 865         return NULL;
 866 }
 867
 868 static struct sk_buff *receive_mergeable(struct net_device *dev,
 869                                          struct virtnet_info *vi,
 870                                          struct receive_queue *rq,
 871                                          void *buf,
 872                                          void *ctx,
 873                                          unsigned int len,
 874                                          unsigned int *xdp_xmit,
 875                                          struct virtnet_rq_stats *stats)
 876 {
 877         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 878         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 879         struct page *page = virt_to_head_page(buf);
 880         int offset = buf - page_address(page);
 881         struct sk_buff *head_skb, *curr_skb;
 882         struct bpf_prog *xdp_prog;
 883         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
 884         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 885         unsigned int metasize = 0;
 886         unsigned int frame_sz;
 887         int err;
 888
 889         head_skb = NULL;
 890         stats->bytes += len - vi->hdr_len;
 891
 892         if (unlikely(len > truesize)) {
 893                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 894                          dev->name, len, (unsigned long)ctx);
 895                 dev->stats.rx_length_errors++;
 896                 goto err_skb;
 897         }
 898         rcu_read_lock();
 899         xdp_prog = rcu_dereference(rq->xdp_prog);
 900         if (xdp_prog) {
 901                 struct xdp_frame *xdpf;
 902                 struct page *xdp_page;
 903                 struct xdp_buff xdp;
 904                 void *data;
 905                 u32 act;
 906
 907                 /* Transient failure which in theory could occur if
 908                  * in-flight packets from before XDP was enabled reach
 909                  * the receive path after XDP is loaded.
 910                  */
 911                 if (unlikely(hdr->hdr.gso_type))
 912                         goto err_xdp;
 913
 914                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 915                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 916                  */
 917                 frame_sz = headroom ? PAGE_SIZE : truesize;
 918
 919                 /* This happens when rx buffer size is underestimated
 920                  * or headroom is not enough because of the buffer
 921                  * was refilled before XDP is set. This should only
 922                  * happen for the first several packets, so we don't
 923                  * care much about its performance.
 924                  */
 925                 if (unlikely(num_buf > 1 ||
 926                              headroom < virtnet_get_headroom(vi))) {
 927                         /* linearize data for XDP */
 928                         xdp_page = xdp_linearize_page(rq, &num_buf,
 929                                                       page, offset,
 930                                                       VIRTIO_XDP_HEADROOM,
 931                                                       &len);
 932                         frame_sz = PAGE_SIZE;
 933
 934                         if (!xdp_page)
 935                                 goto err_xdp;
 936                         offset = VIRTIO_XDP_HEADROOM;
 937                 } else {
 938                         xdp_page = page;
 939                 }
 940
 941                 /* Allow consuming headroom but reserve enough space to push
 942                  * the descriptor on if we get an XDP_TX return code.
 943                  */
 944                 data = page_address(xdp_page) + offset;
 945                 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
 946                 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
 947                                  VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
 948
 949                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 950                 stats->xdp_packets++;
 951
 952                 switch (act) {
 953                 case XDP_PASS:
 954                         metasize = xdp.data - xdp.data_meta;
 955
 956                         /* recalculate offset to account for any header
 957                          * adjustments and minus the metasize to copy the
 958                          * metadata in page_to_skb(). Note other cases do not
 959                          * build an skb and avoid using offset
 960                          */
 961                         offset = xdp.data - page_address(xdp_page) -
 962                                  vi->hdr_len - metasize;
 963
 964                         /* recalculate len if xdp.data, xdp.data_end or
 965                          * xdp.data_meta were adjusted
 966                          */
 967                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 968                         /* We can only create skb based on xdp_page. */
 969                         if (unlikely(xdp_page != page)) {
 970                                 rcu_read_unlock();
 971                                 put_page(page);
 972                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
 973                                                        len, PAGE_SIZE, false,
 974                                                        metasize,
 975                                                        VIRTIO_XDP_HEADROOM);
 976                                 return head_skb;
 977                         }
 978                         break;
 979                 case XDP_TX:
 980                         stats->xdp_tx++;
 981                         xdpf = xdp_convert_buff_to_frame(&xdp);
 982                         if (unlikely(!xdpf))
 983                                 goto err_xdp;
 984                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 985                         if (unlikely(!err)) {
 986                                 xdp_return_frame_rx_napi(xdpf);
 987                         } else if (unlikely(err < 0)) {
 988                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 989                                 if (unlikely(xdp_page != page))
 990                                         put_page(xdp_page);
 991                                 goto err_xdp;
 992                         }
 993                         *xdp_xmit |= VIRTIO_XDP_TX;
 994                         if (unlikely(xdp_page != page))
 995                                 put_page(page);
 996                         rcu_read_unlock();
 997                         goto xdp_xmit;
 998                 case XDP_REDIRECT:
 999                         stats->xdp_redirects++;
1000                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
1001                         if (err) {
1002                                 if (unlikely(xdp_page != page))
1003                                         put_page(xdp_page);
1004                                 goto err_xdp;
1005                         }
1006                         *xdp_xmit |= VIRTIO_XDP_REDIR;
1007                         if (unlikely(xdp_page != page))
1008                                 put_page(page);
1009                         rcu_read_unlock();
1010                         goto xdp_xmit;
1011                 default:
1012                         bpf_warn_invalid_xdp_action(act);
1013                         fallthrough;
1014                 case XDP_ABORTED:
1015                         trace_xdp_exception(vi->dev, xdp_prog, act);
1016                         fallthrough;
1017                 case XDP_DROP:
1018                         if (unlikely(xdp_page != page))
1019                                 __free_pages(xdp_page, 0);
1020                         goto err_xdp;
1021                 }
1022         }
1023         rcu_read_unlock();
1024
1025         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1026                                metasize, headroom);
1027         curr_skb = head_skb;
1028
1029         if (unlikely(!curr_skb))
1030                 goto err_skb;
1031         while (--num_buf) {
1032                 int num_skb_frags;
1033
1034                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1035                 if (unlikely(!buf)) {
1036                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1037                                  dev->name, num_buf,
1038                                  virtio16_to_cpu(vi->vdev,
1039                                                  hdr->num_buffers));
1040                         dev->stats.rx_length_errors++;
1041                         goto err_buf;
1042                 }
1043
1044                 stats->bytes += len;
1045                 page = virt_to_head_page(buf);
1046
1047                 truesize = mergeable_ctx_to_truesize(ctx);
1048                 if (unlikely(len > truesize)) {
1049                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1050                                  dev->name, len, (unsigned long)ctx);
1051                         dev->stats.rx_length_errors++;
1052                         goto err_skb;
1053                 }
1054
1055                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1056                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1057                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1058
1059                         if (unlikely(!nskb))
1060                                 goto err_skb;
1061                         if (curr_skb == head_skb)
1062                                 skb_shinfo(curr_skb)->frag_list = nskb;
1063                         else
1064                                 curr_skb->next = nskb;
1065                         curr_skb = nskb;
1066                         head_skb->truesize += nskb->truesize;
1067                         num_skb_frags = 0;
1068                 }
1069                 if (curr_skb != head_skb) {
1070                         head_skb->data_len += len;
1071                         head_skb->len += len;
1072                         head_skb->truesize += truesize;
1073                 }
1074                 offset = buf - page_address(page);
1075                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1076                         put_page(page);
1077                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1078                                              len, truesize);
1079                 } else {
1080                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1081                                         offset, len, truesize);
1082                 }
1083         }
1084
1085         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1086         return head_skb;
1087
1088 err_xdp:
1089         rcu_read_unlock();
1090         stats->xdp_drops++;
1091 err_skb:
1092         put_page(page);
1093         while (num_buf-- > 1) {
1094                 buf = virtqueue_get_buf(rq->vq, &len);
1095                 if (unlikely(!buf)) {
1096                         pr_debug("%s: rx error: %d buffers missing\n",
1097                                  dev->name, num_buf);
1098                         dev->stats.rx_length_errors++;
1099                         break;
1100                 }
1101                 stats->bytes += len;
1102                 page = virt_to_head_page(buf);
1103                 put_page(page);
1104         }
1105 err_buf:
1106         stats->drops++;
1107         dev_kfree_skb(head_skb);
1108 xdp_xmit:
1109         return NULL;
1110 }
1111
1112 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1113                         void *buf, unsigned int len, void **ctx,
1114                         unsigned int *xdp_xmit,
1115                         struct virtnet_rq_stats *stats)
1116 {
1117         struct net_device *dev = vi->dev;
1118         struct sk_buff *skb;
1119         struct virtio_net_hdr_mrg_rxbuf *hdr;
1120
1121         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1122                 pr_debug("%s: short packet %i\n", dev->name, len);
1123                 dev->stats.rx_length_errors++;
1124                 if (vi->mergeable_rx_bufs) {
1125                         put_page(virt_to_head_page(buf));
1126                 } else if (vi->big_packets) {
1127                         give_pages(rq, buf);
1128                 } else {
1129                         put_page(virt_to_head_page(buf));
1130                 }
1131                 return;
1132         }
1133
1134         if (vi->mergeable_rx_bufs)
1135                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1136                                         stats);
1137         else if (vi->big_packets)
1138                 skb = receive_big(dev, vi, rq, buf, len, stats);
1139         else
1140                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1141
1142         if (unlikely(!skb))
1143                 return;
1144
1145         hdr = skb_vnet_hdr(skb);
1146
1147         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1148                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1149
1150         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1151                                   virtio_is_little_endian(vi->vdev))) {
1152                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1153                                      dev->name, hdr->hdr.gso_type,
1154                                      hdr->hdr.gso_size);
1155                 goto frame_err;
1156         }
1157
1158         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1159         skb->protocol = eth_type_trans(skb, dev);
1160         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1161                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1162
1163         napi_gro_receive(&rq->napi, skb);
1164         return;
1165
1166 frame_err:
1167         dev->stats.rx_frame_errors++;
1168         dev_kfree_skb(skb);
1169 }
1170
1171 /* Unlike mergeable buffers, all buffers are allocated to the
1172  * same size, except for the headroom. For this reason we do
1173  * not need to use  mergeable_len_to_ctx here - it is enough
1174  * to store the headroom as the context ignoring the truesize.
1175  */
1176 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1177                              gfp_t gfp)
1178 {
1179         struct page_frag *alloc_frag = &rq->alloc_frag;
1180         char *buf;
1181         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1182         void *ctx = (void *)(unsigned long)xdp_headroom;
1183         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1184         int err;
1185
1186         len = SKB_DATA_ALIGN(len) +
1187               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1188         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1189                 return -ENOMEM;
1190
1191         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1192         get_page(alloc_frag->page);
1193         alloc_frag->offset += len;
1194         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1195                     vi->hdr_len + GOOD_PACKET_LEN);
1196         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1197         if (err < 0)
1198                 put_page(virt_to_head_page(buf));
1199         return err;
1200 }
1201
1202 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1203                            gfp_t gfp)
1204 {
1205         struct page *first, *list = NULL;
1206         char *p;
1207         int i, err, offset;
1208
1209         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1210
1211         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1212         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1213                 first = get_a_page(rq, gfp);
1214                 if (!first) {
1215                         if (list)
1216                                 give_pages(rq, list);
1217                         return -ENOMEM;
1218                 }
1219                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1220
1221                 /* chain new page in list head to match sg */
1222                 first->private = (unsigned long)list;
1223                 list = first;
1224         }
1225
1226         first = get_a_page(rq, gfp);
1227         if (!first) {
1228                 give_pages(rq, list);
1229                 return -ENOMEM;
1230         }
1231         p = page_address(first);
1232
1233         /* rq->sg[0], rq->sg[1] share the same page */
1234         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1235         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1236
1237         /* rq->sg[1] for data packet, from offset */
1238         offset = sizeof(struct padded_vnet_hdr);
1239         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1240
1241         /* chain first in list head */
1242         first->private = (unsigned long)list;
1243         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1244                                   first, gfp);
1245         if (err < 0)
1246                 give_pages(rq, first);
1247
1248         return err;
1249 }
1250
1251 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1252                                           struct ewma_pkt_len *avg_pkt_len,
1253                                           unsigned int room)
1254 {
1255         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1256         unsigned int len;
1257
1258         if (room)
1259                 return PAGE_SIZE - room;
1260
1261         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1262                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1263
1264         return ALIGN(len, L1_CACHE_BYTES);
1265 }
1266
1267 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1268                                  struct receive_queue *rq, gfp_t gfp)
1269 {
1270         struct page_frag *alloc_frag = &rq->alloc_frag;
1271         unsigned int headroom = virtnet_get_headroom(vi);
1272         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1273         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1274         char *buf;
1275         void *ctx;
1276         int err;
1277         unsigned int len, hole;
1278
1279         /* Extra tailroom is needed to satisfy XDP's assumption. This
1280          * means rx frags coalescing won't work, but consider we've
1281          * disabled GSO for XDP, it won't be a big issue.
1282          */
1283         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1284         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1285                 return -ENOMEM;
1286
1287         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1288         buf += headroom; /* advance address leaving hole at front of pkt */
1289         get_page(alloc_frag->page);
1290         alloc_frag->offset += len + room;
1291         hole = alloc_frag->size - alloc_frag->offset;
1292         if (hole < len + room) {
1293                 /* To avoid internal fragmentation, if there is very likely not
1294                  * enough space for another buffer, add the remaining space to
1295                  * the current buffer.
1296                  */
1297                 len += hole;
1298                 alloc_frag->offset += hole;
1299         }
1300
1301         sg_init_one(rq->sg, buf, len);
1302         ctx = mergeable_len_to_ctx(len, headroom);
1303         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1304         if (err < 0)
1305                 put_page(virt_to_head_page(buf));
1306
1307         return err;
1308 }
1309
1310 /*
1311  * Returns false if we couldn't fill entirely (OOM).
1312  *
1313  * Normally run in the receive path, but can also be run from ndo_open
1314  * before we're receiving packets, or from refill_work which is
1315  * careful to disable receiving (using napi_disable).
1316  */
1317 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1318                           gfp_t gfp)
1319 {
1320         int err;
1321         bool oom;
1322
1323         do {
1324                 if (vi->mergeable_rx_bufs)
1325                         err = add_recvbuf_mergeable(vi, rq, gfp);
1326                 else if (vi->big_packets)
1327                         err = add_recvbuf_big(vi, rq, gfp);
1328                 else
1329                         err = add_recvbuf_small(vi, rq, gfp);
1330
1331                 oom = err == -ENOMEM;
1332                 if (err)
1333                         break;
1334         } while (rq->vq->num_free);
1335         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1336                 unsigned long flags;
1337
1338                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1339                 rq->stats.kicks++;
1340                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1341         }
1342
1343         return !oom;
1344 }
1345
1346 static void skb_recv_done(struct virtqueue *rvq)
1347 {
1348         struct virtnet_info *vi = rvq->vdev->priv;
1349         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1350
1351         virtqueue_napi_schedule(&rq->napi, rvq);
1352 }
1353
1354 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1355 {
1356         napi_enable(napi);
1357
1358         /* If all buffers were filled by other side before we napi_enabled, we
1359          * won't get another interrupt, so process any outstanding packets now.
1360          * Call local_bh_enable after to trigger softIRQ processing.
1361          */
1362         local_bh_disable();
1363         virtqueue_napi_schedule(napi, vq);
1364         local_bh_enable();
1365 }
1366
1367 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1368                                    struct virtqueue *vq,
1369                                    struct napi_struct *napi)
1370 {
1371         if (!napi->weight)
1372                 return;
1373
1374         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1375          * enable the feature if this is likely affine with the transmit path.
1376          */
1377         if (!vi->affinity_hint_set) {
1378                 napi->weight = 0;
1379                 return;
1380         }
1381
1382         return virtnet_napi_enable(vq, napi);
1383 }
1384
1385 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1386 {
1387         if (napi->weight)
1388                 napi_disable(napi);
1389 }
1390
1391 static void refill_work(struct work_struct *work)
1392 {
1393         struct virtnet_info *vi =
1394                 container_of(work, struct virtnet_info, refill.work);
1395         bool still_empty;
1396         int i;
1397
1398         for (i = 0; i < vi->curr_queue_pairs; i++) {
1399                 struct receive_queue *rq = &vi->rq[i];
1400
1401                 napi_disable(&rq->napi);
1402                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1403                 virtnet_napi_enable(rq->vq, &rq->napi);
1404
1405                 /* In theory, this can happen: if we don't get any buffers in
1406                  * we will *never* try to fill again.
1407                  */
1408                 if (still_empty)
1409                         schedule_delayed_work(&vi->refill, HZ/2);
1410         }
1411 }
1412
1413 static int virtnet_receive(struct receive_queue *rq, int budget,
1414                            unsigned int *xdp_xmit)
1415 {
1416         struct virtnet_info *vi = rq->vq->vdev->priv;
1417         struct virtnet_rq_stats stats = {};
1418         unsigned int len;
1419         void *buf;
1420         int i;
1421
1422         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1423                 void *ctx;
1424
1425                 while (stats.packets < budget &&
1426                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1427                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1428                         stats.packets++;
1429                 }
1430         } else {
1431                 while (stats.packets < budget &&
1432                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1433                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1434                         stats.packets++;
1435                 }
1436         }
1437
1438         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1439                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1440                         schedule_delayed_work(&vi->refill, 0);
1441         }
1442
1443         u64_stats_update_begin(&rq->stats.syncp);
1444         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1445                 size_t offset = virtnet_rq_stats_desc[i].offset;
1446                 u64 *item;
1447
1448                 item = (u64 *)((u8 *)&rq->stats + offset);
1449                 *item += *(u64 *)((u8 *)&stats + offset);
1450         }
1451         u64_stats_update_end(&rq->stats.syncp);
1452
1453         return stats.packets;
1454 }
1455
1456 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1457 {
1458         unsigned int len;
1459         unsigned int packets = 0;
1460         unsigned int bytes = 0;
1461         void *ptr;
1462
1463         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1464                 if (likely(!is_xdp_frame(ptr))) {
1465                         struct sk_buff *skb = ptr;
1466
1467                         pr_debug("Sent skb %p\n", skb);
1468
1469                         bytes += skb->len;
1470                         napi_consume_skb(skb, in_napi);
1471                 } else {
1472                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1473
1474                         bytes += frame->len;
1475                         xdp_return_frame(frame);
1476                 }
1477                 packets++;
1478         }
1479
1480         /* Avoid overhead when no packets have been processed
1481          * happens when called speculatively from start_xmit.
1482          */
1483         if (!packets)
1484                 return;
1485
1486         u64_stats_update_begin(&sq->stats.syncp);
1487         sq->stats.bytes += bytes;
1488         sq->stats.packets += packets;
1489         u64_stats_update_end(&sq->stats.syncp);
1490 }
1491
1492 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1493 {
1494         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1495                 return false;
1496         else if (q < vi->curr_queue_pairs)
1497                 return true;
1498         else
1499                 return false;
1500 }
1501
1502 static void virtnet_poll_cleantx(struct receive_queue *rq)
1503 {
1504         struct virtnet_info *vi = rq->vq->vdev->priv;
1505         unsigned int index = vq2rxq(rq->vq);
1506         struct send_queue *sq = &vi->sq[index];
1507         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1508
1509         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1510                 return;
1511
1512         if (__netif_tx_trylock(txq)) {
1513                 do {
1514                         virtqueue_disable_cb(sq->vq);
1515                         free_old_xmit_skbs(sq, true);
1516                 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1517
1518                 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1519                         netif_tx_wake_queue(txq);
1520
1521                 __netif_tx_unlock(txq);
1522         }
1523 }
1524
1525 static int virtnet_poll(struct napi_struct *napi, int budget)
1526 {
1527         struct receive_queue *rq =
1528                 container_of(napi, struct receive_queue, napi);
1529         struct virtnet_info *vi = rq->vq->vdev->priv;
1530         struct send_queue *sq;
1531         unsigned int received;
1532         unsigned int xdp_xmit = 0;
1533
1534         virtnet_poll_cleantx(rq);
1535
1536         received = virtnet_receive(rq, budget, &xdp_xmit);
1537
1538         /* Out of packets? */
1539         if (received < budget)
1540                 virtqueue_napi_complete(napi, rq->vq, received);
1541
1542         if (xdp_xmit & VIRTIO_XDP_REDIR)
1543                 xdp_do_flush();
1544
1545         if (xdp_xmit & VIRTIO_XDP_TX) {
1546                 sq = virtnet_xdp_get_sq(vi);
1547                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1548                         u64_stats_update_begin(&sq->stats.syncp);
1549                         sq->stats.kicks++;
1550                         u64_stats_update_end(&sq->stats.syncp);
1551                 }
1552                 virtnet_xdp_put_sq(vi, sq);
1553         }
1554
1555         return received;
1556 }
1557
1558 static int virtnet_open(struct net_device *dev)
1559 {
1560         struct virtnet_info *vi = netdev_priv(dev);
1561         int i, err;
1562
1563         for (i = 0; i < vi->max_queue_pairs; i++) {
1564                 if (i < vi->curr_queue_pairs)
1565                         /* Make sure we have some buffers: if oom use wq. */
1566                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1567                                 schedule_delayed_work(&vi->refill, 0);
1568
1569                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1570                 if (err < 0)
1571                         return err;
1572
1573                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1574                                                  MEM_TYPE_PAGE_SHARED, NULL);
1575                 if (err < 0) {
1576                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1577                         return err;
1578                 }
1579
1580                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1581                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1582         }
1583
1584         return 0;
1585 }
1586
1587 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1588 {
1589         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1590         struct virtnet_info *vi = sq->vq->vdev->priv;
1591         unsigned int index = vq2txq(sq->vq);
1592         struct netdev_queue *txq;
1593         int opaque;
1594         bool done;
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         virtqueue_disable_cb(sq->vq);
1605         free_old_xmit_skbs(sq, true);
1606
1607         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1608                 netif_tx_wake_queue(txq);
1609
1610         opaque = virtqueue_enable_cb_prepare(sq->vq);
1611
1612         done = napi_complete_done(napi, 0);
1613
1614         if (!done)
1615                 virtqueue_disable_cb(sq->vq);
1616
1617         __netif_tx_unlock(txq);
1618
1619         if (done) {
1620                 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1621                         if (napi_schedule_prep(napi)) {
1622                                 __netif_tx_lock(txq, raw_smp_processor_id());
1623                                 virtqueue_disable_cb(sq->vq);
1624                                 __netif_tx_unlock(txq);
1625                                 __napi_schedule(napi);
1626                         }
1627                 }
1628         }
1629
1630         return 0;
1631 }
1632
1633 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1634 {
1635         struct virtio_net_hdr_mrg_rxbuf *hdr;
1636         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1637         struct virtnet_info *vi = sq->vq->vdev->priv;
1638         int num_sg;
1639         unsigned hdr_len = vi->hdr_len;
1640         bool can_push;
1641
1642         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1643
1644         can_push = vi->any_header_sg &&
1645                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1646                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1647         /* Even if we can, don't push here yet as this would skew
1648          * csum_start offset below. */
1649         if (can_push)
1650                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1651         else
1652                 hdr = skb_vnet_hdr(skb);
1653
1654         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1655                                     virtio_is_little_endian(vi->vdev), false,
1656                                     0))
1657                 return -EPROTO;
1658
1659         if (vi->mergeable_rx_bufs)
1660                 hdr->num_buffers = 0;
1661
1662         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1663         if (can_push) {
1664                 __skb_push(skb, hdr_len);
1665                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1666                 if (unlikely(num_sg < 0))
1667                         return num_sg;
1668                 /* Pull header back to avoid skew in tx bytes calculations. */
1669                 __skb_pull(skb, hdr_len);
1670         } else {
1671                 sg_set_buf(sq->sg, hdr, hdr_len);
1672                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1673                 if (unlikely(num_sg < 0))
1674                         return num_sg;
1675                 num_sg++;
1676         }
1677         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1678 }
1679
1680 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1681 {
1682         struct virtnet_info *vi = netdev_priv(dev);
1683         int qnum = skb_get_queue_mapping(skb);
1684         struct send_queue *sq = &vi->sq[qnum];
1685         int err;
1686         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1687         bool kick = !netdev_xmit_more();
1688         bool use_napi = sq->napi.weight;
1689
1690         /* Free up any pending old buffers before queueing new ones. */
1691         do {
1692                 if (use_napi)
1693                         virtqueue_disable_cb(sq->vq);
1694
1695                 free_old_xmit_skbs(sq, false);
1696
1697         } while (use_napi && kick &&
1698                unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1699
1700         /* timestamp packet in software */
1701         skb_tx_timestamp(skb);
1702
1703         /* Try to transmit */
1704         err = xmit_skb(sq, skb);
1705
1706         /* This should not happen! */
1707         if (unlikely(err)) {
1708                 dev->stats.tx_fifo_errors++;
1709                 if (net_ratelimit())
1710                         dev_warn(&dev->dev,
1711                                  "Unexpected TXQ (%d) queue failure: %d\n",
1712                                  qnum, err);
1713                 dev->stats.tx_dropped++;
1714                 dev_kfree_skb_any(skb);
1715                 return NETDEV_TX_OK;
1716         }
1717
1718         /* Don't wait up for transmitted skbs to be freed. */
1719         if (!use_napi) {
1720                 skb_orphan(skb);
1721                 nf_reset_ct(skb);
1722         }
1723
1724         /* If running out of space, stop queue to avoid getting packets that we
1725          * are then unable to transmit.
1726          * An alternative would be to force queuing layer to requeue the skb by
1727          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1728          * returned in a normal path of operation: it means that driver is not
1729          * maintaining the TX queue stop/start state properly, and causes
1730          * the stack to do a non-trivial amount of useless work.
1731          * Since most packets only take 1 or 2 ring slots, stopping the queue
1732          * early means 16 slots are typically wasted.
1733          */
1734         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1735                 netif_stop_subqueue(dev, qnum);
1736                 if (!use_napi &&
1737                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1738                         /* More just got used, free them then recheck. */
1739                         free_old_xmit_skbs(sq, false);
1740                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1741                                 netif_start_subqueue(dev, qnum);
1742                                 virtqueue_disable_cb(sq->vq);
1743                         }
1744                 }
1745         }
1746
1747         if (kick || netif_xmit_stopped(txq)) {
1748                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1749                         u64_stats_update_begin(&sq->stats.syncp);
1750                         sq->stats.kicks++;
1751                         u64_stats_update_end(&sq->stats.syncp);
1752                 }
1753         }
1754
1755         return NETDEV_TX_OK;
1756 }
1757
1758 /*
1759  * Send command via the control virtqueue and check status.  Commands
1760  * supported by the hypervisor, as indicated by feature bits, should
1761  * never fail unless improperly formatted.
1762  */
1763 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1764                                  struct scatterlist *out)
1765 {
1766         struct scatterlist *sgs[4], hdr, stat;
1767         unsigned out_num = 0, tmp;
1768         int ret;
1769
1770         /* Caller should know better */
1771         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1772
1773         vi->ctrl->status = ~0;
1774         vi->ctrl->hdr.class = class;
1775         vi->ctrl->hdr.cmd = cmd;
1776         /* Add header */
1777         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1778         sgs[out_num++] = &hdr;
1779
1780         if (out)
1781                 sgs[out_num++] = out;
1782
1783         /* Add return status. */
1784         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1785         sgs[out_num] = &stat;
1786
1787         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1788         ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1789         if (ret < 0) {
1790                 dev_warn(&vi->vdev->dev,
1791                          "Failed to add sgs for command vq: %d\n.", ret);
1792                 return false;
1793         }
1794
1795         if (unlikely(!virtqueue_kick(vi->cvq)))
1796                 return vi->ctrl->status == VIRTIO_NET_OK;
1797
1798         /* Spin for a response, the kick causes an ioport write, trapping
1799          * into the hypervisor, so the request should be handled immediately.
1800          */
1801         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1802                !virtqueue_is_broken(vi->cvq))
1803                 cpu_relax();
1804
1805         return vi->ctrl->status == VIRTIO_NET_OK;
1806 }
1807
1808 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1809 {
1810         struct virtnet_info *vi = netdev_priv(dev);
1811         struct virtio_device *vdev = vi->vdev;
1812         int ret;
1813         struct sockaddr *addr;
1814         struct scatterlist sg;
1815
1816         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1817                 return -EOPNOTSUPP;
1818
1819         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1820         if (!addr)
1821                 return -ENOMEM;
1822
1823         ret = eth_prepare_mac_addr_change(dev, addr);
1824         if (ret)
1825                 goto out;
1826
1827         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1828                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1829                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1830                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1831                         dev_warn(&vdev->dev,
1832                                  "Failed to set mac address by vq command.\n");
1833                         ret = -EINVAL;
1834                         goto out;
1835                 }
1836         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1837                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1838                 unsigned int i;
1839
1840                 /* Naturally, this has an atomicity problem. */
1841                 for (i = 0; i < dev->addr_len; i++)
1842                         virtio_cwrite8(vdev,
1843                                        offsetof(struct virtio_net_config, mac) +
1844                                        i, addr->sa_data[i]);
1845         }
1846
1847         eth_commit_mac_addr_change(dev, p);
1848         ret = 0;
1849
1850 out:
1851         kfree(addr);
1852         return ret;
1853 }
1854
1855 static void virtnet_stats(struct net_device *dev,
1856                           struct rtnl_link_stats64 *tot)
1857 {
1858         struct virtnet_info *vi = netdev_priv(dev);
1859         unsigned int start;
1860         int i;
1861
1862         for (i = 0; i < vi->max_queue_pairs; i++) {
1863                 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1864                 struct receive_queue *rq = &vi->rq[i];
1865                 struct send_queue *sq = &vi->sq[i];
1866
1867                 do {
1868                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1869                         tpackets = sq->stats.packets;
1870                         tbytes   = sq->stats.bytes;
1871                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1872
1873                 do {
1874                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1875                         rpackets = rq->stats.packets;
1876                         rbytes   = rq->stats.bytes;
1877                         rdrops   = rq->stats.drops;
1878                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1879
1880                 tot->rx_packets += rpackets;
1881                 tot->tx_packets += tpackets;
1882                 tot->rx_bytes   += rbytes;
1883                 tot->tx_bytes   += tbytes;
1884                 tot->rx_dropped += rdrops;
1885         }
1886
1887         tot->tx_dropped = dev->stats.tx_dropped;
1888         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1889         tot->rx_length_errors = dev->stats.rx_length_errors;
1890         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1891 }
1892
1893 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1894 {
1895         rtnl_lock();
1896         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1897                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1898                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1899         rtnl_unlock();
1900 }
1901
1902 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1903 {
1904         struct scatterlist sg;
1905         struct net_device *dev = vi->dev;
1906
1907         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1908                 return 0;
1909
1910         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1911         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1912
1913         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1914                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1915                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1916                          queue_pairs);
1917                 return -EINVAL;
1918         } else {
1919                 vi->curr_queue_pairs = queue_pairs;
1920                 /* virtnet_open() will refill when device is going to up. */
1921                 if (dev->flags & IFF_UP)
1922                         schedule_delayed_work(&vi->refill, 0);
1923         }
1924
1925         return 0;
1926 }
1927
1928 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1929 {
1930         int err;
1931
1932         rtnl_lock();
1933         err = _virtnet_set_queues(vi, queue_pairs);
1934         rtnl_unlock();
1935         return err;
1936 }
1937
1938 static int virtnet_close(struct net_device *dev)
1939 {
1940         struct virtnet_info *vi = netdev_priv(dev);
1941         int i;
1942
1943         /* Make sure refill_work doesn't re-enable napi! */
1944         cancel_delayed_work_sync(&vi->refill);
1945
1946         for (i = 0; i < vi->max_queue_pairs; i++) {
1947                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1948                 napi_disable(&vi->rq[i].napi);
1949                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1950         }
1951
1952         return 0;
1953 }
1954
1955 static void virtnet_set_rx_mode(struct net_device *dev)
1956 {
1957         struct virtnet_info *vi = netdev_priv(dev);
1958         struct scatterlist sg[2];
1959         struct virtio_net_ctrl_mac *mac_data;
1960         struct netdev_hw_addr *ha;
1961         int uc_count;
1962         int mc_count;
1963         void *buf;
1964         int i;
1965
1966         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1967         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1968                 return;
1969
1970         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1971         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1972
1973         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1974
1975         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1976                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1977                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1978                          vi->ctrl->promisc ? "en" : "dis");
1979
1980         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1981
1982         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1983                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1984                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1985                          vi->ctrl->allmulti ? "en" : "dis");
1986
1987         uc_count = netdev_uc_count(dev);
1988         mc_count = netdev_mc_count(dev);
1989         /* MAC filter - use one buffer for both lists */
1990         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1991                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1992         mac_data = buf;
1993         if (!buf)
1994                 return;
1995
1996         sg_init_table(sg, 2);
1997
1998         /* Store the unicast list and count in the front of the buffer */
1999         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2000         i = 0;
2001         netdev_for_each_uc_addr(ha, dev)
2002                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2003
2004         sg_set_buf(&sg[0], mac_data,
2005                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2006
2007         /* multicast list and count fill the end */
2008         mac_data = (void *)&mac_data->macs[uc_count][0];
2009
2010         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2011         i = 0;
2012         netdev_for_each_mc_addr(ha, dev)
2013                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2014
2015         sg_set_buf(&sg[1], mac_data,
2016                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2017
2018         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2019                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2020                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2021
2022         kfree(buf);
2023 }
2024
2025 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2026                                    __be16 proto, u16 vid)
2027 {
2028         struct virtnet_info *vi = netdev_priv(dev);
2029         struct scatterlist sg;
2030
2031         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2032         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2033
2034         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2035                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2036                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2037         return 0;
2038 }
2039
2040 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2041                                     __be16 proto, u16 vid)
2042 {
2043         struct virtnet_info *vi = netdev_priv(dev);
2044         struct scatterlist sg;
2045
2046         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2047         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2048
2049         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2050                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2051                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2052         return 0;
2053 }
2054
2055 static void virtnet_clean_affinity(struct virtnet_info *vi)
2056 {
2057         int i;
2058
2059         if (vi->affinity_hint_set) {
2060                 for (i = 0; i < vi->max_queue_pairs; i++) {
2061                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
2062                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
2063                 }
2064
2065                 vi->affinity_hint_set = false;
2066         }
2067 }
2068
2069 static void virtnet_set_affinity(struct virtnet_info *vi)
2070 {
2071         cpumask_var_t mask;
2072         int stragglers;
2073         int group_size;
2074         int i, j, cpu;
2075         int num_cpu;
2076         int stride;
2077
2078         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2079                 virtnet_clean_affinity(vi);
2080                 return;
2081         }
2082
2083         num_cpu = num_online_cpus();
2084         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2085         stragglers = num_cpu >= vi->curr_queue_pairs ?
2086                         num_cpu % vi->curr_queue_pairs :
2087                         0;
2088         cpu = cpumask_next(-1, cpu_online_mask);
2089
2090         for (i = 0; i < vi->curr_queue_pairs; i++) {
2091                 group_size = stride + (i < stragglers ? 1 : 0);
2092
2093                 for (j = 0; j < group_size; j++) {
2094                         cpumask_set_cpu(cpu, mask);
2095                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2096                                                 nr_cpu_ids, false);
2097                 }
2098                 virtqueue_set_affinity(vi->rq[i].vq, mask);
2099                 virtqueue_set_affinity(vi->sq[i].vq, mask);
2100                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2101                 cpumask_clear(mask);
2102         }
2103
2104         vi->affinity_hint_set = true;
2105         free_cpumask_var(mask);
2106 }
2107
2108 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2109 {
2110         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2111                                                    node);
2112         virtnet_set_affinity(vi);
2113         return 0;
2114 }
2115
2116 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2117 {
2118         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2119                                                    node_dead);
2120         virtnet_set_affinity(vi);
2121         return 0;
2122 }
2123
2124 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2125 {
2126         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2127                                                    node);
2128
2129         virtnet_clean_affinity(vi);
2130         return 0;
2131 }
2132
2133 static enum cpuhp_state virtionet_online;
2134
2135 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2136 {
2137         int ret;
2138
2139         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2140         if (ret)
2141                 return ret;
2142         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2143                                                &vi->node_dead);
2144         if (!ret)
2145                 return ret;
2146         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2147         return ret;
2148 }
2149
2150 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2151 {
2152         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2153         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2154                                             &vi->node_dead);
2155 }
2156
2157 static void virtnet_get_ringparam(struct net_device *dev,
2158                                 struct ethtool_ringparam *ring)
2159 {
2160         struct virtnet_info *vi = netdev_priv(dev);
2161
2162         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2163         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2164         ring->rx_pending = ring->rx_max_pending;
2165         ring->tx_pending = ring->tx_max_pending;
2166 }
2167
2168
2169 static void virtnet_get_drvinfo(struct net_device *dev,
2170                                 struct ethtool_drvinfo *info)
2171 {
2172         struct virtnet_info *vi = netdev_priv(dev);
2173         struct virtio_device *vdev = vi->vdev;
2174
2175         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2176         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2177         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2178
2179 }
2180
2181 /* TODO: Eliminate OOO packets during switching */
2182 static int virtnet_set_channels(struct net_device *dev,
2183                                 struct ethtool_channels *channels)
2184 {
2185         struct virtnet_info *vi = netdev_priv(dev);
2186         u16 queue_pairs = channels->combined_count;
2187         int err;
2188
2189         /* We don't support separate rx/tx channels.
2190          * We don't allow setting 'other' channels.
2191          */
2192         if (channels->rx_count || channels->tx_count || channels->other_count)
2193                 return -EINVAL;
2194
2195         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2196                 return -EINVAL;
2197
2198         /* For now we don't support modifying channels while XDP is loaded
2199          * also when XDP is loaded all RX queues have XDP programs so we only
2200          * need to check a single RX queue.
2201          */
2202         if (vi->rq[0].xdp_prog)
2203                 return -EINVAL;
2204
2205         cpus_read_lock();
2206         err = _virtnet_set_queues(vi, queue_pairs);
2207         if (err) {
2208                 cpus_read_unlock();
2209                 goto err;
2210         }
2211         virtnet_set_affinity(vi);
2212         cpus_read_unlock();
2213
2214         netif_set_real_num_tx_queues(dev, queue_pairs);
2215         netif_set_real_num_rx_queues(dev, queue_pairs);
2216  err:
2217         return err;
2218 }
2219
2220 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2221 {
2222         struct virtnet_info *vi = netdev_priv(dev);
2223         unsigned int i, j;
2224         u8 *p = data;
2225
2226         switch (stringset) {
2227         case ETH_SS_STATS:
2228                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2229                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2230                                 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2231                                                 virtnet_rq_stats_desc[j].desc);
2232                 }
2233
2234                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2235                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2236                                 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2237                                                 virtnet_sq_stats_desc[j].desc);
2238                 }
2239                 break;
2240         }
2241 }
2242
2243 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2244 {
2245         struct virtnet_info *vi = netdev_priv(dev);
2246
2247         switch (sset) {
2248         case ETH_SS_STATS:
2249                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2250                                                VIRTNET_SQ_STATS_LEN);
2251         default:
2252                 return -EOPNOTSUPP;
2253         }
2254 }
2255
2256 static void virtnet_get_ethtool_stats(struct net_device *dev,
2257                                       struct ethtool_stats *stats, u64 *data)
2258 {
2259         struct virtnet_info *vi = netdev_priv(dev);
2260         unsigned int idx = 0, start, i, j;
2261         const u8 *stats_base;
2262         size_t offset;
2263
2264         for (i = 0; i < vi->curr_queue_pairs; i++) {
2265                 struct receive_queue *rq = &vi->rq[i];
2266
2267                 stats_base = (u8 *)&rq->stats;
2268                 do {
2269                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2270                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2271                                 offset = virtnet_rq_stats_desc[j].offset;
2272                                 data[idx + j] = *(u64 *)(stats_base + offset);
2273                         }
2274                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2275                 idx += VIRTNET_RQ_STATS_LEN;
2276         }
2277
2278         for (i = 0; i < vi->curr_queue_pairs; i++) {
2279                 struct send_queue *sq = &vi->sq[i];
2280
2281                 stats_base = (u8 *)&sq->stats;
2282                 do {
2283                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2284                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2285                                 offset = virtnet_sq_stats_desc[j].offset;
2286                                 data[idx + j] = *(u64 *)(stats_base + offset);
2287                         }
2288                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2289                 idx += VIRTNET_SQ_STATS_LEN;
2290         }
2291 }
2292
2293 static void virtnet_get_channels(struct net_device *dev,
2294                                  struct ethtool_channels *channels)
2295 {
2296         struct virtnet_info *vi = netdev_priv(dev);
2297
2298         channels->combined_count = vi->curr_queue_pairs;
2299         channels->max_combined = vi->max_queue_pairs;
2300         channels->max_other = 0;
2301         channels->rx_count = 0;
2302         channels->tx_count = 0;
2303         channels->other_count = 0;
2304 }
2305
2306 static int virtnet_set_link_ksettings(struct net_device *dev,
2307                                       const struct ethtool_link_ksettings *cmd)
2308 {
2309         struct virtnet_info *vi = netdev_priv(dev);
2310
2311         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2312                                                   &vi->speed, &vi->duplex);
2313 }
2314
2315 static int virtnet_get_link_ksettings(struct net_device *dev,
2316                                       struct ethtool_link_ksettings *cmd)
2317 {
2318         struct virtnet_info *vi = netdev_priv(dev);
2319
2320         cmd->base.speed = vi->speed;
2321         cmd->base.duplex = vi->duplex;
2322         cmd->base.port = PORT_OTHER;
2323
2324         return 0;
2325 }
2326
2327 static int virtnet_set_coalesce(struct net_device *dev,
2328                                 struct ethtool_coalesce *ec,
2329                                 struct kernel_ethtool_coalesce *kernel_coal,
2330                                 struct netlink_ext_ack *extack)
2331 {
2332         struct virtnet_info *vi = netdev_priv(dev);
2333         int i, napi_weight;
2334
2335         if (ec->tx_max_coalesced_frames > 1 ||
2336             ec->rx_max_coalesced_frames != 1)
2337                 return -EINVAL;
2338
2339         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2340         if (napi_weight ^ vi->sq[0].napi.weight) {
2341                 if (dev->flags & IFF_UP)
2342                         return -EBUSY;
2343                 for (i = 0; i < vi->max_queue_pairs; i++)
2344                         vi->sq[i].napi.weight = napi_weight;
2345         }
2346
2347         return 0;
2348 }
2349
2350 static int virtnet_get_coalesce(struct net_device *dev,
2351                                 struct ethtool_coalesce *ec,
2352                                 struct kernel_ethtool_coalesce *kernel_coal,
2353                                 struct netlink_ext_ack *extack)
2354 {
2355         struct ethtool_coalesce ec_default = {
2356                 .cmd = ETHTOOL_GCOALESCE,
2357                 .rx_max_coalesced_frames = 1,
2358         };
2359         struct virtnet_info *vi = netdev_priv(dev);
2360
2361         memcpy(ec, &ec_default, sizeof(ec_default));
2362
2363         if (vi->sq[0].napi.weight)
2364                 ec->tx_max_coalesced_frames = 1;
2365
2366         return 0;
2367 }
2368
2369 static void virtnet_init_settings(struct net_device *dev)
2370 {
2371         struct virtnet_info *vi = netdev_priv(dev);
2372
2373         vi->speed = SPEED_UNKNOWN;
2374         vi->duplex = DUPLEX_UNKNOWN;
2375 }
2376
2377 static void virtnet_update_settings(struct virtnet_info *vi)
2378 {
2379         u32 speed;
2380         u8 duplex;
2381
2382         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2383                 return;
2384
2385         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2386
2387         if (ethtool_validate_speed(speed))
2388                 vi->speed = speed;
2389
2390         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2391
2392         if (ethtool_validate_duplex(duplex))
2393                 vi->duplex = duplex;
2394 }
2395
2396 static const struct ethtool_ops virtnet_ethtool_ops = {
2397         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2398         .get_drvinfo = virtnet_get_drvinfo,
2399         .get_link = ethtool_op_get_link,
2400         .get_ringparam = virtnet_get_ringparam,
2401         .get_strings = virtnet_get_strings,
2402         .get_sset_count = virtnet_get_sset_count,
2403         .get_ethtool_stats = virtnet_get_ethtool_stats,
2404         .set_channels = virtnet_set_channels,
2405         .get_channels = virtnet_get_channels,
2406         .get_ts_info = ethtool_op_get_ts_info,
2407         .get_link_ksettings = virtnet_get_link_ksettings,
2408         .set_link_ksettings = virtnet_set_link_ksettings,
2409         .set_coalesce = virtnet_set_coalesce,
2410         .get_coalesce = virtnet_get_coalesce,
2411 };
2412
2413 static void virtnet_freeze_down(struct virtio_device *vdev)
2414 {
2415         struct virtnet_info *vi = vdev->priv;
2416         int i;
2417
2418         /* Make sure no work handler is accessing the device */
2419         flush_work(&vi->config_work);
2420
2421         netif_tx_lock_bh(vi->dev);
2422         netif_device_detach(vi->dev);
2423         netif_tx_unlock_bh(vi->dev);
2424         cancel_delayed_work_sync(&vi->refill);
2425
2426         if (netif_running(vi->dev)) {
2427                 for (i = 0; i < vi->max_queue_pairs; i++) {
2428                         napi_disable(&vi->rq[i].napi);
2429                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2430                 }
2431         }
2432 }
2433
2434 static int init_vqs(struct virtnet_info *vi);
2435
2436 static int virtnet_restore_up(struct virtio_device *vdev)
2437 {
2438         struct virtnet_info *vi = vdev->priv;
2439         int err, i;
2440
2441         err = init_vqs(vi);
2442         if (err)
2443                 return err;
2444
2445         virtio_device_ready(vdev);
2446
2447         if (netif_running(vi->dev)) {
2448                 for (i = 0; i < vi->curr_queue_pairs; i++)
2449                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2450                                 schedule_delayed_work(&vi->refill, 0);
2451
2452                 for (i = 0; i < vi->max_queue_pairs; i++) {
2453                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2454                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2455                                                &vi->sq[i].napi);
2456                 }
2457         }
2458
2459         netif_tx_lock_bh(vi->dev);
2460         netif_device_attach(vi->dev);
2461         netif_tx_unlock_bh(vi->dev);
2462         return err;
2463 }
2464
2465 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2466 {
2467         struct scatterlist sg;
2468         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2469
2470         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2471
2472         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2473                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2474                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2475                 return -EINVAL;
2476         }
2477
2478         return 0;
2479 }
2480
2481 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2482 {
2483         u64 offloads = 0;
2484
2485         if (!vi->guest_offloads)
2486                 return 0;
2487
2488         return virtnet_set_guest_offloads(vi, offloads);
2489 }
2490
2491 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2492 {
2493         u64 offloads = vi->guest_offloads;
2494
2495         if (!vi->guest_offloads)
2496                 return 0;
2497
2498         return virtnet_set_guest_offloads(vi, offloads);
2499 }
2500
2501 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2502                            struct netlink_ext_ack *extack)
2503 {
2504         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2505         struct virtnet_info *vi = netdev_priv(dev);
2506         struct bpf_prog *old_prog;
2507         u16 xdp_qp = 0, curr_qp;
2508         int i, err;
2509
2510         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2511             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2512                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2513                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2514                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2515                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2516                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
2517                 return -EOPNOTSUPP;
2518         }
2519
2520         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2521                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2522                 return -EINVAL;
2523         }
2524
2525         if (dev->mtu > max_sz) {
2526                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2527                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2528                 return -EINVAL;
2529         }
2530
2531         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2532         if (prog)
2533                 xdp_qp = nr_cpu_ids;
2534
2535         /* XDP requires extra queues for XDP_TX */
2536         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2537                 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",
2538                             curr_qp + xdp_qp, vi->max_queue_pairs);
2539                 xdp_qp = 0;
2540         }
2541
2542         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2543         if (!prog && !old_prog)
2544                 return 0;
2545
2546         if (prog)
2547                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2548
2549         /* Make sure NAPI is not using any XDP TX queues for RX. */
2550         if (netif_running(dev)) {
2551                 for (i = 0; i < vi->max_queue_pairs; i++) {
2552                         napi_disable(&vi->rq[i].napi);
2553                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2554                 }
2555         }
2556
2557         if (!prog) {
2558                 for (i = 0; i < vi->max_queue_pairs; i++) {
2559                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2560                         if (i == 0)
2561                                 virtnet_restore_guest_offloads(vi);
2562                 }
2563                 synchronize_net();
2564         }
2565
2566         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2567         if (err)
2568                 goto err;
2569         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2570         vi->xdp_queue_pairs = xdp_qp;
2571
2572         if (prog) {
2573                 vi->xdp_enabled = true;
2574                 for (i = 0; i < vi->max_queue_pairs; i++) {
2575                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2576                         if (i == 0 && !old_prog)
2577                                 virtnet_clear_guest_offloads(vi);
2578                 }
2579         } else {
2580                 vi->xdp_enabled = false;
2581         }
2582
2583         for (i = 0; i < vi->max_queue_pairs; i++) {
2584                 if (old_prog)
2585                         bpf_prog_put(old_prog);
2586                 if (netif_running(dev)) {
2587                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2588                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2589                                                &vi->sq[i].napi);
2590                 }
2591         }
2592
2593         return 0;
2594
2595 err:
2596         if (!prog) {
2597                 virtnet_clear_guest_offloads(vi);
2598                 for (i = 0; i < vi->max_queue_pairs; i++)
2599                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2600         }
2601
2602         if (netif_running(dev)) {
2603                 for (i = 0; i < vi->max_queue_pairs; i++) {
2604                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2605                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2606                                                &vi->sq[i].napi);
2607                 }
2608         }
2609         if (prog)
2610                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2611         return err;
2612 }
2613
2614 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2615 {
2616         switch (xdp->command) {
2617         case XDP_SETUP_PROG:
2618                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2619         default:
2620                 return -EINVAL;
2621         }
2622 }
2623
2624 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2625                                       size_t len)
2626 {
2627         struct virtnet_info *vi = netdev_priv(dev);
2628         int ret;
2629
2630         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2631                 return -EOPNOTSUPP;
2632
2633         ret = snprintf(buf, len, "sby");
2634         if (ret >= len)
2635                 return -EOPNOTSUPP;
2636
2637         return 0;
2638 }
2639
2640 static int virtnet_set_features(struct net_device *dev,
2641                                 netdev_features_t features)
2642 {
2643         struct virtnet_info *vi = netdev_priv(dev);
2644         u64 offloads;
2645         int err;
2646
2647         if ((dev->features ^ features) & NETIF_F_GRO_HW) {
2648                 if (vi->xdp_enabled)
2649                         return -EBUSY;
2650
2651                 if (features & NETIF_F_GRO_HW)
2652                         offloads = vi->guest_offloads_capable;
2653                 else
2654                         offloads = vi->guest_offloads_capable &
2655                                    ~GUEST_OFFLOAD_GRO_HW_MASK;
2656
2657                 err = virtnet_set_guest_offloads(vi, offloads);
2658                 if (err)
2659                         return err;
2660                 vi->guest_offloads = offloads;
2661         }
2662
2663         return 0;
2664 }
2665
2666 static const struct net_device_ops virtnet_netdev = {
2667         .ndo_open            = virtnet_open,
2668         .ndo_stop            = virtnet_close,
2669         .ndo_start_xmit      = start_xmit,
2670         .ndo_validate_addr   = eth_validate_addr,
2671         .ndo_set_mac_address = virtnet_set_mac_address,
2672         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2673         .ndo_get_stats64     = virtnet_stats,
2674         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2675         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2676         .ndo_bpf                = virtnet_xdp,
2677         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2678         .ndo_features_check     = passthru_features_check,
2679         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2680         .ndo_set_features       = virtnet_set_features,
2681 };
2682
2683 static void virtnet_config_changed_work(struct work_struct *work)
2684 {
2685         struct virtnet_info *vi =
2686                 container_of(work, struct virtnet_info, config_work);
2687         u16 v;
2688
2689         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2690                                  struct virtio_net_config, status, &v) < 0)
2691                 return;
2692
2693         if (v & VIRTIO_NET_S_ANNOUNCE) {
2694                 netdev_notify_peers(vi->dev);
2695                 virtnet_ack_link_announce(vi);
2696         }
2697
2698         /* Ignore unknown (future) status bits */
2699         v &= VIRTIO_NET_S_LINK_UP;
2700
2701         if (vi->status == v)
2702                 return;
2703
2704         vi->status = v;
2705
2706         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2707                 virtnet_update_settings(vi);
2708                 netif_carrier_on(vi->dev);
2709                 netif_tx_wake_all_queues(vi->dev);
2710         } else {
2711                 netif_carrier_off(vi->dev);
2712                 netif_tx_stop_all_queues(vi->dev);
2713         }
2714 }
2715
2716 static void virtnet_config_changed(struct virtio_device *vdev)
2717 {
2718         struct virtnet_info *vi = vdev->priv;
2719
2720         schedule_work(&vi->config_work);
2721 }
2722
2723 static void virtnet_free_queues(struct virtnet_info *vi)
2724 {
2725         int i;
2726
2727         for (i = 0; i < vi->max_queue_pairs; i++) {
2728                 __netif_napi_del(&vi->rq[i].napi);
2729                 __netif_napi_del(&vi->sq[i].napi);
2730         }
2731
2732         /* We called __netif_napi_del(),
2733          * we need to respect an RCU grace period before freeing vi->rq
2734          */
2735         synchronize_net();
2736
2737         kfree(vi->rq);
2738         kfree(vi->sq);
2739         kfree(vi->ctrl);
2740 }
2741
2742 static void _free_receive_bufs(struct virtnet_info *vi)
2743 {
2744         struct bpf_prog *old_prog;
2745         int i;
2746
2747         for (i = 0; i < vi->max_queue_pairs; i++) {
2748                 while (vi->rq[i].pages)
2749                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2750
2751                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2752                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2753                 if (old_prog)
2754                         bpf_prog_put(old_prog);
2755         }
2756 }
2757
2758 static void free_receive_bufs(struct virtnet_info *vi)
2759 {
2760         rtnl_lock();
2761         _free_receive_bufs(vi);
2762         rtnl_unlock();
2763 }
2764
2765 static void free_receive_page_frags(struct virtnet_info *vi)
2766 {
2767         int i;
2768         for (i = 0; i < vi->max_queue_pairs; i++)
2769                 if (vi->rq[i].alloc_frag.page)
2770                         put_page(vi->rq[i].alloc_frag.page);
2771 }
2772
2773 static void free_unused_bufs(struct virtnet_info *vi)
2774 {
2775         void *buf;
2776         int i;
2777
2778         for (i = 0; i < vi->max_queue_pairs; i++) {
2779                 struct virtqueue *vq = vi->sq[i].vq;
2780                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2781                         if (!is_xdp_frame(buf))
2782                                 dev_kfree_skb(buf);
2783                         else
2784                                 xdp_return_frame(ptr_to_xdp(buf));
2785                 }
2786         }
2787
2788         for (i = 0; i < vi->max_queue_pairs; i++) {
2789                 struct virtqueue *vq = vi->rq[i].vq;
2790
2791                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2792                         if (vi->mergeable_rx_bufs) {
2793                                 put_page(virt_to_head_page(buf));
2794                         } else if (vi->big_packets) {
2795                                 give_pages(&vi->rq[i], buf);
2796                         } else {
2797                                 put_page(virt_to_head_page(buf));
2798                         }
2799                 }
2800         }
2801 }
2802
2803 static void virtnet_del_vqs(struct virtnet_info *vi)
2804 {
2805         struct virtio_device *vdev = vi->vdev;
2806
2807         virtnet_clean_affinity(vi);
2808
2809         vdev->config->del_vqs(vdev);
2810
2811         virtnet_free_queues(vi);
2812 }
2813
2814 /* How large should a single buffer be so a queue full of these can fit at
2815  * least one full packet?
2816  * Logic below assumes the mergeable buffer header is used.
2817  */
2818 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2819 {
2820         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2821         unsigned int rq_size = virtqueue_get_vring_size(vq);
2822         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2823         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2824         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2825
2826         return max(max(min_buf_len, hdr_len) - hdr_len,
2827                    (unsigned int)GOOD_PACKET_LEN);
2828 }
2829
2830 static int virtnet_find_vqs(struct virtnet_info *vi)
2831 {
2832         vq_callback_t **callbacks;
2833         struct virtqueue **vqs;
2834         int ret = -ENOMEM;
2835         int i, total_vqs;
2836         const char **names;
2837         bool *ctx;
2838
2839         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2840          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2841          * possible control vq.
2842          */
2843         total_vqs = vi->max_queue_pairs * 2 +
2844                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2845
2846         /* Allocate space for find_vqs parameters */
2847         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2848         if (!vqs)
2849                 goto err_vq;
2850         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2851         if (!callbacks)
2852                 goto err_callback;
2853         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2854         if (!names)
2855                 goto err_names;
2856         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2857                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2858                 if (!ctx)
2859                         goto err_ctx;
2860         } else {
2861                 ctx = NULL;
2862         }
2863
2864         /* Parameters for control virtqueue, if any */
2865         if (vi->has_cvq) {
2866                 callbacks[total_vqs - 1] = NULL;
2867                 names[total_vqs - 1] = "control";
2868         }
2869
2870         /* Allocate/initialize parameters for send/receive virtqueues */
2871         for (i = 0; i < vi->max_queue_pairs; i++) {
2872                 callbacks[rxq2vq(i)] = skb_recv_done;
2873                 callbacks[txq2vq(i)] = skb_xmit_done;
2874                 sprintf(vi->rq[i].name, "input.%d", i);
2875                 sprintf(vi->sq[i].name, "output.%d", i);
2876                 names[rxq2vq(i)] = vi->rq[i].name;
2877                 names[txq2vq(i)] = vi->sq[i].name;
2878                 if (ctx)
2879                         ctx[rxq2vq(i)] = true;
2880         }
2881
2882         ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
2883                                   names, ctx, NULL);
2884         if (ret)
2885                 goto err_find;
2886
2887         if (vi->has_cvq) {
2888                 vi->cvq = vqs[total_vqs - 1];
2889                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2890                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2891         }
2892
2893         for (i = 0; i < vi->max_queue_pairs; i++) {
2894                 vi->rq[i].vq = vqs[rxq2vq(i)];
2895                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2896                 vi->sq[i].vq = vqs[txq2vq(i)];
2897         }
2898
2899         /* run here: ret == 0. */
2900
2901
2902 err_find:
2903         kfree(ctx);
2904 err_ctx:
2905         kfree(names);
2906 err_names:
2907         kfree(callbacks);
2908 err_callback:
2909         kfree(vqs);
2910 err_vq:
2911         return ret;
2912 }
2913
2914 static int virtnet_alloc_queues(struct virtnet_info *vi)
2915 {
2916         int i;
2917
2918         if (vi->has_cvq) {
2919                 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2920                 if (!vi->ctrl)
2921                         goto err_ctrl;
2922         } else {
2923                 vi->ctrl = NULL;
2924         }
2925         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2926         if (!vi->sq)
2927                 goto err_sq;
2928         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2929         if (!vi->rq)
2930                 goto err_rq;
2931
2932         INIT_DELAYED_WORK(&vi->refill, refill_work);
2933         for (i = 0; i < vi->max_queue_pairs; i++) {
2934                 vi->rq[i].pages = NULL;
2935                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2936                                napi_weight);
2937                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2938                                   napi_tx ? napi_weight : 0);
2939
2940                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2941                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2942                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2943
2944                 u64_stats_init(&vi->rq[i].stats.syncp);
2945                 u64_stats_init(&vi->sq[i].stats.syncp);
2946         }
2947
2948         return 0;
2949
2950 err_rq:
2951         kfree(vi->sq);
2952 err_sq:
2953         kfree(vi->ctrl);
2954 err_ctrl:
2955         return -ENOMEM;
2956 }
2957
2958 static int init_vqs(struct virtnet_info *vi)
2959 {
2960         int ret;
2961
2962         /* Allocate send & receive queues */
2963         ret = virtnet_alloc_queues(vi);
2964         if (ret)
2965                 goto err;
2966
2967         ret = virtnet_find_vqs(vi);
2968         if (ret)
2969                 goto err_free;
2970
2971         cpus_read_lock();
2972         virtnet_set_affinity(vi);
2973         cpus_read_unlock();
2974
2975         return 0;
2976
2977 err_free:
2978         virtnet_free_queues(vi);
2979 err:
2980         return ret;
2981 }
2982
2983 #ifdef CONFIG_SYSFS
2984 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2985                 char *buf)
2986 {
2987         struct virtnet_info *vi = netdev_priv(queue->dev);
2988         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2989         unsigned int headroom = virtnet_get_headroom(vi);
2990         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2991         struct ewma_pkt_len *avg;
2992
2993         BUG_ON(queue_index >= vi->max_queue_pairs);
2994         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2995         return sprintf(buf, "%u\n",
2996                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
2997                                        SKB_DATA_ALIGN(headroom + tailroom)));
2998 }
2999
3000 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
3001         __ATTR_RO(mergeable_rx_buffer_size);
3002
3003 static struct attribute *virtio_net_mrg_rx_attrs[] = {
3004         &mergeable_rx_buffer_size_attribute.attr,
3005         NULL
3006 };
3007
3008 static const struct attribute_group virtio_net_mrg_rx_group = {
3009         .name = "virtio_net",
3010         .attrs = virtio_net_mrg_rx_attrs
3011 };
3012 #endif
3013
3014 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3015                                     unsigned int fbit,
3016                                     const char *fname, const char *dname)
3017 {
3018         if (!virtio_has_feature(vdev, fbit))
3019                 return false;
3020
3021         dev_err(&vdev->dev, "device advertises feature %s but not %s",
3022                 fname, dname);
3023
3024         return true;
3025 }
3026
3027 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
3028         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3029
3030 static bool virtnet_validate_features(struct virtio_device *vdev)
3031 {
3032         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3033             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3034                              "VIRTIO_NET_F_CTRL_VQ") ||
3035              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3036                              "VIRTIO_NET_F_CTRL_VQ") ||
3037              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3038                              "VIRTIO_NET_F_CTRL_VQ") ||
3039              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3040              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3041                              "VIRTIO_NET_F_CTRL_VQ"))) {
3042                 return false;
3043         }
3044
3045         return true;
3046 }
3047
3048 #define MIN_MTU ETH_MIN_MTU
3049 #define MAX_MTU ETH_MAX_MTU
3050
3051 static int virtnet_validate(struct virtio_device *vdev)
3052 {
3053         if (!vdev->config->get) {
3054                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3055                         __func__);
3056                 return -EINVAL;
3057         }
3058
3059         if (!virtnet_validate_features(vdev))
3060                 return -EINVAL;
3061
3062         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3063                 int mtu = virtio_cread16(vdev,
3064                                          offsetof(struct virtio_net_config,
3065                                                   mtu));
3066                 if (mtu < MIN_MTU)
3067                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3068         }
3069
3070         return 0;
3071 }
3072
3073 static int virtnet_probe(struct virtio_device *vdev)
3074 {
3075         int i, err = -ENOMEM;
3076         struct net_device *dev;
3077         struct virtnet_info *vi;
3078         u16 max_queue_pairs;
3079         int mtu;
3080
3081         /* Find if host supports multiqueue virtio_net device */
3082         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3083                                    struct virtio_net_config,
3084                                    max_virtqueue_pairs, &max_queue_pairs);
3085
3086         /* We need at least 2 queue's */
3087         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3088             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3089             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3090                 max_queue_pairs = 1;
3091
3092         /* Allocate ourselves a network device with room for our info */
3093         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3094         if (!dev)
3095                 return -ENOMEM;
3096
3097         /* Set up network device as normal. */
3098         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3099                            IFF_TX_SKB_NO_LINEAR;
3100         dev->netdev_ops = &virtnet_netdev;
3101         dev->features = NETIF_F_HIGHDMA;
3102
3103         dev->ethtool_ops = &virtnet_ethtool_ops;
3104         SET_NETDEV_DEV(dev, &vdev->dev);
3105
3106         /* Do we support "hardware" checksums? */
3107         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3108                 /* This opens up the world of extra features. */
3109                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3110                 if (csum)
3111                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3112
3113                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3114                         dev->hw_features |= NETIF_F_TSO
3115                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3116                 }
3117                 /* Individual feature bits: what can host handle? */
3118                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3119                         dev->hw_features |= NETIF_F_TSO;
3120                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3121                         dev->hw_features |= NETIF_F_TSO6;
3122                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3123                         dev->hw_features |= NETIF_F_TSO_ECN;
3124
3125                 dev->features |= NETIF_F_GSO_ROBUST;
3126
3127                 if (gso)
3128                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3129                 /* (!csum && gso) case will be fixed by register_netdev() */
3130         }
3131         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3132                 dev->features |= NETIF_F_RXCSUM;
3133         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3134             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3135                 dev->features |= NETIF_F_GRO_HW;
3136         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3137                 dev->hw_features |= NETIF_F_GRO_HW;
3138
3139         dev->vlan_features = dev->features;
3140
3141         /* MTU range: 68 - 65535 */
3142         dev->min_mtu = MIN_MTU;
3143         dev->max_mtu = MAX_MTU;
3144
3145         /* Configuration may specify what MAC to use.  Otherwise random. */
3146         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3147                 virtio_cread_bytes(vdev,
3148                                    offsetof(struct virtio_net_config, mac),
3149                                    dev->dev_addr, dev->addr_len);
3150         else
3151                 eth_hw_addr_random(dev);
3152
3153         /* Set up our device-specific information */
3154         vi = netdev_priv(dev);
3155         vi->dev = dev;
3156         vi->vdev = vdev;
3157         vdev->priv = vi;
3158
3159         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3160
3161         /* If we can receive ANY GSO packets, we must allocate large ones. */
3162         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3163             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3164             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3165             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3166                 vi->big_packets = true;
3167
3168         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3169                 vi->mergeable_rx_bufs = true;
3170
3171         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3172             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3173                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3174         else
3175                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3176
3177         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3178             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3179                 vi->any_header_sg = true;
3180
3181         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3182                 vi->has_cvq = true;
3183
3184         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3185                 mtu = virtio_cread16(vdev,
3186                                      offsetof(struct virtio_net_config,
3187                                               mtu));
3188                 if (mtu < dev->min_mtu) {
3189                         /* Should never trigger: MTU was previously validated
3190                          * in virtnet_validate.
3191                          */
3192                         dev_err(&vdev->dev,
3193                                 "device MTU appears to have changed it is now %d < %d",
3194                                 mtu, dev->min_mtu);
3195                         err = -EINVAL;
3196                         goto free;
3197                 }
3198
3199                 dev->mtu = mtu;
3200                 dev->max_mtu = mtu;
3201
3202                 /* TODO: size buffers correctly in this case. */
3203                 if (dev->mtu > ETH_DATA_LEN)
3204                         vi->big_packets = true;
3205         }
3206
3207         if (vi->any_header_sg)
3208                 dev->needed_headroom = vi->hdr_len;
3209
3210         /* Enable multiqueue by default */
3211         if (num_online_cpus() >= max_queue_pairs)
3212                 vi->curr_queue_pairs = max_queue_pairs;
3213         else
3214                 vi->curr_queue_pairs = num_online_cpus();
3215         vi->max_queue_pairs = max_queue_pairs;
3216
3217         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3218         err = init_vqs(vi);
3219         if (err)
3220                 goto free;
3221
3222 #ifdef CONFIG_SYSFS
3223         if (vi->mergeable_rx_bufs)
3224                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3225 #endif
3226         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3227         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3228
3229         virtnet_init_settings(dev);
3230
3231         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3232                 vi->failover = net_failover_create(vi->dev);
3233                 if (IS_ERR(vi->failover)) {
3234                         err = PTR_ERR(vi->failover);
3235                         goto free_vqs;
3236                 }
3237         }
3238
3239         err = register_netdev(dev);
3240         if (err) {
3241                 pr_debug("virtio_net: registering device failed\n");
3242                 goto free_failover;
3243         }
3244
3245         virtio_device_ready(vdev);
3246
3247         err = virtnet_cpu_notif_add(vi);
3248         if (err) {
3249                 pr_debug("virtio_net: registering cpu notifier failed\n");
3250                 goto free_unregister_netdev;
3251         }
3252
3253         virtnet_set_queues(vi, vi->curr_queue_pairs);
3254
3255         /* Assume link up if device can't report link status,
3256            otherwise get link status from config. */
3257         netif_carrier_off(dev);
3258         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3259                 schedule_work(&vi->config_work);
3260         } else {
3261                 vi->status = VIRTIO_NET_S_LINK_UP;
3262                 virtnet_update_settings(vi);
3263                 netif_carrier_on(dev);
3264         }
3265
3266         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3267                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3268                         set_bit(guest_offloads[i], &vi->guest_offloads);
3269         vi->guest_offloads_capable = vi->guest_offloads;
3270
3271         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3272                  dev->name, max_queue_pairs);
3273
3274         return 0;
3275
3276 free_unregister_netdev:
3277         vi->vdev->config->reset(vdev);
3278
3279         unregister_netdev(dev);
3280 free_failover:
3281         net_failover_destroy(vi->failover);
3282 free_vqs:
3283         cancel_delayed_work_sync(&vi->refill);
3284         free_receive_page_frags(vi);
3285         virtnet_del_vqs(vi);
3286 free:
3287         free_netdev(dev);
3288         return err;
3289 }
3290
3291 static void remove_vq_common(struct virtnet_info *vi)
3292 {
3293         vi->vdev->config->reset(vi->vdev);
3294
3295         /* Free unused buffers in both send and recv, if any. */
3296         free_unused_bufs(vi);
3297
3298         free_receive_bufs(vi);
3299
3300         free_receive_page_frags(vi);
3301
3302         virtnet_del_vqs(vi);
3303 }
3304
3305 static void virtnet_remove(struct virtio_device *vdev)
3306 {
3307         struct virtnet_info *vi = vdev->priv;
3308
3309         virtnet_cpu_notif_remove(vi);
3310
3311         /* Make sure no work handler is accessing the device. */
3312         flush_work(&vi->config_work);
3313
3314         unregister_netdev(vi->dev);
3315
3316         net_failover_destroy(vi->failover);
3317
3318         remove_vq_common(vi);
3319
3320         free_netdev(vi->dev);
3321 }
3322
3323 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3324 {
3325         struct virtnet_info *vi = vdev->priv;
3326
3327         virtnet_cpu_notif_remove(vi);
3328         virtnet_freeze_down(vdev);
3329         remove_vq_common(vi);
3330
3331         return 0;
3332 }
3333
3334 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3335 {
3336         struct virtnet_info *vi = vdev->priv;
3337         int err;
3338
3339         err = virtnet_restore_up(vdev);
3340         if (err)
3341                 return err;
3342         virtnet_set_queues(vi, vi->curr_queue_pairs);
3343
3344         err = virtnet_cpu_notif_add(vi);
3345         if (err) {
3346                 virtnet_freeze_down(vdev);
3347                 remove_vq_common(vi);
3348                 return err;
3349         }
3350
3351         return 0;
3352 }
3353
3354 static struct virtio_device_id id_table[] = {
3355         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3356         { 0 },
3357 };
3358
3359 #define VIRTNET_FEATURES \
3360         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3361         VIRTIO_NET_F_MAC, \
3362         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3363         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3364         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3365         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3366         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3367         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3368         VIRTIO_NET_F_CTRL_MAC_ADDR, \
3369         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3370         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3371
3372 static unsigned int features[] = {
3373         VIRTNET_FEATURES,
3374 };
3375
3376 static unsigned int features_legacy[] = {
3377         VIRTNET_FEATURES,
3378         VIRTIO_NET_F_GSO,
3379         VIRTIO_F_ANY_LAYOUT,
3380 };
3381
3382 static struct virtio_driver virtio_net_driver = {
3383         .feature_table = features,
3384         .feature_table_size = ARRAY_SIZE(features),
3385         .feature_table_legacy = features_legacy,
3386         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3387         .driver.name =  KBUILD_MODNAME,
3388         .driver.owner = THIS_MODULE,
3389         .id_table =     id_table,
3390         .validate =     virtnet_validate,
3391         .probe =        virtnet_probe,
3392         .remove =       virtnet_remove,
3393         .config_changed = virtnet_config_changed,
3394 #ifdef CONFIG_PM_SLEEP
3395         .freeze =       virtnet_freeze,
3396         .restore =      virtnet_restore,
3397 #endif
3398 };
3399
3400 static __init int virtio_net_driver_init(void)
3401 {
3402         int ret;
3403
3404         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3405                                       virtnet_cpu_online,
3406                                       virtnet_cpu_down_prep);
3407         if (ret < 0)
3408                 goto out;
3409         virtionet_online = ret;
3410         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3411                                       NULL, virtnet_cpu_dead);
3412         if (ret)
3413                 goto err_dead;
3414
3415         ret = register_virtio_driver(&virtio_net_driver);
3416         if (ret)
3417                 goto err_virtio;
3418         return 0;
3419 err_virtio:
3420         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3421 err_dead:
3422         cpuhp_remove_multi_state(virtionet_online);
3423 out:
3424         return ret;
3425 }
3426 module_init(virtio_net_driver_init);
3427
3428 static __exit void virtio_net_driver_exit(void)
3429 {
3430         unregister_virtio_driver(&virtio_net_driver);
3431         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3432         cpuhp_remove_multi_state(virtionet_online);
3433 }
3434 module_exit(virtio_net_driver_exit);
3435
3436 MODULE_DEVICE_TABLE(virtio, id_table);
3437 MODULE_DESCRIPTION("Virtio network driver");
3438 MODULE_LICENSE("GPL");