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 <[email protected]> IBM Corporation
   5  */
   6 //#define DEBUG
   7 #include <linux/netdevice.h>
   8 #include <linux/etherdevice.h>
   9 #include <linux/ethtool.h>
  10 #include <linux/module.h>
  11 #include <linux/virtio.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/bpf.h>
  14 #include <linux/bpf_trace.h>
  15 #include <linux/scatterlist.h>
  16 #include <linux/if_vlan.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/average.h>
  20 #include <linux/filter.h>
  21 #include <linux/kernel.h>
  22 #include <net/route.h>
  23 #include <net/xdp.h>
  24 #include <net/net_failover.h>
  25
  26 static int napi_weight = NAPI_POLL_WEIGHT;
  27 module_param(napi_weight, int, 0444);
  28
  29 static bool csum = true, gso = true, napi_tx = true;
  30 module_param(csum, bool, 0444);
  31 module_param(gso, bool, 0444);
  32 module_param(napi_tx, bool, 0644);
  33
  34 /* FIXME: MTU in config. */
  35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  36 #define GOOD_COPY_LEN   128
  37
  38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  39
  40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  41 #define VIRTIO_XDP_HEADROOM 256
  42
  43 /* Separating two types of XDP xmit */
  44 #define VIRTIO_XDP_TX           BIT(0)
  45 #define VIRTIO_XDP_REDIR        BIT(1)
  46
  47 #define VIRTIO_XDP_FLAG BIT(0)
  48
  49 /* RX packet size EWMA. The average packet size is used to determine the packet
  50  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  51  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  52  * term, transient changes in packet size.
  53  */
  54 DECLARE_EWMA(pkt_len, 0, 64)
  55
  56 #define VIRTNET_DRIVER_VERSION "1.0.0"
  57
  58 static const unsigned long guest_offloads[] = {
  59         VIRTIO_NET_F_GUEST_TSO4,
  60         VIRTIO_NET_F_GUEST_TSO6,
  61         VIRTIO_NET_F_GUEST_ECN,
  62         VIRTIO_NET_F_GUEST_UFO,
  63         VIRTIO_NET_F_GUEST_CSUM
  64 };
  65
  66 #define GUEST_OFFLOAD_LRO_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
  67                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
  68                                 (1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
  69                                 (1ULL << VIRTIO_NET_F_GUEST_UFO))
  70
  71 struct virtnet_stat_desc {
  72         char desc[ETH_GSTRING_LEN];
  73         size_t offset;
  74 };
  75
  76 struct virtnet_sq_stats {
  77         struct u64_stats_sync syncp;
  78         u64 packets;
  79         u64 bytes;
  80         u64 xdp_tx;
  81         u64 xdp_tx_drops;
  82         u64 kicks;
  83 };
  84
  85 struct virtnet_rq_stats {
  86         struct u64_stats_sync syncp;
  87         u64 packets;
  88         u64 bytes;
  89         u64 drops;
  90         u64 xdp_packets;
  91         u64 xdp_tx;
  92         u64 xdp_redirects;
  93         u64 xdp_drops;
  94         u64 kicks;
  95 };
  96
  97 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
  98 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
  99
 100 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
 101         { "packets",            VIRTNET_SQ_STAT(packets) },
 102         { "bytes",              VIRTNET_SQ_STAT(bytes) },
 103         { "xdp_tx",             VIRTNET_SQ_STAT(xdp_tx) },
 104         { "xdp_tx_drops",       VIRTNET_SQ_STAT(xdp_tx_drops) },
 105         { "kicks",              VIRTNET_SQ_STAT(kicks) },
 106 };
 107
 108 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
 109         { "packets",            VIRTNET_RQ_STAT(packets) },
 110         { "bytes",              VIRTNET_RQ_STAT(bytes) },
 111         { "drops",              VIRTNET_RQ_STAT(drops) },
 112         { "xdp_packets",        VIRTNET_RQ_STAT(xdp_packets) },
 113         { "xdp_tx",             VIRTNET_RQ_STAT(xdp_tx) },
 114         { "xdp_redirects",      VIRTNET_RQ_STAT(xdp_redirects) },
 115         { "xdp_drops",          VIRTNET_RQ_STAT(xdp_drops) },
 116         { "kicks",              VIRTNET_RQ_STAT(kicks) },
 117 };
 118
 119 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 120 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 121
 122 /* Internal representation of a send virtqueue */
 123 struct send_queue {
 124         /* Virtqueue associated with this send _queue */
 125         struct virtqueue *vq;
 126
 127         /* TX: fragments + linear part + virtio header */
 128         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 129
 130         /* Name of the send queue: output.$index */
 131         char name[40];
 132
 133         struct virtnet_sq_stats stats;
 134
 135         struct napi_struct napi;
 136 };
 137
 138 /* Internal representation of a receive virtqueue */
 139 struct receive_queue {
 140         /* Virtqueue associated with this receive_queue */
 141         struct virtqueue *vq;
 142
 143         struct napi_struct napi;
 144
 145         struct bpf_prog __rcu *xdp_prog;
 146
 147         struct virtnet_rq_stats stats;
 148
 149         /* Chain pages by the private ptr. */
 150         struct page *pages;
 151
 152         /* Average packet length for mergeable receive buffers. */
 153         struct ewma_pkt_len mrg_avg_pkt_len;
 154
 155         /* Page frag for packet buffer allocation. */
 156         struct page_frag alloc_frag;
 157
 158         /* RX: fragments + linear part + virtio header */
 159         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 160
 161         /* Min single buffer size for mergeable buffers case. */
 162         unsigned int min_buf_len;
 163
 164         /* Name of this receive queue: input.$index */
 165         char name[40];
 166
 167         struct xdp_rxq_info xdp_rxq;
 168 };
 169
 170 /* Control VQ buffers: protected by the rtnl lock */
 171 struct control_buf {
 172         struct virtio_net_ctrl_hdr hdr;
 173         virtio_net_ctrl_ack status;
 174         struct virtio_net_ctrl_mq mq;
 175         u8 promisc;
 176         u8 allmulti;
 177         __virtio16 vid;
 178         __virtio64 offloads;
 179 };
 180
 181 struct virtnet_info {
 182         struct virtio_device *vdev;
 183         struct virtqueue *cvq;
 184         struct net_device *dev;
 185         struct send_queue *sq;
 186         struct receive_queue *rq;
 187         unsigned int status;
 188
 189         /* Max # of queue pairs supported by the device */
 190         u16 max_queue_pairs;
 191
 192         /* # of queue pairs currently used by the driver */
 193         u16 curr_queue_pairs;
 194
 195         /* # of XDP queue pairs currently used by the driver */
 196         u16 xdp_queue_pairs;
 197
 198         /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
 199         bool xdp_enabled;
 200
 201         /* I like... big packets and I cannot lie! */
 202         bool big_packets;
 203
 204         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 205         bool mergeable_rx_bufs;
 206
 207         /* Has control virtqueue */
 208         bool has_cvq;
 209
 210         /* Host can handle any s/g split between our header and packet data */
 211         bool any_header_sg;
 212
 213         /* Packet virtio header size */
 214         u8 hdr_len;
 215
 216         /* Work struct for refilling if we run low on memory. */
 217         struct delayed_work refill;
 218
 219         /* Work struct for config space updates */
 220         struct work_struct config_work;
 221
 222         /* Does the affinity hint is set for virtqueues? */
 223         bool affinity_hint_set;
 224
 225         /* CPU hotplug instances for online & dead */
 226         struct hlist_node node;
 227         struct hlist_node node_dead;
 228
 229         struct control_buf *ctrl;
 230
 231         /* Ethtool settings */
 232         u8 duplex;
 233         u32 speed;
 234
 235         unsigned long guest_offloads;
 236         unsigned long guest_offloads_capable;
 237
 238         /* failover when STANDBY feature enabled */
 239         struct failover *failover;
 240 };
 241
 242 struct padded_vnet_hdr {
 243         struct virtio_net_hdr_mrg_rxbuf hdr;
 244         /*
 245          * hdr is in a separate sg buffer, and data sg buffer shares same page
 246          * with this header sg. This padding makes next sg 16 byte aligned
 247          * after the header.
 248          */
 249         char padding[4];
 250 };
 251
 252 static bool is_xdp_frame(void *ptr)
 253 {
 254         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 255 }
 256
 257 static void *xdp_to_ptr(struct xdp_frame *ptr)
 258 {
 259         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 260 }
 261
 262 static struct xdp_frame *ptr_to_xdp(void *ptr)
 263 {
 264         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 265 }
 266
 267 /* Converting between virtqueue no. and kernel tx/rx queue no.
 268  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 269  */
 270 static int vq2txq(struct virtqueue *vq)
 271 {
 272         return (vq->index - 1) / 2;
 273 }
 274
 275 static int txq2vq(int txq)
 276 {
 277         return txq * 2 + 1;
 278 }
 279
 280 static int vq2rxq(struct virtqueue *vq)
 281 {
 282         return vq->index / 2;
 283 }
 284
 285 static int rxq2vq(int rxq)
 286 {
 287         return rxq * 2;
 288 }
 289
 290 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 291 {
 292         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 293 }
 294
 295 /*
 296  * private is used to chain pages for big packets, put the whole
 297  * most recent used list in the beginning for reuse
 298  */
 299 static void give_pages(struct receive_queue *rq, struct page *page)
 300 {
 301         struct page *end;
 302
 303         /* Find end of list, sew whole thing into vi->rq.pages. */
 304         for (end = page; end->private; end = (struct page *)end->private);
 305         end->private = (unsigned long)rq->pages;
 306         rq->pages = page;
 307 }
 308
 309 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 310 {
 311         struct page *p = rq->pages;
 312
 313         if (p) {
 314                 rq->pages = (struct page *)p->private;
 315                 /* clear private here, it is used to chain pages */
 316                 p->private = 0;
 317         } else
 318                 p = alloc_page(gfp_mask);
 319         return p;
 320 }
 321
 322 static void virtqueue_napi_schedule(struct napi_struct *napi,
 323                                     struct virtqueue *vq)
 324 {
 325         if (napi_schedule_prep(napi)) {
 326                 virtqueue_disable_cb(vq);
 327                 __napi_schedule(napi);
 328         }
 329 }
 330
 331 static void virtqueue_napi_complete(struct napi_struct *napi,
 332                                     struct virtqueue *vq, int processed)
 333 {
 334         int opaque;
 335
 336         opaque = virtqueue_enable_cb_prepare(vq);
 337         if (napi_complete_done(napi, processed)) {
 338                 if (unlikely(virtqueue_poll(vq, opaque)))
 339                         virtqueue_napi_schedule(napi, vq);
 340         } else {
 341                 virtqueue_disable_cb(vq);
 342         }
 343 }
 344
 345 static void skb_xmit_done(struct virtqueue *vq)
 346 {
 347         struct virtnet_info *vi = vq->vdev->priv;
 348         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 349
 350         /* Suppress further interrupts. */
 351         virtqueue_disable_cb(vq);
 352
 353         if (napi->weight)
 354                 virtqueue_napi_schedule(napi, vq);
 355         else
 356                 /* We were probably waiting for more output buffers. */
 357                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 358 }
 359
 360 #define MRG_CTX_HEADER_SHIFT 22
 361 static void *mergeable_len_to_ctx(unsigned int truesize,
 362                                   unsigned int headroom)
 363 {
 364         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 365 }
 366
 367 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 368 {
 369         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 370 }
 371
 372 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 373 {
 374         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 375 }
 376
 377 /* Called from bottom half context */
 378 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 379                                    struct receive_queue *rq,
 380                                    struct page *page, unsigned int offset,
 381                                    unsigned int len, unsigned int truesize,
 382                                    bool hdr_valid, unsigned int metasize,
 383                                    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         if (headroom) {
 406                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 407                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 408                  */
 409                 truesize = PAGE_SIZE;
 410                 tailroom = truesize - len - offset;
 411                 buf = page_address(page);
 412         } else {
 413                 tailroom = truesize - len;
 414                 buf = p;
 415         }
 416
 417         len -= hdr_len;
 418         offset += hdr_padded_len;
 419         p += hdr_padded_len;
 420
 421         shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 422
 423         /* copy small packet so we can reuse these pages */
 424         if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
 425                 skb = build_skb(buf, truesize);
 426                 if (unlikely(!skb))
 427                         return NULL;
 428
 429                 skb_reserve(skb, p - buf);
 430                 skb_put(skb, len);
 431                 goto ok;
 432         }
 433
 434         /* copy small packet so we can reuse these pages for small data */
 435         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 436         if (unlikely(!skb))
 437                 return NULL;
 438
 439         /* Copy all frame if it fits skb->head, otherwise
 440          * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
 441          */
 442         if (len <= skb_tailroom(skb))
 443                 copy = len;
 444         else
 445                 copy = ETH_HLEN + metasize;
 446         skb_put_data(skb, p, copy);
 447
 448         len -= copy;
 449         offset += copy;
 450
 451         if (vi->mergeable_rx_bufs) {
 452                 if (len)
 453                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 454                 else
 455                         page_to_free = page;
 456                 goto ok;
 457         }
 458
 459         /*
 460          * Verify that we can indeed put this data into a skb.
 461          * This is here to handle cases when the device erroneously
 462          * tries to receive more than is possible. This is usually
 463          * the case of a broken device.
 464          */
 465         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 466                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 467                 dev_kfree_skb(skb);
 468                 return NULL;
 469         }
 470         BUG_ON(offset >= PAGE_SIZE);
 471         while (len) {
 472                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 473                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 474                                 frag_size, truesize);
 475                 len -= frag_size;
 476                 page = (struct page *)page->private;
 477                 offset = 0;
 478         }
 479
 480         if (page)
 481                 give_pages(rq, page);
 482
 483 ok:
 484         /* hdr_valid means no XDP, so we can copy the vnet header */
 485         if (hdr_valid) {
 486                 hdr = skb_vnet_hdr(skb);
 487                 memcpy(hdr, hdr_p, hdr_len);
 488         }
 489         if (page_to_free)
 490                 put_page(page_to_free);
 491
 492         if (metasize) {
 493                 __skb_pull(skb, metasize);
 494                 skb_metadata_set(skb, metasize);
 495         }
 496
 497         return skb;
 498 }
 499
 500 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 501                                    struct send_queue *sq,
 502                                    struct xdp_frame *xdpf)
 503 {
 504         struct virtio_net_hdr_mrg_rxbuf *hdr;
 505         int err;
 506
 507         if (unlikely(xdpf->headroom < vi->hdr_len))
 508                 return -EOVERFLOW;
 509
 510         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 511         xdpf->data -= vi->hdr_len;
 512         /* Zero header and leave csum up to XDP layers */
 513         hdr = xdpf->data;
 514         memset(hdr, 0, vi->hdr_len);
 515         xdpf->len   += vi->hdr_len;
 516
 517         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 518
 519         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
 520                                    GFP_ATOMIC);
 521         if (unlikely(err))
 522                 return -ENOSPC; /* Caller handle free/refcnt */
 523
 524         return 0;
 525 }
 526
 527 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
 528  * the current cpu, so it does not need to be locked.
 529  *
 530  * Here we use marco instead of inline functions because we have to deal with
 531  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
 532  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
 533  * functions to perfectly solve these three problems at the same time.
 534  */
 535 #define virtnet_xdp_get_sq(vi) ({                                       \
 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 += smp_processor_id();                               \
 543                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 544                 __netif_tx_acquire(txq);                                \
 545         } else {                                                        \
 546                 qp = smp_processor_id() % v->curr_queue_pairs;          \
 547                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 548                 __netif_tx_lock(txq, raw_smp_processor_id());           \
 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         rcu_read_lock();
 730         xdp_prog = rcu_dereference(rq->xdp_prog);
 731         if (xdp_prog) {
 732                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 733                 struct xdp_frame *xdpf;
 734                 struct xdp_buff xdp;
 735                 void *orig_data;
 736                 u32 act;
 737
 738                 if (unlikely(hdr->hdr.gso_type))
 739                         goto err_xdp;
 740
 741                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 742                         int offset = buf - page_address(page) + header_offset;
 743                         unsigned int tlen = len + vi->hdr_len;
 744                         u16 num_buf = 1;
 745
 746                         xdp_headroom = virtnet_get_headroom(vi);
 747                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 748                         headroom = vi->hdr_len + header_offset;
 749                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 750                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 751                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 752                                                       offset, header_offset,
 753                                                       &tlen);
 754                         if (!xdp_page)
 755                                 goto err_xdp;
 756
 757                         buf = page_address(xdp_page);
 758                         put_page(page);
 759                         page = xdp_page;
 760                 }
 761
 762                 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
 763                 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
 764                                  xdp_headroom, len, true);
 765                 orig_data = xdp.data;
 766                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 767                 stats->xdp_packets++;
 768
 769                 switch (act) {
 770                 case XDP_PASS:
 771                         /* Recalculate length in case bpf program changed it */
 772                         delta = orig_data - xdp.data;
 773                         len = xdp.data_end - xdp.data;
 774                         metasize = xdp.data - xdp.data_meta;
 775                         break;
 776                 case XDP_TX:
 777                         stats->xdp_tx++;
 778                         xdpf = xdp_convert_buff_to_frame(&xdp);
 779                         if (unlikely(!xdpf))
 780                                 goto err_xdp;
 781                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 782                         if (unlikely(!err)) {
 783                                 xdp_return_frame_rx_napi(xdpf);
 784                         } else if (unlikely(err < 0)) {
 785                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 786                                 goto err_xdp;
 787                         }
 788                         *xdp_xmit |= VIRTIO_XDP_TX;
 789                         rcu_read_unlock();
 790                         goto xdp_xmit;
 791                 case XDP_REDIRECT:
 792                         stats->xdp_redirects++;
 793                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 794                         if (err)
 795                                 goto err_xdp;
 796                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 797                         rcu_read_unlock();
 798                         goto xdp_xmit;
 799                 default:
 800                         bpf_warn_invalid_xdp_action(act);
 801                         fallthrough;
 802                 case XDP_ABORTED:
 803                         trace_xdp_exception(vi->dev, xdp_prog, act);
 804                         goto err_xdp;
 805                 case XDP_DROP:
 806                         goto err_xdp;
 807                 }
 808         }
 809         rcu_read_unlock();
 810
 811         skb = build_skb(buf, buflen);
 812         if (!skb) {
 813                 put_page(page);
 814                 goto err;
 815         }
 816         skb_reserve(skb, headroom - delta);
 817         skb_put(skb, len);
 818         if (!xdp_prog) {
 819                 buf += header_offset;
 820                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 821         } /* keep zeroed vnet hdr since XDP is loaded */
 822
 823         if (metasize)
 824                 skb_metadata_set(skb, metasize);
 825
 826 err:
 827         return skb;
 828
 829 err_xdp:
 830         rcu_read_unlock();
 831         stats->xdp_drops++;
 832         stats->drops++;
 833         put_page(page);
 834 xdp_xmit:
 835         return NULL;
 836 }
 837
 838 static struct sk_buff *receive_big(struct net_device *dev,
 839                                    struct virtnet_info *vi,
 840                                    struct receive_queue *rq,
 841                                    void *buf,
 842                                    unsigned int len,
 843                                    struct virtnet_rq_stats *stats)
 844 {
 845         struct page *page = buf;
 846         struct sk_buff *skb =
 847                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
 848
 849         stats->bytes += len - vi->hdr_len;
 850         if (unlikely(!skb))
 851                 goto err;
 852
 853         return skb;
 854
 855 err:
 856         stats->drops++;
 857         give_pages(rq, page);
 858         return NULL;
 859 }
 860
 861 static struct sk_buff *receive_mergeable(struct net_device *dev,
 862                                          struct virtnet_info *vi,
 863                                          struct receive_queue *rq,
 864                                          void *buf,
 865                                          void *ctx,
 866                                          unsigned int len,
 867                                          unsigned int *xdp_xmit,
 868                                          struct virtnet_rq_stats *stats)
 869 {
 870         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 871         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 872         struct page *page = virt_to_head_page(buf);
 873         int offset = buf - page_address(page);
 874         struct sk_buff *head_skb, *curr_skb;
 875         struct bpf_prog *xdp_prog;
 876         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
 877         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 878         unsigned int metasize = 0;
 879         unsigned int frame_sz;
 880         int err;
 881
 882         head_skb = NULL;
 883         stats->bytes += len - vi->hdr_len;
 884
 885         rcu_read_lock();
 886         xdp_prog = rcu_dereference(rq->xdp_prog);
 887         if (xdp_prog) {
 888                 struct xdp_frame *xdpf;
 889                 struct page *xdp_page;
 890                 struct xdp_buff xdp;
 891                 void *data;
 892                 u32 act;
 893
 894                 /* Transient failure which in theory could occur if
 895                  * in-flight packets from before XDP was enabled reach
 896                  * the receive path after XDP is loaded.
 897                  */
 898                 if (unlikely(hdr->hdr.gso_type))
 899                         goto err_xdp;
 900
 901                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 902                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 903                  */
 904                 frame_sz = headroom ? PAGE_SIZE : truesize;
 905
 906                 /* This happens when rx buffer size is underestimated
 907                  * or headroom is not enough because of the buffer
 908                  * was refilled before XDP is set. This should only
 909                  * happen for the first several packets, so we don't
 910                  * care much about its performance.
 911                  */
 912                 if (unlikely(num_buf > 1 ||
 913                              headroom < virtnet_get_headroom(vi))) {
 914                         /* linearize data for XDP */
 915                         xdp_page = xdp_linearize_page(rq, &num_buf,
 916                                                       page, offset,
 917                                                       VIRTIO_XDP_HEADROOM,
 918                                                       &len);
 919                         frame_sz = PAGE_SIZE;
 920
 921                         if (!xdp_page)
 922                                 goto err_xdp;
 923                         offset = VIRTIO_XDP_HEADROOM;
 924                 } else {
 925                         xdp_page = page;
 926                 }
 927
 928                 /* Allow consuming headroom but reserve enough space to push
 929                  * the descriptor on if we get an XDP_TX return code.
 930                  */
 931                 data = page_address(xdp_page) + offset;
 932                 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
 933                 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
 934                                  VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
 935
 936                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 937                 stats->xdp_packets++;
 938
 939                 switch (act) {
 940                 case XDP_PASS:
 941                         metasize = xdp.data - xdp.data_meta;
 942
 943                         /* recalculate offset to account for any header
 944                          * adjustments and minus the metasize to copy the
 945                          * metadata in page_to_skb(). Note other cases do not
 946                          * build an skb and avoid using offset
 947                          */
 948                         offset = xdp.data - page_address(xdp_page) -
 949                                  vi->hdr_len - metasize;
 950
 951                         /* recalculate len if xdp.data, xdp.data_end or
 952                          * xdp.data_meta were adjusted
 953                          */
 954                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 955                         /* We can only create skb based on xdp_page. */
 956                         if (unlikely(xdp_page != page)) {
 957                                 rcu_read_unlock();
 958                                 put_page(page);
 959                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
 960                                                        len, PAGE_SIZE, false,
 961                                                        metasize, headroom);
 962                                 return head_skb;
 963                         }
 964                         break;
 965                 case XDP_TX:
 966                         stats->xdp_tx++;
 967                         xdpf = xdp_convert_buff_to_frame(&xdp);
 968                         if (unlikely(!xdpf))
 969                                 goto err_xdp;
 970                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 971                         if (unlikely(!err)) {
 972                                 xdp_return_frame_rx_napi(xdpf);
 973                         } else if (unlikely(err < 0)) {
 974                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 975                                 if (unlikely(xdp_page != page))
 976                                         put_page(xdp_page);
 977                                 goto err_xdp;
 978                         }
 979                         *xdp_xmit |= VIRTIO_XDP_TX;
 980                         if (unlikely(xdp_page != page))
 981                                 put_page(page);
 982                         rcu_read_unlock();
 983                         goto xdp_xmit;
 984                 case XDP_REDIRECT:
 985                         stats->xdp_redirects++;
 986                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 987                         if (err) {
 988                                 if (unlikely(xdp_page != page))
 989                                         put_page(xdp_page);
 990                                 goto err_xdp;
 991                         }
 992                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 993                         if (unlikely(xdp_page != page))
 994                                 put_page(page);
 995                         rcu_read_unlock();
 996                         goto xdp_xmit;
 997                 default:
 998                         bpf_warn_invalid_xdp_action(act);
 999                         fallthrough;
1000                 case XDP_ABORTED:
1001                         trace_xdp_exception(vi->dev, xdp_prog, act);
1002                         fallthrough;
1003                 case XDP_DROP:
1004                         if (unlikely(xdp_page != page))
1005                                 __free_pages(xdp_page, 0);
1006                         goto err_xdp;
1007                 }
1008         }
1009         rcu_read_unlock();
1010
1011         if (unlikely(len > truesize)) {
1012                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1013                          dev->name, len, (unsigned long)ctx);
1014                 dev->stats.rx_length_errors++;
1015                 goto err_skb;
1016         }
1017
1018         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1019                                metasize, headroom);
1020         curr_skb = head_skb;
1021
1022         if (unlikely(!curr_skb))
1023                 goto err_skb;
1024         while (--num_buf) {
1025                 int num_skb_frags;
1026
1027                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1028                 if (unlikely(!buf)) {
1029                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1030                                  dev->name, num_buf,
1031                                  virtio16_to_cpu(vi->vdev,
1032                                                  hdr->num_buffers));
1033                         dev->stats.rx_length_errors++;
1034                         goto err_buf;
1035                 }
1036
1037                 stats->bytes += len;
1038                 page = virt_to_head_page(buf);
1039
1040                 truesize = mergeable_ctx_to_truesize(ctx);
1041                 if (unlikely(len > truesize)) {
1042                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1043                                  dev->name, len, (unsigned long)ctx);
1044                         dev->stats.rx_length_errors++;
1045                         goto err_skb;
1046                 }
1047
1048                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1049                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1050                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1051
1052                         if (unlikely(!nskb))
1053                                 goto err_skb;
1054                         if (curr_skb == head_skb)
1055                                 skb_shinfo(curr_skb)->frag_list = nskb;
1056                         else
1057                                 curr_skb->next = nskb;
1058                         curr_skb = nskb;
1059                         head_skb->truesize += nskb->truesize;
1060                         num_skb_frags = 0;
1061                 }
1062                 if (curr_skb != head_skb) {
1063                         head_skb->data_len += len;
1064                         head_skb->len += len;
1065                         head_skb->truesize += truesize;
1066                 }
1067                 offset = buf - page_address(page);
1068                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1069                         put_page(page);
1070                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1071                                              len, truesize);
1072                 } else {
1073                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1074                                         offset, len, truesize);
1075                 }
1076         }
1077
1078         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1079         return head_skb;
1080
1081 err_xdp:
1082         rcu_read_unlock();
1083         stats->xdp_drops++;
1084 err_skb:
1085         put_page(page);
1086         while (num_buf-- > 1) {
1087                 buf = virtqueue_get_buf(rq->vq, &len);
1088                 if (unlikely(!buf)) {
1089                         pr_debug("%s: rx error: %d buffers missing\n",
1090                                  dev->name, num_buf);
1091                         dev->stats.rx_length_errors++;
1092                         break;
1093                 }
1094                 stats->bytes += len;
1095                 page = virt_to_head_page(buf);
1096                 put_page(page);
1097         }
1098 err_buf:
1099         stats->drops++;
1100         dev_kfree_skb(head_skb);
1101 xdp_xmit:
1102         return NULL;
1103 }
1104
1105 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1106                         void *buf, unsigned int len, void **ctx,
1107                         unsigned int *xdp_xmit,
1108                         struct virtnet_rq_stats *stats)
1109 {
1110         struct net_device *dev = vi->dev;
1111         struct sk_buff *skb;
1112         struct virtio_net_hdr_mrg_rxbuf *hdr;
1113
1114         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1115                 pr_debug("%s: short packet %i\n", dev->name, len);
1116                 dev->stats.rx_length_errors++;
1117                 if (vi->mergeable_rx_bufs) {
1118                         put_page(virt_to_head_page(buf));
1119                 } else if (vi->big_packets) {
1120                         give_pages(rq, buf);
1121                 } else {
1122                         put_page(virt_to_head_page(buf));
1123                 }
1124                 return;
1125         }
1126
1127         if (vi->mergeable_rx_bufs)
1128                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1129                                         stats);
1130         else if (vi->big_packets)
1131                 skb = receive_big(dev, vi, rq, buf, len, stats);
1132         else
1133                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1134
1135         if (unlikely(!skb))
1136                 return;
1137
1138         hdr = skb_vnet_hdr(skb);
1139
1140         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1141                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1142
1143         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1144                                   virtio_is_little_endian(vi->vdev))) {
1145                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1146                                      dev->name, hdr->hdr.gso_type,
1147                                      hdr->hdr.gso_size);
1148                 goto frame_err;
1149         }
1150
1151         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1152         skb->protocol = eth_type_trans(skb, dev);
1153         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1154                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1155
1156         napi_gro_receive(&rq->napi, skb);
1157         return;
1158
1159 frame_err:
1160         dev->stats.rx_frame_errors++;
1161         dev_kfree_skb(skb);
1162 }
1163
1164 /* Unlike mergeable buffers, all buffers are allocated to the
1165  * same size, except for the headroom. For this reason we do
1166  * not need to use  mergeable_len_to_ctx here - it is enough
1167  * to store the headroom as the context ignoring the truesize.
1168  */
1169 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1170                              gfp_t gfp)
1171 {
1172         struct page_frag *alloc_frag = &rq->alloc_frag;
1173         char *buf;
1174         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1175         void *ctx = (void *)(unsigned long)xdp_headroom;
1176         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1177         int err;
1178
1179         len = SKB_DATA_ALIGN(len) +
1180               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1181         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1182                 return -ENOMEM;
1183
1184         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1185         get_page(alloc_frag->page);
1186         alloc_frag->offset += len;
1187         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1188                     vi->hdr_len + GOOD_PACKET_LEN);
1189         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1190         if (err < 0)
1191                 put_page(virt_to_head_page(buf));
1192         return err;
1193 }
1194
1195 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1196                            gfp_t gfp)
1197 {
1198         struct page *first, *list = NULL;
1199         char *p;
1200         int i, err, offset;
1201
1202         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1203
1204         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1205         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1206                 first = get_a_page(rq, gfp);
1207                 if (!first) {
1208                         if (list)
1209                                 give_pages(rq, list);
1210                         return -ENOMEM;
1211                 }
1212                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1213
1214                 /* chain new page in list head to match sg */
1215                 first->private = (unsigned long)list;
1216                 list = first;
1217         }
1218
1219         first = get_a_page(rq, gfp);
1220         if (!first) {
1221                 give_pages(rq, list);
1222                 return -ENOMEM;
1223         }
1224         p = page_address(first);
1225
1226         /* rq->sg[0], rq->sg[1] share the same page */
1227         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1228         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1229
1230         /* rq->sg[1] for data packet, from offset */
1231         offset = sizeof(struct padded_vnet_hdr);
1232         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1233
1234         /* chain first in list head */
1235         first->private = (unsigned long)list;
1236         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1237                                   first, gfp);
1238         if (err < 0)
1239                 give_pages(rq, first);
1240
1241         return err;
1242 }
1243
1244 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1245                                           struct ewma_pkt_len *avg_pkt_len,
1246                                           unsigned int room)
1247 {
1248         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1249         unsigned int len;
1250
1251         if (room)
1252                 return PAGE_SIZE - room;
1253
1254         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1255                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1256
1257         return ALIGN(len, L1_CACHE_BYTES);
1258 }
1259
1260 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1261                                  struct receive_queue *rq, gfp_t gfp)
1262 {
1263         struct page_frag *alloc_frag = &rq->alloc_frag;
1264         unsigned int headroom = virtnet_get_headroom(vi);
1265         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1266         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1267         char *buf;
1268         void *ctx;
1269         int err;
1270         unsigned int len, hole;
1271
1272         /* Extra tailroom is needed to satisfy XDP's assumption. This
1273          * means rx frags coalescing won't work, but consider we've
1274          * disabled GSO for XDP, it won't be a big issue.
1275          */
1276         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1277         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1278                 return -ENOMEM;
1279
1280         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1281         buf += headroom; /* advance address leaving hole at front of pkt */
1282         get_page(alloc_frag->page);
1283         alloc_frag->offset += len + room;
1284         hole = alloc_frag->size - alloc_frag->offset;
1285         if (hole < len + room) {
1286                 /* To avoid internal fragmentation, if there is very likely not
1287                  * enough space for another buffer, add the remaining space to
1288                  * the current buffer.
1289                  */
1290                 len += hole;
1291                 alloc_frag->offset += hole;
1292         }
1293
1294         sg_init_one(rq->sg, buf, len);
1295         ctx = mergeable_len_to_ctx(len, headroom);
1296         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1297         if (err < 0)
1298                 put_page(virt_to_head_page(buf));
1299
1300         return err;
1301 }
1302
1303 /*
1304  * Returns false if we couldn't fill entirely (OOM).
1305  *
1306  * Normally run in the receive path, but can also be run from ndo_open
1307  * before we're receiving packets, or from refill_work which is
1308  * careful to disable receiving (using napi_disable).
1309  */
1310 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1311                           gfp_t gfp)
1312 {
1313         int err;
1314         bool oom;
1315
1316         do {
1317                 if (vi->mergeable_rx_bufs)
1318                         err = add_recvbuf_mergeable(vi, rq, gfp);
1319                 else if (vi->big_packets)
1320                         err = add_recvbuf_big(vi, rq, gfp);
1321                 else
1322                         err = add_recvbuf_small(vi, rq, gfp);
1323
1324                 oom = err == -ENOMEM;
1325                 if (err)
1326                         break;
1327         } while (rq->vq->num_free);
1328         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1329                 unsigned long flags;
1330
1331                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1332                 rq->stats.kicks++;
1333                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1334         }
1335
1336         return !oom;
1337 }
1338
1339 static void skb_recv_done(struct virtqueue *rvq)
1340 {
1341         struct virtnet_info *vi = rvq->vdev->priv;
1342         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1343
1344         virtqueue_napi_schedule(&rq->napi, rvq);
1345 }
1346
1347 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1348 {
1349         napi_enable(napi);
1350
1351         /* If all buffers were filled by other side before we napi_enabled, we
1352          * won't get another interrupt, so process any outstanding packets now.
1353          * Call local_bh_enable after to trigger softIRQ processing.
1354          */
1355         local_bh_disable();
1356         virtqueue_napi_schedule(napi, vq);
1357         local_bh_enable();
1358 }
1359
1360 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1361                                    struct virtqueue *vq,
1362                                    struct napi_struct *napi)
1363 {
1364         if (!napi->weight)
1365                 return;
1366
1367         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1368          * enable the feature if this is likely affine with the transmit path.
1369          */
1370         if (!vi->affinity_hint_set) {
1371                 napi->weight = 0;
1372                 return;
1373         }
1374
1375         return virtnet_napi_enable(vq, napi);
1376 }
1377
1378 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1379 {
1380         if (napi->weight)
1381                 napi_disable(napi);
1382 }
1383
1384 static void refill_work(struct work_struct *work)
1385 {
1386         struct virtnet_info *vi =
1387                 container_of(work, struct virtnet_info, refill.work);
1388         bool still_empty;
1389         int i;
1390
1391         for (i = 0; i < vi->curr_queue_pairs; i++) {
1392                 struct receive_queue *rq = &vi->rq[i];
1393
1394                 napi_disable(&rq->napi);
1395                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1396                 virtnet_napi_enable(rq->vq, &rq->napi);
1397
1398                 /* In theory, this can happen: if we don't get any buffers in
1399                  * we will *never* try to fill again.
1400                  */
1401                 if (still_empty)
1402                         schedule_delayed_work(&vi->refill, HZ/2);
1403         }
1404 }
1405
1406 static int virtnet_receive(struct receive_queue *rq, int budget,
1407                            unsigned int *xdp_xmit)
1408 {
1409         struct virtnet_info *vi = rq->vq->vdev->priv;
1410         struct virtnet_rq_stats stats = {};
1411         unsigned int len;
1412         void *buf;
1413         int i;
1414
1415         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1416                 void *ctx;
1417
1418                 while (stats.packets < budget &&
1419                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1420                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1421                         stats.packets++;
1422                 }
1423         } else {
1424                 while (stats.packets < budget &&
1425                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1426                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1427                         stats.packets++;
1428                 }
1429         }
1430
1431         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1432                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1433                         schedule_delayed_work(&vi->refill, 0);
1434         }
1435
1436         u64_stats_update_begin(&rq->stats.syncp);
1437         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1438                 size_t offset = virtnet_rq_stats_desc[i].offset;
1439                 u64 *item;
1440
1441                 item = (u64 *)((u8 *)&rq->stats + offset);
1442                 *item += *(u64 *)((u8 *)&stats + offset);
1443         }
1444         u64_stats_update_end(&rq->stats.syncp);
1445
1446         return stats.packets;
1447 }
1448
1449 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1450 {
1451         unsigned int len;
1452         unsigned int packets = 0;
1453         unsigned int bytes = 0;
1454         void *ptr;
1455
1456         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1457                 if (likely(!is_xdp_frame(ptr))) {
1458                         struct sk_buff *skb = ptr;
1459
1460                         pr_debug("Sent skb %p\n", skb);
1461
1462                         bytes += skb->len;
1463                         napi_consume_skb(skb, in_napi);
1464                 } else {
1465                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1466
1467                         bytes += frame->len;
1468                         xdp_return_frame(frame);
1469                 }
1470                 packets++;
1471         }
1472
1473         /* Avoid overhead when no packets have been processed
1474          * happens when called speculatively from start_xmit.
1475          */
1476         if (!packets)
1477                 return;
1478
1479         u64_stats_update_begin(&sq->stats.syncp);
1480         sq->stats.bytes += bytes;
1481         sq->stats.packets += packets;
1482         u64_stats_update_end(&sq->stats.syncp);
1483 }
1484
1485 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1486 {
1487         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1488                 return false;
1489         else if (q < vi->curr_queue_pairs)
1490                 return true;
1491         else
1492                 return false;
1493 }
1494
1495 static void virtnet_poll_cleantx(struct receive_queue *rq)
1496 {
1497         struct virtnet_info *vi = rq->vq->vdev->priv;
1498         unsigned int index = vq2rxq(rq->vq);
1499         struct send_queue *sq = &vi->sq[index];
1500         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1501
1502         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1503                 return;
1504
1505         if (__netif_tx_trylock(txq)) {
1506                 free_old_xmit_skbs(sq, true);
1507                 __netif_tx_unlock(txq);
1508         }
1509
1510         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1511                 netif_tx_wake_queue(txq);
1512 }
1513
1514 static int virtnet_poll(struct napi_struct *napi, int budget)
1515 {
1516         struct receive_queue *rq =
1517                 container_of(napi, struct receive_queue, napi);
1518         struct virtnet_info *vi = rq->vq->vdev->priv;
1519         struct send_queue *sq;
1520         unsigned int received;
1521         unsigned int xdp_xmit = 0;
1522
1523         virtnet_poll_cleantx(rq);
1524
1525         received = virtnet_receive(rq, budget, &xdp_xmit);
1526
1527         /* Out of packets? */
1528         if (received < budget)
1529                 virtqueue_napi_complete(napi, rq->vq, received);
1530
1531         if (xdp_xmit & VIRTIO_XDP_REDIR)
1532                 xdp_do_flush();
1533
1534         if (xdp_xmit & VIRTIO_XDP_TX) {
1535                 sq = virtnet_xdp_get_sq(vi);
1536                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1537                         u64_stats_update_begin(&sq->stats.syncp);
1538                         sq->stats.kicks++;
1539                         u64_stats_update_end(&sq->stats.syncp);
1540                 }
1541                 virtnet_xdp_put_sq(vi, sq);
1542         }
1543
1544         return received;
1545 }
1546
1547 static int virtnet_open(struct net_device *dev)
1548 {
1549         struct virtnet_info *vi = netdev_priv(dev);
1550         int i, err;
1551
1552         for (i = 0; i < vi->max_queue_pairs; i++) {
1553                 if (i < vi->curr_queue_pairs)
1554                         /* Make sure we have some buffers: if oom use wq. */
1555                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1556                                 schedule_delayed_work(&vi->refill, 0);
1557
1558                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1559                 if (err < 0)
1560                         return err;
1561
1562                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1563                                                  MEM_TYPE_PAGE_SHARED, NULL);
1564                 if (err < 0) {
1565                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1566                         return err;
1567                 }
1568
1569                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1570                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1571         }
1572
1573         return 0;
1574 }
1575
1576 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1577 {
1578         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1579         struct virtnet_info *vi = sq->vq->vdev->priv;
1580         unsigned int index = vq2txq(sq->vq);
1581         struct netdev_queue *txq;
1582
1583         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1584                 /* We don't need to enable cb for XDP */
1585                 napi_complete_done(napi, 0);
1586                 return 0;
1587         }
1588
1589         txq = netdev_get_tx_queue(vi->dev, index);
1590         __netif_tx_lock(txq, raw_smp_processor_id());
1591         free_old_xmit_skbs(sq, true);
1592         __netif_tx_unlock(txq);
1593
1594         virtqueue_napi_complete(napi, sq->vq, 0);
1595
1596         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1597                 netif_tx_wake_queue(txq);
1598
1599         return 0;
1600 }
1601
1602 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1603 {
1604         struct virtio_net_hdr_mrg_rxbuf *hdr;
1605         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1606         struct virtnet_info *vi = sq->vq->vdev->priv;
1607         int num_sg;
1608         unsigned hdr_len = vi->hdr_len;
1609         bool can_push;
1610
1611         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1612
1613         can_push = vi->any_header_sg &&
1614                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1615                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1616         /* Even if we can, don't push here yet as this would skew
1617          * csum_start offset below. */
1618         if (can_push)
1619                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1620         else
1621                 hdr = skb_vnet_hdr(skb);
1622
1623         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1624                                     virtio_is_little_endian(vi->vdev), false,
1625                                     0))
1626                 BUG();
1627
1628         if (vi->mergeable_rx_bufs)
1629                 hdr->num_buffers = 0;
1630
1631         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1632         if (can_push) {
1633                 __skb_push(skb, hdr_len);
1634                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1635                 if (unlikely(num_sg < 0))
1636                         return num_sg;
1637                 /* Pull header back to avoid skew in tx bytes calculations. */
1638                 __skb_pull(skb, hdr_len);
1639         } else {
1640                 sg_set_buf(sq->sg, hdr, hdr_len);
1641                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1642                 if (unlikely(num_sg < 0))
1643                         return num_sg;
1644                 num_sg++;
1645         }
1646         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1647 }
1648
1649 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1650 {
1651         struct virtnet_info *vi = netdev_priv(dev);
1652         int qnum = skb_get_queue_mapping(skb);
1653         struct send_queue *sq = &vi->sq[qnum];
1654         int err;
1655         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1656         bool kick = !netdev_xmit_more();
1657         bool use_napi = sq->napi.weight;
1658
1659         /* Free up any pending old buffers before queueing new ones. */
1660         free_old_xmit_skbs(sq, false);
1661
1662         if (use_napi && kick)
1663                 virtqueue_enable_cb_delayed(sq->vq);
1664
1665         /* timestamp packet in software */
1666         skb_tx_timestamp(skb);
1667
1668         /* Try to transmit */
1669         err = xmit_skb(sq, skb);
1670
1671         /* This should not happen! */
1672         if (unlikely(err)) {
1673                 dev->stats.tx_fifo_errors++;
1674                 if (net_ratelimit())
1675                         dev_warn(&dev->dev,
1676                                  "Unexpected TXQ (%d) queue failure: %d\n",
1677                                  qnum, err);
1678                 dev->stats.tx_dropped++;
1679                 dev_kfree_skb_any(skb);
1680                 return NETDEV_TX_OK;
1681         }
1682
1683         /* Don't wait up for transmitted skbs to be freed. */
1684         if (!use_napi) {
1685                 skb_orphan(skb);
1686                 nf_reset_ct(skb);
1687         }
1688
1689         /* If running out of space, stop queue to avoid getting packets that we
1690          * are then unable to transmit.
1691          * An alternative would be to force queuing layer to requeue the skb by
1692          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1693          * returned in a normal path of operation: it means that driver is not
1694          * maintaining the TX queue stop/start state properly, and causes
1695          * the stack to do a non-trivial amount of useless work.
1696          * Since most packets only take 1 or 2 ring slots, stopping the queue
1697          * early means 16 slots are typically wasted.
1698          */
1699         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1700                 netif_stop_subqueue(dev, qnum);
1701                 if (!use_napi &&
1702                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1703                         /* More just got used, free them then recheck. */
1704                         free_old_xmit_skbs(sq, false);
1705                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1706                                 netif_start_subqueue(dev, qnum);
1707                                 virtqueue_disable_cb(sq->vq);
1708                         }
1709                 }
1710         }
1711
1712         if (kick || netif_xmit_stopped(txq)) {
1713                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1714                         u64_stats_update_begin(&sq->stats.syncp);
1715                         sq->stats.kicks++;
1716                         u64_stats_update_end(&sq->stats.syncp);
1717                 }
1718         }
1719
1720         return NETDEV_TX_OK;
1721 }
1722
1723 /*
1724  * Send command via the control virtqueue and check status.  Commands
1725  * supported by the hypervisor, as indicated by feature bits, should
1726  * never fail unless improperly formatted.
1727  */
1728 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1729                                  struct scatterlist *out)
1730 {
1731         struct scatterlist *sgs[4], hdr, stat;
1732         unsigned out_num = 0, tmp;
1733
1734         /* Caller should know better */
1735         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1736
1737         vi->ctrl->status = ~0;
1738         vi->ctrl->hdr.class = class;
1739         vi->ctrl->hdr.cmd = cmd;
1740         /* Add header */
1741         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1742         sgs[out_num++] = &hdr;
1743
1744         if (out)
1745                 sgs[out_num++] = out;
1746
1747         /* Add return status. */
1748         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1749         sgs[out_num] = &stat;
1750
1751         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1752         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1753
1754         if (unlikely(!virtqueue_kick(vi->cvq)))
1755                 return vi->ctrl->status == VIRTIO_NET_OK;
1756
1757         /* Spin for a response, the kick causes an ioport write, trapping
1758          * into the hypervisor, so the request should be handled immediately.
1759          */
1760         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1761                !virtqueue_is_broken(vi->cvq))
1762                 cpu_relax();
1763
1764         return vi->ctrl->status == VIRTIO_NET_OK;
1765 }
1766
1767 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1768 {
1769         struct virtnet_info *vi = netdev_priv(dev);
1770         struct virtio_device *vdev = vi->vdev;
1771         int ret;
1772         struct sockaddr *addr;
1773         struct scatterlist sg;
1774
1775         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1776                 return -EOPNOTSUPP;
1777
1778         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1779         if (!addr)
1780                 return -ENOMEM;
1781
1782         ret = eth_prepare_mac_addr_change(dev, addr);
1783         if (ret)
1784                 goto out;
1785
1786         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1787                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1788                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1789                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1790                         dev_warn(&vdev->dev,
1791                                  "Failed to set mac address by vq command.\n");
1792                         ret = -EINVAL;
1793                         goto out;
1794                 }
1795         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1796                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1797                 unsigned int i;
1798
1799                 /* Naturally, this has an atomicity problem. */
1800                 for (i = 0; i < dev->addr_len; i++)
1801                         virtio_cwrite8(vdev,
1802                                        offsetof(struct virtio_net_config, mac) +
1803                                        i, addr->sa_data[i]);
1804         }
1805
1806         eth_commit_mac_addr_change(dev, p);
1807         ret = 0;
1808
1809 out:
1810         kfree(addr);
1811         return ret;
1812 }
1813
1814 static void virtnet_stats(struct net_device *dev,
1815                           struct rtnl_link_stats64 *tot)
1816 {
1817         struct virtnet_info *vi = netdev_priv(dev);
1818         unsigned int start;
1819         int i;
1820
1821         for (i = 0; i < vi->max_queue_pairs; i++) {
1822                 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1823                 struct receive_queue *rq = &vi->rq[i];
1824                 struct send_queue *sq = &vi->sq[i];
1825
1826                 do {
1827                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1828                         tpackets = sq->stats.packets;
1829                         tbytes   = sq->stats.bytes;
1830                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1831
1832                 do {
1833                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1834                         rpackets = rq->stats.packets;
1835                         rbytes   = rq->stats.bytes;
1836                         rdrops   = rq->stats.drops;
1837                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1838
1839                 tot->rx_packets += rpackets;
1840                 tot->tx_packets += tpackets;
1841                 tot->rx_bytes   += rbytes;
1842                 tot->tx_bytes   += tbytes;
1843                 tot->rx_dropped += rdrops;
1844         }
1845
1846         tot->tx_dropped = dev->stats.tx_dropped;
1847         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1848         tot->rx_length_errors = dev->stats.rx_length_errors;
1849         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1850 }
1851
1852 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1853 {
1854         rtnl_lock();
1855         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1856                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1857                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1858         rtnl_unlock();
1859 }
1860
1861 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1862 {
1863         struct scatterlist sg;
1864         struct net_device *dev = vi->dev;
1865
1866         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1867                 return 0;
1868
1869         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1870         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1871
1872         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1873                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1874                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1875                          queue_pairs);
1876                 return -EINVAL;
1877         } else {
1878                 vi->curr_queue_pairs = queue_pairs;
1879                 /* virtnet_open() will refill when device is going to up. */
1880                 if (dev->flags & IFF_UP)
1881                         schedule_delayed_work(&vi->refill, 0);
1882         }
1883
1884         return 0;
1885 }
1886
1887 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1888 {
1889         int err;
1890
1891         rtnl_lock();
1892         err = _virtnet_set_queues(vi, queue_pairs);
1893         rtnl_unlock();
1894         return err;
1895 }
1896
1897 static int virtnet_close(struct net_device *dev)
1898 {
1899         struct virtnet_info *vi = netdev_priv(dev);
1900         int i;
1901
1902         /* Make sure refill_work doesn't re-enable napi! */
1903         cancel_delayed_work_sync(&vi->refill);
1904
1905         for (i = 0; i < vi->max_queue_pairs; i++) {
1906                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1907                 napi_disable(&vi->rq[i].napi);
1908                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1909         }
1910
1911         return 0;
1912 }
1913
1914 static void virtnet_set_rx_mode(struct net_device *dev)
1915 {
1916         struct virtnet_info *vi = netdev_priv(dev);
1917         struct scatterlist sg[2];
1918         struct virtio_net_ctrl_mac *mac_data;
1919         struct netdev_hw_addr *ha;
1920         int uc_count;
1921         int mc_count;
1922         void *buf;
1923         int i;
1924
1925         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1926         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1927                 return;
1928
1929         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1930         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1931
1932         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1933
1934         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1935                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1936                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1937                          vi->ctrl->promisc ? "en" : "dis");
1938
1939         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1940
1941         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1942                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1943                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1944                          vi->ctrl->allmulti ? "en" : "dis");
1945
1946         uc_count = netdev_uc_count(dev);
1947         mc_count = netdev_mc_count(dev);
1948         /* MAC filter - use one buffer for both lists */
1949         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1950                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1951         mac_data = buf;
1952         if (!buf)
1953                 return;
1954
1955         sg_init_table(sg, 2);
1956
1957         /* Store the unicast list and count in the front of the buffer */
1958         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1959         i = 0;
1960         netdev_for_each_uc_addr(ha, dev)
1961                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1962
1963         sg_set_buf(&sg[0], mac_data,
1964                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1965
1966         /* multicast list and count fill the end */
1967         mac_data = (void *)&mac_data->macs[uc_count][0];
1968
1969         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1970         i = 0;
1971         netdev_for_each_mc_addr(ha, dev)
1972                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1973
1974         sg_set_buf(&sg[1], mac_data,
1975                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1976
1977         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1978                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1979                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1980
1981         kfree(buf);
1982 }
1983
1984 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1985                                    __be16 proto, u16 vid)
1986 {
1987         struct virtnet_info *vi = netdev_priv(dev);
1988         struct scatterlist sg;
1989
1990         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1991         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1992
1993         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1994                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1995                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1996         return 0;
1997 }
1998
1999 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2000                                     __be16 proto, u16 vid)
2001 {
2002         struct virtnet_info *vi = netdev_priv(dev);
2003         struct scatterlist sg;
2004
2005         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2006         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2007
2008         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2009                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2010                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2011         return 0;
2012 }
2013
2014 static void virtnet_clean_affinity(struct virtnet_info *vi)
2015 {
2016         int i;
2017
2018         if (vi->affinity_hint_set) {
2019                 for (i = 0; i < vi->max_queue_pairs; i++) {
2020                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
2021                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
2022                 }
2023
2024                 vi->affinity_hint_set = false;
2025         }
2026 }
2027
2028 static void virtnet_set_affinity(struct virtnet_info *vi)
2029 {
2030         cpumask_var_t mask;
2031         int stragglers;
2032         int group_size;
2033         int i, j, cpu;
2034         int num_cpu;
2035         int stride;
2036
2037         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2038                 virtnet_clean_affinity(vi);
2039                 return;
2040         }
2041
2042         num_cpu = num_online_cpus();
2043         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2044         stragglers = num_cpu >= vi->curr_queue_pairs ?
2045                         num_cpu % vi->curr_queue_pairs :
2046                         0;
2047         cpu = cpumask_next(-1, cpu_online_mask);
2048
2049         for (i = 0; i < vi->curr_queue_pairs; i++) {
2050                 group_size = stride + (i < stragglers ? 1 : 0);
2051
2052                 for (j = 0; j < group_size; j++) {
2053                         cpumask_set_cpu(cpu, mask);
2054                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2055                                                 nr_cpu_ids, false);
2056                 }
2057                 virtqueue_set_affinity(vi->rq[i].vq, mask);
2058                 virtqueue_set_affinity(vi->sq[i].vq, mask);
2059                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2060                 cpumask_clear(mask);
2061         }
2062
2063         vi->affinity_hint_set = true;
2064         free_cpumask_var(mask);
2065 }
2066
2067 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2068 {
2069         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2070                                                    node);
2071         virtnet_set_affinity(vi);
2072         return 0;
2073 }
2074
2075 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2076 {
2077         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2078                                                    node_dead);
2079         virtnet_set_affinity(vi);
2080         return 0;
2081 }
2082
2083 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2084 {
2085         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2086                                                    node);
2087
2088         virtnet_clean_affinity(vi);
2089         return 0;
2090 }
2091
2092 static enum cpuhp_state virtionet_online;
2093
2094 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2095 {
2096         int ret;
2097
2098         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2099         if (ret)
2100                 return ret;
2101         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2102                                                &vi->node_dead);
2103         if (!ret)
2104                 return ret;
2105         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2106         return ret;
2107 }
2108
2109 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2110 {
2111         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2112         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2113                                             &vi->node_dead);
2114 }
2115
2116 static void virtnet_get_ringparam(struct net_device *dev,
2117                                 struct ethtool_ringparam *ring)
2118 {
2119         struct virtnet_info *vi = netdev_priv(dev);
2120
2121         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2122         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2123         ring->rx_pending = ring->rx_max_pending;
2124         ring->tx_pending = ring->tx_max_pending;
2125 }
2126
2127
2128 static void virtnet_get_drvinfo(struct net_device *dev,
2129                                 struct ethtool_drvinfo *info)
2130 {
2131         struct virtnet_info *vi = netdev_priv(dev);
2132         struct virtio_device *vdev = vi->vdev;
2133
2134         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2135         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2136         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2137
2138 }
2139
2140 /* TODO: Eliminate OOO packets during switching */
2141 static int virtnet_set_channels(struct net_device *dev,
2142                                 struct ethtool_channels *channels)
2143 {
2144         struct virtnet_info *vi = netdev_priv(dev);
2145         u16 queue_pairs = channels->combined_count;
2146         int err;
2147
2148         /* We don't support separate rx/tx channels.
2149          * We don't allow setting 'other' channels.
2150          */
2151         if (channels->rx_count || channels->tx_count || channels->other_count)
2152                 return -EINVAL;
2153
2154         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2155                 return -EINVAL;
2156
2157         /* For now we don't support modifying channels while XDP is loaded
2158          * also when XDP is loaded all RX queues have XDP programs so we only
2159          * need to check a single RX queue.
2160          */
2161         if (vi->rq[0].xdp_prog)
2162                 return -EINVAL;
2163
2164         get_online_cpus();
2165         err = _virtnet_set_queues(vi, queue_pairs);
2166         if (err) {
2167                 put_online_cpus();
2168                 goto err;
2169         }
2170         virtnet_set_affinity(vi);
2171         put_online_cpus();
2172
2173         netif_set_real_num_tx_queues(dev, queue_pairs);
2174         netif_set_real_num_rx_queues(dev, queue_pairs);
2175  err:
2176         return err;
2177 }
2178
2179 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2180 {
2181         struct virtnet_info *vi = netdev_priv(dev);
2182         unsigned int i, j;
2183         u8 *p = data;
2184
2185         switch (stringset) {
2186         case ETH_SS_STATS:
2187                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2188                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2189                                 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2190                                                 virtnet_rq_stats_desc[j].desc);
2191                 }
2192
2193                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2194                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2195                                 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2196                                                 virtnet_sq_stats_desc[j].desc);
2197                 }
2198                 break;
2199         }
2200 }
2201
2202 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2203 {
2204         struct virtnet_info *vi = netdev_priv(dev);
2205
2206         switch (sset) {
2207         case ETH_SS_STATS:
2208                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2209                                                VIRTNET_SQ_STATS_LEN);
2210         default:
2211                 return -EOPNOTSUPP;
2212         }
2213 }
2214
2215 static void virtnet_get_ethtool_stats(struct net_device *dev,
2216                                       struct ethtool_stats *stats, u64 *data)
2217 {
2218         struct virtnet_info *vi = netdev_priv(dev);
2219         unsigned int idx = 0, start, i, j;
2220         const u8 *stats_base;
2221         size_t offset;
2222
2223         for (i = 0; i < vi->curr_queue_pairs; i++) {
2224                 struct receive_queue *rq = &vi->rq[i];
2225
2226                 stats_base = (u8 *)&rq->stats;
2227                 do {
2228                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2229                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2230                                 offset = virtnet_rq_stats_desc[j].offset;
2231                                 data[idx + j] = *(u64 *)(stats_base + offset);
2232                         }
2233                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2234                 idx += VIRTNET_RQ_STATS_LEN;
2235         }
2236
2237         for (i = 0; i < vi->curr_queue_pairs; i++) {
2238                 struct send_queue *sq = &vi->sq[i];
2239
2240                 stats_base = (u8 *)&sq->stats;
2241                 do {
2242                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2243                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2244                                 offset = virtnet_sq_stats_desc[j].offset;
2245                                 data[idx + j] = *(u64 *)(stats_base + offset);
2246                         }
2247                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2248                 idx += VIRTNET_SQ_STATS_LEN;
2249         }
2250 }
2251
2252 static void virtnet_get_channels(struct net_device *dev,
2253                                  struct ethtool_channels *channels)
2254 {
2255         struct virtnet_info *vi = netdev_priv(dev);
2256
2257         channels->combined_count = vi->curr_queue_pairs;
2258         channels->max_combined = vi->max_queue_pairs;
2259         channels->max_other = 0;
2260         channels->rx_count = 0;
2261         channels->tx_count = 0;
2262         channels->other_count = 0;
2263 }
2264
2265 static int virtnet_set_link_ksettings(struct net_device *dev,
2266                                       const struct ethtool_link_ksettings *cmd)
2267 {
2268         struct virtnet_info *vi = netdev_priv(dev);
2269
2270         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2271                                                   &vi->speed, &vi->duplex);
2272 }
2273
2274 static int virtnet_get_link_ksettings(struct net_device *dev,
2275                                       struct ethtool_link_ksettings *cmd)
2276 {
2277         struct virtnet_info *vi = netdev_priv(dev);
2278
2279         cmd->base.speed = vi->speed;
2280         cmd->base.duplex = vi->duplex;
2281         cmd->base.port = PORT_OTHER;
2282
2283         return 0;
2284 }
2285
2286 static int virtnet_set_coalesce(struct net_device *dev,
2287                                 struct ethtool_coalesce *ec)
2288 {
2289         struct virtnet_info *vi = netdev_priv(dev);
2290         int i, napi_weight;
2291
2292         if (ec->tx_max_coalesced_frames > 1 ||
2293             ec->rx_max_coalesced_frames != 1)
2294                 return -EINVAL;
2295
2296         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2297         if (napi_weight ^ vi->sq[0].napi.weight) {
2298                 if (dev->flags & IFF_UP)
2299                         return -EBUSY;
2300                 for (i = 0; i < vi->max_queue_pairs; i++)
2301                         vi->sq[i].napi.weight = napi_weight;
2302         }
2303
2304         return 0;
2305 }
2306
2307 static int virtnet_get_coalesce(struct net_device *dev,
2308                                 struct ethtool_coalesce *ec)
2309 {
2310         struct ethtool_coalesce ec_default = {
2311                 .cmd = ETHTOOL_GCOALESCE,
2312                 .rx_max_coalesced_frames = 1,
2313         };
2314         struct virtnet_info *vi = netdev_priv(dev);
2315
2316         memcpy(ec, &ec_default, sizeof(ec_default));
2317
2318         if (vi->sq[0].napi.weight)
2319                 ec->tx_max_coalesced_frames = 1;
2320
2321         return 0;
2322 }
2323
2324 static void virtnet_init_settings(struct net_device *dev)
2325 {
2326         struct virtnet_info *vi = netdev_priv(dev);
2327
2328         vi->speed = SPEED_UNKNOWN;
2329         vi->duplex = DUPLEX_UNKNOWN;
2330 }
2331
2332 static void virtnet_update_settings(struct virtnet_info *vi)
2333 {
2334         u32 speed;
2335         u8 duplex;
2336
2337         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2338                 return;
2339
2340         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2341
2342         if (ethtool_validate_speed(speed))
2343                 vi->speed = speed;
2344
2345         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2346
2347         if (ethtool_validate_duplex(duplex))
2348                 vi->duplex = duplex;
2349 }
2350
2351 static const struct ethtool_ops virtnet_ethtool_ops = {
2352         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2353         .get_drvinfo = virtnet_get_drvinfo,
2354         .get_link = ethtool_op_get_link,
2355         .get_ringparam = virtnet_get_ringparam,
2356         .get_strings = virtnet_get_strings,
2357         .get_sset_count = virtnet_get_sset_count,
2358         .get_ethtool_stats = virtnet_get_ethtool_stats,
2359         .set_channels = virtnet_set_channels,
2360         .get_channels = virtnet_get_channels,
2361         .get_ts_info = ethtool_op_get_ts_info,
2362         .get_link_ksettings = virtnet_get_link_ksettings,
2363         .set_link_ksettings = virtnet_set_link_ksettings,
2364         .set_coalesce = virtnet_set_coalesce,
2365         .get_coalesce = virtnet_get_coalesce,
2366 };
2367
2368 static void virtnet_freeze_down(struct virtio_device *vdev)
2369 {
2370         struct virtnet_info *vi = vdev->priv;
2371         int i;
2372
2373         /* Make sure no work handler is accessing the device */
2374         flush_work(&vi->config_work);
2375
2376         netif_tx_lock_bh(vi->dev);
2377         netif_device_detach(vi->dev);
2378         netif_tx_unlock_bh(vi->dev);
2379         cancel_delayed_work_sync(&vi->refill);
2380
2381         if (netif_running(vi->dev)) {
2382                 for (i = 0; i < vi->max_queue_pairs; i++) {
2383                         napi_disable(&vi->rq[i].napi);
2384                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2385                 }
2386         }
2387 }
2388
2389 static int init_vqs(struct virtnet_info *vi);
2390
2391 static int virtnet_restore_up(struct virtio_device *vdev)
2392 {
2393         struct virtnet_info *vi = vdev->priv;
2394         int err, i;
2395
2396         err = init_vqs(vi);
2397         if (err)
2398                 return err;
2399
2400         virtio_device_ready(vdev);
2401
2402         if (netif_running(vi->dev)) {
2403                 for (i = 0; i < vi->curr_queue_pairs; i++)
2404                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2405                                 schedule_delayed_work(&vi->refill, 0);
2406
2407                 for (i = 0; i < vi->max_queue_pairs; i++) {
2408                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2409                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2410                                                &vi->sq[i].napi);
2411                 }
2412         }
2413
2414         netif_tx_lock_bh(vi->dev);
2415         netif_device_attach(vi->dev);
2416         netif_tx_unlock_bh(vi->dev);
2417         return err;
2418 }
2419
2420 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2421 {
2422         struct scatterlist sg;
2423         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2424
2425         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2426
2427         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2428                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2429                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2430                 return -EINVAL;
2431         }
2432
2433         return 0;
2434 }
2435
2436 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2437 {
2438         u64 offloads = 0;
2439
2440         if (!vi->guest_offloads)
2441                 return 0;
2442
2443         return virtnet_set_guest_offloads(vi, offloads);
2444 }
2445
2446 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2447 {
2448         u64 offloads = vi->guest_offloads;
2449
2450         if (!vi->guest_offloads)
2451                 return 0;
2452
2453         return virtnet_set_guest_offloads(vi, offloads);
2454 }
2455
2456 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2457                            struct netlink_ext_ack *extack)
2458 {
2459         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2460         struct virtnet_info *vi = netdev_priv(dev);
2461         struct bpf_prog *old_prog;
2462         u16 xdp_qp = 0, curr_qp;
2463         int i, err;
2464
2465         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2466             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2467                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2468                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2469                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2470                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2471                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2472                 return -EOPNOTSUPP;
2473         }
2474
2475         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2476                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2477                 return -EINVAL;
2478         }
2479
2480         if (dev->mtu > max_sz) {
2481                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2482                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2483                 return -EINVAL;
2484         }
2485
2486         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2487         if (prog)
2488                 xdp_qp = nr_cpu_ids;
2489
2490         /* XDP requires extra queues for XDP_TX */
2491         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2492                 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",
2493                             curr_qp + xdp_qp, vi->max_queue_pairs);
2494                 xdp_qp = 0;
2495         }
2496
2497         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2498         if (!prog && !old_prog)
2499                 return 0;
2500
2501         if (prog)
2502                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2503
2504         /* Make sure NAPI is not using any XDP TX queues for RX. */
2505         if (netif_running(dev)) {
2506                 for (i = 0; i < vi->max_queue_pairs; i++) {
2507                         napi_disable(&vi->rq[i].napi);
2508                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2509                 }
2510         }
2511
2512         if (!prog) {
2513                 for (i = 0; i < vi->max_queue_pairs; i++) {
2514                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2515                         if (i == 0)
2516                                 virtnet_restore_guest_offloads(vi);
2517                 }
2518                 synchronize_net();
2519         }
2520
2521         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2522         if (err)
2523                 goto err;
2524         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2525         vi->xdp_queue_pairs = xdp_qp;
2526
2527         if (prog) {
2528                 vi->xdp_enabled = true;
2529                 for (i = 0; i < vi->max_queue_pairs; i++) {
2530                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2531                         if (i == 0 && !old_prog)
2532                                 virtnet_clear_guest_offloads(vi);
2533                 }
2534         } else {
2535                 vi->xdp_enabled = false;
2536         }
2537
2538         for (i = 0; i < vi->max_queue_pairs; i++) {
2539                 if (old_prog)
2540                         bpf_prog_put(old_prog);
2541                 if (netif_running(dev)) {
2542                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2543                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2544                                                &vi->sq[i].napi);
2545                 }
2546         }
2547
2548         return 0;
2549
2550 err:
2551         if (!prog) {
2552                 virtnet_clear_guest_offloads(vi);
2553                 for (i = 0; i < vi->max_queue_pairs; i++)
2554                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2555         }
2556
2557         if (netif_running(dev)) {
2558                 for (i = 0; i < vi->max_queue_pairs; i++) {
2559                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2560                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2561                                                &vi->sq[i].napi);
2562                 }
2563         }
2564         if (prog)
2565                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2566         return err;
2567 }
2568
2569 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2570 {
2571         switch (xdp->command) {
2572         case XDP_SETUP_PROG:
2573                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2574         default:
2575                 return -EINVAL;
2576         }
2577 }
2578
2579 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2580                                       size_t len)
2581 {
2582         struct virtnet_info *vi = netdev_priv(dev);
2583         int ret;
2584
2585         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2586                 return -EOPNOTSUPP;
2587
2588         ret = snprintf(buf, len, "sby");
2589         if (ret >= len)
2590                 return -EOPNOTSUPP;
2591
2592         return 0;
2593 }
2594
2595 static int virtnet_set_features(struct net_device *dev,
2596                                 netdev_features_t features)
2597 {
2598         struct virtnet_info *vi = netdev_priv(dev);
2599         u64 offloads;
2600         int err;
2601
2602         if ((dev->features ^ features) & NETIF_F_LRO) {
2603                 if (vi->xdp_enabled)
2604                         return -EBUSY;
2605
2606                 if (features & NETIF_F_LRO)
2607                         offloads = vi->guest_offloads_capable;
2608                 else
2609                         offloads = vi->guest_offloads_capable &
2610                                    ~GUEST_OFFLOAD_LRO_MASK;
2611
2612                 err = virtnet_set_guest_offloads(vi, offloads);
2613                 if (err)
2614                         return err;
2615                 vi->guest_offloads = offloads;
2616         }
2617
2618         return 0;
2619 }
2620
2621 static const struct net_device_ops virtnet_netdev = {
2622         .ndo_open            = virtnet_open,
2623         .ndo_stop            = virtnet_close,
2624         .ndo_start_xmit      = start_xmit,
2625         .ndo_validate_addr   = eth_validate_addr,
2626         .ndo_set_mac_address = virtnet_set_mac_address,
2627         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2628         .ndo_get_stats64     = virtnet_stats,
2629         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2630         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2631         .ndo_bpf                = virtnet_xdp,
2632         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2633         .ndo_features_check     = passthru_features_check,
2634         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2635         .ndo_set_features       = virtnet_set_features,
2636 };
2637
2638 static void virtnet_config_changed_work(struct work_struct *work)
2639 {
2640         struct virtnet_info *vi =
2641                 container_of(work, struct virtnet_info, config_work);
2642         u16 v;
2643
2644         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2645                                  struct virtio_net_config, status, &v) < 0)
2646                 return;
2647
2648         if (v & VIRTIO_NET_S_ANNOUNCE) {
2649                 netdev_notify_peers(vi->dev);
2650                 virtnet_ack_link_announce(vi);
2651         }
2652
2653         /* Ignore unknown (future) status bits */
2654         v &= VIRTIO_NET_S_LINK_UP;
2655
2656         if (vi->status == v)
2657                 return;
2658
2659         vi->status = v;
2660
2661         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2662                 virtnet_update_settings(vi);
2663                 netif_carrier_on(vi->dev);
2664                 netif_tx_wake_all_queues(vi->dev);
2665         } else {
2666                 netif_carrier_off(vi->dev);
2667                 netif_tx_stop_all_queues(vi->dev);
2668         }
2669 }
2670
2671 static void virtnet_config_changed(struct virtio_device *vdev)
2672 {
2673         struct virtnet_info *vi = vdev->priv;
2674
2675         schedule_work(&vi->config_work);
2676 }
2677
2678 static void virtnet_free_queues(struct virtnet_info *vi)
2679 {
2680         int i;
2681
2682         for (i = 0; i < vi->max_queue_pairs; i++) {
2683                 __netif_napi_del(&vi->rq[i].napi);
2684                 __netif_napi_del(&vi->sq[i].napi);
2685         }
2686
2687         /* We called __netif_napi_del(),
2688          * we need to respect an RCU grace period before freeing vi->rq
2689          */
2690         synchronize_net();
2691
2692         kfree(vi->rq);
2693         kfree(vi->sq);
2694         kfree(vi->ctrl);
2695 }
2696
2697 static void _free_receive_bufs(struct virtnet_info *vi)
2698 {
2699         struct bpf_prog *old_prog;
2700         int i;
2701
2702         for (i = 0; i < vi->max_queue_pairs; i++) {
2703                 while (vi->rq[i].pages)
2704                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2705
2706                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2707                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2708                 if (old_prog)
2709                         bpf_prog_put(old_prog);
2710         }
2711 }
2712
2713 static void free_receive_bufs(struct virtnet_info *vi)
2714 {
2715         rtnl_lock();
2716         _free_receive_bufs(vi);
2717         rtnl_unlock();
2718 }
2719
2720 static void free_receive_page_frags(struct virtnet_info *vi)
2721 {
2722         int i;
2723         for (i = 0; i < vi->max_queue_pairs; i++)
2724                 if (vi->rq[i].alloc_frag.page)
2725                         put_page(vi->rq[i].alloc_frag.page);
2726 }
2727
2728 static void free_unused_bufs(struct virtnet_info *vi)
2729 {
2730         void *buf;
2731         int i;
2732
2733         for (i = 0; i < vi->max_queue_pairs; i++) {
2734                 struct virtqueue *vq = vi->sq[i].vq;
2735                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2736                         if (!is_xdp_frame(buf))
2737                                 dev_kfree_skb(buf);
2738                         else
2739                                 xdp_return_frame(ptr_to_xdp(buf));
2740                 }
2741         }
2742
2743         for (i = 0; i < vi->max_queue_pairs; i++) {
2744                 struct virtqueue *vq = vi->rq[i].vq;
2745
2746                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2747                         if (vi->mergeable_rx_bufs) {
2748                                 put_page(virt_to_head_page(buf));
2749                         } else if (vi->big_packets) {
2750                                 give_pages(&vi->rq[i], buf);
2751                         } else {
2752                                 put_page(virt_to_head_page(buf));
2753                         }
2754                 }
2755         }
2756 }
2757
2758 static void virtnet_del_vqs(struct virtnet_info *vi)
2759 {
2760         struct virtio_device *vdev = vi->vdev;
2761
2762         virtnet_clean_affinity(vi);
2763
2764         vdev->config->del_vqs(vdev);
2765
2766         virtnet_free_queues(vi);
2767 }
2768
2769 /* How large should a single buffer be so a queue full of these can fit at
2770  * least one full packet?
2771  * Logic below assumes the mergeable buffer header is used.
2772  */
2773 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2774 {
2775         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2776         unsigned int rq_size = virtqueue_get_vring_size(vq);
2777         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2778         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2779         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2780
2781         return max(max(min_buf_len, hdr_len) - hdr_len,
2782                    (unsigned int)GOOD_PACKET_LEN);
2783 }
2784
2785 static int virtnet_find_vqs(struct virtnet_info *vi)
2786 {
2787         vq_callback_t **callbacks;
2788         struct virtqueue **vqs;
2789         int ret = -ENOMEM;
2790         int i, total_vqs;
2791         const char **names;
2792         bool *ctx;
2793
2794         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2795          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2796          * possible control vq.
2797          */
2798         total_vqs = vi->max_queue_pairs * 2 +
2799                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2800
2801         /* Allocate space for find_vqs parameters */
2802         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2803         if (!vqs)
2804                 goto err_vq;
2805         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2806         if (!callbacks)
2807                 goto err_callback;
2808         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2809         if (!names)
2810                 goto err_names;
2811         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2812                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2813                 if (!ctx)
2814                         goto err_ctx;
2815         } else {
2816                 ctx = NULL;
2817         }
2818
2819         /* Parameters for control virtqueue, if any */
2820         if (vi->has_cvq) {
2821                 callbacks[total_vqs - 1] = NULL;
2822                 names[total_vqs - 1] = "control";
2823         }
2824
2825         /* Allocate/initialize parameters for send/receive virtqueues */
2826         for (i = 0; i < vi->max_queue_pairs; i++) {
2827                 callbacks[rxq2vq(i)] = skb_recv_done;
2828                 callbacks[txq2vq(i)] = skb_xmit_done;
2829                 sprintf(vi->rq[i].name, "input.%d", i);
2830                 sprintf(vi->sq[i].name, "output.%d", i);
2831                 names[rxq2vq(i)] = vi->rq[i].name;
2832                 names[txq2vq(i)] = vi->sq[i].name;
2833                 if (ctx)
2834                         ctx[rxq2vq(i)] = true;
2835         }
2836
2837         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2838                                          names, ctx, NULL);
2839         if (ret)
2840                 goto err_find;
2841
2842         if (vi->has_cvq) {
2843                 vi->cvq = vqs[total_vqs - 1];
2844                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2845                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2846         }
2847
2848         for (i = 0; i < vi->max_queue_pairs; i++) {
2849                 vi->rq[i].vq = vqs[rxq2vq(i)];
2850                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2851                 vi->sq[i].vq = vqs[txq2vq(i)];
2852         }
2853
2854         /* run here: ret == 0. */
2855
2856
2857 err_find:
2858         kfree(ctx);
2859 err_ctx:
2860         kfree(names);
2861 err_names:
2862         kfree(callbacks);
2863 err_callback:
2864         kfree(vqs);
2865 err_vq:
2866         return ret;
2867 }
2868
2869 static int virtnet_alloc_queues(struct virtnet_info *vi)
2870 {
2871         int i;
2872
2873         if (vi->has_cvq) {
2874                 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2875                 if (!vi->ctrl)
2876                         goto err_ctrl;
2877         } else {
2878                 vi->ctrl = NULL;
2879         }
2880         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2881         if (!vi->sq)
2882                 goto err_sq;
2883         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2884         if (!vi->rq)
2885                 goto err_rq;
2886
2887         INIT_DELAYED_WORK(&vi->refill, refill_work);
2888         for (i = 0; i < vi->max_queue_pairs; i++) {
2889                 vi->rq[i].pages = NULL;
2890                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2891                                napi_weight);
2892                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2893                                   napi_tx ? napi_weight : 0);
2894
2895                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2896                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2897                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2898
2899                 u64_stats_init(&vi->rq[i].stats.syncp);
2900                 u64_stats_init(&vi->sq[i].stats.syncp);
2901         }
2902
2903         return 0;
2904
2905 err_rq:
2906         kfree(vi->sq);
2907 err_sq:
2908         kfree(vi->ctrl);
2909 err_ctrl:
2910         return -ENOMEM;
2911 }
2912
2913 static int init_vqs(struct virtnet_info *vi)
2914 {
2915         int ret;
2916
2917         /* Allocate send & receive queues */
2918         ret = virtnet_alloc_queues(vi);
2919         if (ret)
2920                 goto err;
2921
2922         ret = virtnet_find_vqs(vi);
2923         if (ret)
2924                 goto err_free;
2925
2926         get_online_cpus();
2927         virtnet_set_affinity(vi);
2928         put_online_cpus();
2929
2930         return 0;
2931
2932 err_free:
2933         virtnet_free_queues(vi);
2934 err:
2935         return ret;
2936 }
2937
2938 #ifdef CONFIG_SYSFS
2939 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2940                 char *buf)
2941 {
2942         struct virtnet_info *vi = netdev_priv(queue->dev);
2943         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2944         unsigned int headroom = virtnet_get_headroom(vi);
2945         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2946         struct ewma_pkt_len *avg;
2947
2948         BUG_ON(queue_index >= vi->max_queue_pairs);
2949         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2950         return sprintf(buf, "%u\n",
2951                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
2952                                        SKB_DATA_ALIGN(headroom + tailroom)));
2953 }
2954
2955 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2956         __ATTR_RO(mergeable_rx_buffer_size);
2957
2958 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2959         &mergeable_rx_buffer_size_attribute.attr,
2960         NULL
2961 };
2962
2963 static const struct attribute_group virtio_net_mrg_rx_group = {
2964         .name = "virtio_net",
2965         .attrs = virtio_net_mrg_rx_attrs
2966 };
2967 #endif
2968
2969 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2970                                     unsigned int fbit,
2971                                     const char *fname, const char *dname)
2972 {
2973         if (!virtio_has_feature(vdev, fbit))
2974                 return false;
2975
2976         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2977                 fname, dname);
2978
2979         return true;
2980 }
2981
2982 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2983         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2984
2985 static bool virtnet_validate_features(struct virtio_device *vdev)
2986 {
2987         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2988             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2989                              "VIRTIO_NET_F_CTRL_VQ") ||
2990              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2991                              "VIRTIO_NET_F_CTRL_VQ") ||
2992              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2993                              "VIRTIO_NET_F_CTRL_VQ") ||
2994              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2995              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2996                              "VIRTIO_NET_F_CTRL_VQ"))) {
2997                 return false;
2998         }
2999
3000         return true;
3001 }
3002
3003 #define MIN_MTU ETH_MIN_MTU
3004 #define MAX_MTU ETH_MAX_MTU
3005
3006 static int virtnet_validate(struct virtio_device *vdev)
3007 {
3008         if (!vdev->config->get) {
3009                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3010                         __func__);
3011                 return -EINVAL;
3012         }
3013
3014         if (!virtnet_validate_features(vdev))
3015                 return -EINVAL;
3016
3017         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3018                 int mtu = virtio_cread16(vdev,
3019                                          offsetof(struct virtio_net_config,
3020                                                   mtu));
3021                 if (mtu < MIN_MTU)
3022                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3023         }
3024
3025         return 0;
3026 }
3027
3028 static int virtnet_probe(struct virtio_device *vdev)
3029 {
3030         int i, err = -ENOMEM;
3031         struct net_device *dev;
3032         struct virtnet_info *vi;
3033         u16 max_queue_pairs;
3034         int mtu;
3035
3036         /* Find if host supports multiqueue virtio_net device */
3037         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3038                                    struct virtio_net_config,
3039                                    max_virtqueue_pairs, &max_queue_pairs);
3040
3041         /* We need at least 2 queue's */
3042         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3043             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3044             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3045                 max_queue_pairs = 1;
3046
3047         /* Allocate ourselves a network device with room for our info */
3048         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3049         if (!dev)
3050                 return -ENOMEM;
3051
3052         /* Set up network device as normal. */
3053         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3054                            IFF_TX_SKB_NO_LINEAR;
3055         dev->netdev_ops = &virtnet_netdev;
3056         dev->features = NETIF_F_HIGHDMA;
3057
3058         dev->ethtool_ops = &virtnet_ethtool_ops;
3059         SET_NETDEV_DEV(dev, &vdev->dev);
3060
3061         /* Do we support "hardware" checksums? */
3062         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3063                 /* This opens up the world of extra features. */
3064                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3065                 if (csum)
3066                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3067
3068                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3069                         dev->hw_features |= NETIF_F_TSO
3070                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3071                 }
3072                 /* Individual feature bits: what can host handle? */
3073                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3074                         dev->hw_features |= NETIF_F_TSO;
3075                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3076                         dev->hw_features |= NETIF_F_TSO6;
3077                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3078                         dev->hw_features |= NETIF_F_TSO_ECN;
3079
3080                 dev->features |= NETIF_F_GSO_ROBUST;
3081
3082                 if (gso)
3083                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3084                 /* (!csum && gso) case will be fixed by register_netdev() */
3085         }
3086         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3087                 dev->features |= NETIF_F_RXCSUM;
3088         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3089             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3090                 dev->features |= NETIF_F_LRO;
3091         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3092                 dev->hw_features |= NETIF_F_LRO;
3093
3094         dev->vlan_features = dev->features;
3095
3096         /* MTU range: 68 - 65535 */
3097         dev->min_mtu = MIN_MTU;
3098         dev->max_mtu = MAX_MTU;
3099
3100         /* Configuration may specify what MAC to use.  Otherwise random. */
3101         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3102                 virtio_cread_bytes(vdev,
3103                                    offsetof(struct virtio_net_config, mac),
3104                                    dev->dev_addr, dev->addr_len);
3105         else
3106                 eth_hw_addr_random(dev);
3107
3108         /* Set up our device-specific information */
3109         vi = netdev_priv(dev);
3110         vi->dev = dev;
3111         vi->vdev = vdev;
3112         vdev->priv = vi;
3113
3114         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3115
3116         /* If we can receive ANY GSO packets, we must allocate large ones. */
3117         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3118             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3119             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3120             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3121                 vi->big_packets = true;
3122
3123         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3124                 vi->mergeable_rx_bufs = true;
3125
3126         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3127             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3128                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3129         else
3130                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3131
3132         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3133             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3134                 vi->any_header_sg = true;
3135
3136         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3137                 vi->has_cvq = true;
3138
3139         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3140                 mtu = virtio_cread16(vdev,
3141                                      offsetof(struct virtio_net_config,
3142                                               mtu));
3143                 if (mtu < dev->min_mtu) {
3144                         /* Should never trigger: MTU was previously validated
3145                          * in virtnet_validate.
3146                          */
3147                         dev_err(&vdev->dev,
3148                                 "device MTU appears to have changed it is now %d < %d",
3149                                 mtu, dev->min_mtu);
3150                         err = -EINVAL;
3151                         goto free;
3152                 }
3153
3154                 dev->mtu = mtu;
3155                 dev->max_mtu = mtu;
3156
3157                 /* TODO: size buffers correctly in this case. */
3158                 if (dev->mtu > ETH_DATA_LEN)
3159                         vi->big_packets = true;
3160         }
3161
3162         if (vi->any_header_sg)
3163                 dev->needed_headroom = vi->hdr_len;
3164
3165         /* Enable multiqueue by default */
3166         if (num_online_cpus() >= max_queue_pairs)
3167                 vi->curr_queue_pairs = max_queue_pairs;
3168         else
3169                 vi->curr_queue_pairs = num_online_cpus();
3170         vi->max_queue_pairs = max_queue_pairs;
3171
3172         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3173         err = init_vqs(vi);
3174         if (err)
3175                 goto free;
3176
3177 #ifdef CONFIG_SYSFS
3178         if (vi->mergeable_rx_bufs)
3179                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3180 #endif
3181         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3182         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3183
3184         virtnet_init_settings(dev);
3185
3186         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3187                 vi->failover = net_failover_create(vi->dev);
3188                 if (IS_ERR(vi->failover)) {
3189                         err = PTR_ERR(vi->failover);
3190                         goto free_vqs;
3191                 }
3192         }
3193
3194         err = register_netdev(dev);
3195         if (err) {
3196                 pr_debug("virtio_net: registering device failed\n");
3197                 goto free_failover;
3198         }
3199
3200         virtio_device_ready(vdev);
3201
3202         err = virtnet_cpu_notif_add(vi);
3203         if (err) {
3204                 pr_debug("virtio_net: registering cpu notifier failed\n");
3205                 goto free_unregister_netdev;
3206         }
3207
3208         virtnet_set_queues(vi, vi->curr_queue_pairs);
3209
3210         /* Assume link up if device can't report link status,
3211            otherwise get link status from config. */
3212         netif_carrier_off(dev);
3213         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3214                 schedule_work(&vi->config_work);
3215         } else {
3216                 vi->status = VIRTIO_NET_S_LINK_UP;
3217                 virtnet_update_settings(vi);
3218                 netif_carrier_on(dev);
3219         }
3220
3221         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3222                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3223                         set_bit(guest_offloads[i], &vi->guest_offloads);
3224         vi->guest_offloads_capable = vi->guest_offloads;
3225
3226         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3227                  dev->name, max_queue_pairs);
3228
3229         return 0;
3230
3231 free_unregister_netdev:
3232         vi->vdev->config->reset(vdev);
3233
3234         unregister_netdev(dev);
3235 free_failover:
3236         net_failover_destroy(vi->failover);
3237 free_vqs:
3238         cancel_delayed_work_sync(&vi->refill);
3239         free_receive_page_frags(vi);
3240         virtnet_del_vqs(vi);
3241 free:
3242         free_netdev(dev);
3243         return err;
3244 }
3245
3246 static void remove_vq_common(struct virtnet_info *vi)
3247 {
3248         vi->vdev->config->reset(vi->vdev);
3249
3250         /* Free unused buffers in both send and recv, if any. */
3251         free_unused_bufs(vi);
3252
3253         free_receive_bufs(vi);
3254
3255         free_receive_page_frags(vi);
3256
3257         virtnet_del_vqs(vi);
3258 }
3259
3260 static void virtnet_remove(struct virtio_device *vdev)
3261 {
3262         struct virtnet_info *vi = vdev->priv;
3263
3264         virtnet_cpu_notif_remove(vi);
3265
3266         /* Make sure no work handler is accessing the device. */
3267         flush_work(&vi->config_work);
3268
3269         unregister_netdev(vi->dev);
3270
3271         net_failover_destroy(vi->failover);
3272
3273         remove_vq_common(vi);
3274
3275         free_netdev(vi->dev);
3276 }
3277
3278 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3279 {
3280         struct virtnet_info *vi = vdev->priv;
3281
3282         virtnet_cpu_notif_remove(vi);
3283         virtnet_freeze_down(vdev);
3284         remove_vq_common(vi);
3285
3286         return 0;
3287 }
3288
3289 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3290 {
3291         struct virtnet_info *vi = vdev->priv;
3292         int err;
3293
3294         err = virtnet_restore_up(vdev);
3295         if (err)
3296                 return err;
3297         virtnet_set_queues(vi, vi->curr_queue_pairs);
3298
3299         err = virtnet_cpu_notif_add(vi);
3300         if (err)
3301                 return err;
3302
3303         return 0;
3304 }
3305
3306 static struct virtio_device_id id_table[] = {
3307         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3308         { 0 },
3309 };
3310
3311 #define VIRTNET_FEATURES \
3312         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3313         VIRTIO_NET_F_MAC, \
3314         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3315         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3316         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3317         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3318         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3319         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3320         VIRTIO_NET_F_CTRL_MAC_ADDR, \
3321         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3322         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3323
3324 static unsigned int features[] = {
3325         VIRTNET_FEATURES,
3326 };
3327
3328 static unsigned int features_legacy[] = {
3329         VIRTNET_FEATURES,
3330         VIRTIO_NET_F_GSO,
3331         VIRTIO_F_ANY_LAYOUT,
3332 };
3333
3334 static struct virtio_driver virtio_net_driver = {
3335         .feature_table = features,
3336         .feature_table_size = ARRAY_SIZE(features),
3337         .feature_table_legacy = features_legacy,
3338         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3339         .driver.name =  KBUILD_MODNAME,
3340         .driver.owner = THIS_MODULE,
3341         .id_table =     id_table,
3342         .validate =     virtnet_validate,
3343         .probe =        virtnet_probe,
3344         .remove =       virtnet_remove,
3345         .config_changed = virtnet_config_changed,
3346 #ifdef CONFIG_PM_SLEEP
3347         .freeze =       virtnet_freeze,
3348         .restore =      virtnet_restore,
3349 #endif
3350 };
3351
3352 static __init int virtio_net_driver_init(void)
3353 {
3354         int ret;
3355
3356         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3357                                       virtnet_cpu_online,
3358                                       virtnet_cpu_down_prep);
3359         if (ret < 0)
3360                 goto out;
3361         virtionet_online = ret;
3362         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3363                                       NULL, virtnet_cpu_dead);
3364         if (ret)
3365                 goto err_dead;
3366
3367         ret = register_virtio_driver(&virtio_net_driver);
3368         if (ret)
3369                 goto err_virtio;
3370         return 0;
3371 err_virtio:
3372         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3373 err_dead:
3374         cpuhp_remove_multi_state(virtionet_online);
3375 out:
3376         return ret;
3377 }
3378 module_init(virtio_net_driver_init);
3379
3380 static __exit void virtio_net_driver_exit(void)
3381 {
3382         unregister_virtio_driver(&virtio_net_driver);
3383         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3384         cpuhp_remove_multi_state(virtionet_online);
3385 }
3386 module_exit(virtio_net_driver_exit);
3387
3388 MODULE_DEVICE_TABLE(virtio, id_table);
3389 MODULE_DESCRIPTION("Virtio network driver");
3390 MODULE_LICENSE("GPL");