drivers/net/virtio_net.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* A network driver using virtio.
   3  *
   4  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   5  */
   6 //#define DEBUG
   7 #include <linux/netdevice.h>
   8 #include <linux/etherdevice.h>
   9 #include <linux/ethtool.h>
  10 #include <linux/module.h>
  11 #include <linux/virtio.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/bpf.h>
  14 #include <linux/bpf_trace.h>
  15 #include <linux/scatterlist.h>
  16 #include <linux/if_vlan.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/average.h>
  20 #include <linux/filter.h>
  21 #include <linux/kernel.h>
  22 #include <net/route.h>
  23 #include <net/xdp.h>
  24 #include <net/net_failover.h>
  25
  26 static int napi_weight = NAPI_POLL_WEIGHT;
  27 module_param(napi_weight, int, 0444);
  28
  29 static bool csum = true, gso = true, napi_tx = true;
  30 module_param(csum, bool, 0444);
  31 module_param(gso, bool, 0444);
  32 module_param(napi_tx, bool, 0644);
  33
  34 /* FIXME: MTU in config. */
  35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  36 #define GOOD_COPY_LEN   128
  37
  38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  39
  40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  41 #define VIRTIO_XDP_HEADROOM 256
  42
  43 /* Separating two types of XDP xmit */
  44 #define VIRTIO_XDP_TX           BIT(0)
  45 #define VIRTIO_XDP_REDIR        BIT(1)
  46
  47 #define VIRTIO_XDP_FLAG BIT(0)
  48
  49 /* RX packet size EWMA. The average packet size is used to determine the packet
  50  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  51  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  52  * term, transient changes in packet size.
  53  */
  54 DECLARE_EWMA(pkt_len, 0, 64)
  55
  56 #define VIRTNET_DRIVER_VERSION "1.0.0"
  57
  58 static const unsigned long guest_offloads[] = {
  59         VIRTIO_NET_F_GUEST_TSO4,
  60         VIRTIO_NET_F_GUEST_TSO6,
  61         VIRTIO_NET_F_GUEST_ECN,
  62         VIRTIO_NET_F_GUEST_UFO,
  63         VIRTIO_NET_F_GUEST_CSUM
  64 };
  65
  66 #define GUEST_OFFLOAD_LRO_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
  67                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
  68                                 (1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
  69                                 (1ULL << VIRTIO_NET_F_GUEST_UFO))
  70
  71 struct virtnet_stat_desc {
  72         char desc[ETH_GSTRING_LEN];
  73         size_t offset;
  74 };
  75
  76 struct virtnet_sq_stats {
  77         struct u64_stats_sync syncp;
  78         u64 packets;
  79         u64 bytes;
  80         u64 xdp_tx;
  81         u64 xdp_tx_drops;
  82         u64 kicks;
  83 };
  84
  85 struct virtnet_rq_stats {
  86         struct u64_stats_sync syncp;
  87         u64 packets;
  88         u64 bytes;
  89         u64 drops;
  90         u64 xdp_packets;
  91         u64 xdp_tx;
  92         u64 xdp_redirects;
  93         u64 xdp_drops;
  94         u64 kicks;
  95 };
  96
  97 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
  98 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
  99
 100 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
 101         { "packets",            VIRTNET_SQ_STAT(packets) },
 102         { "bytes",              VIRTNET_SQ_STAT(bytes) },
 103         { "xdp_tx",             VIRTNET_SQ_STAT(xdp_tx) },
 104         { "xdp_tx_drops",       VIRTNET_SQ_STAT(xdp_tx_drops) },
 105         { "kicks",              VIRTNET_SQ_STAT(kicks) },
 106 };
 107
 108 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
 109         { "packets",            VIRTNET_RQ_STAT(packets) },
 110         { "bytes",              VIRTNET_RQ_STAT(bytes) },
 111         { "drops",              VIRTNET_RQ_STAT(drops) },
 112         { "xdp_packets",        VIRTNET_RQ_STAT(xdp_packets) },
 113         { "xdp_tx",             VIRTNET_RQ_STAT(xdp_tx) },
 114         { "xdp_redirects",      VIRTNET_RQ_STAT(xdp_redirects) },
 115         { "xdp_drops",          VIRTNET_RQ_STAT(xdp_drops) },
 116         { "kicks",              VIRTNET_RQ_STAT(kicks) },
 117 };
 118
 119 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 120 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 121
 122 /* Internal representation of a send virtqueue */
 123 struct send_queue {
 124         /* Virtqueue associated with this send _queue */
 125         struct virtqueue *vq;
 126
 127         /* TX: fragments + linear part + virtio header */
 128         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 129
 130         /* Name of the send queue: output.$index */
 131         char name[40];
 132
 133         struct virtnet_sq_stats stats;
 134
 135         struct napi_struct napi;
 136 };
 137
 138 /* Internal representation of a receive virtqueue */
 139 struct receive_queue {
 140         /* Virtqueue associated with this receive_queue */
 141         struct virtqueue *vq;
 142
 143         struct napi_struct napi;
 144
 145         struct bpf_prog __rcu *xdp_prog;
 146
 147         struct virtnet_rq_stats stats;
 148
 149         /* Chain pages by the private ptr. */
 150         struct page *pages;
 151
 152         /* Average packet length for mergeable receive buffers. */
 153         struct ewma_pkt_len mrg_avg_pkt_len;
 154
 155         /* Page frag for packet buffer allocation. */
 156         struct page_frag alloc_frag;
 157
 158         /* RX: fragments + linear part + virtio header */
 159         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 160
 161         /* Min single buffer size for mergeable buffers case. */
 162         unsigned int min_buf_len;
 163
 164         /* Name of this receive queue: input.$index */
 165         char name[40];
 166
 167         struct xdp_rxq_info xdp_rxq;
 168 };
 169
 170 /* Control VQ buffers: protected by the rtnl lock */
 171 struct control_buf {
 172         struct virtio_net_ctrl_hdr hdr;
 173         virtio_net_ctrl_ack status;
 174         struct virtio_net_ctrl_mq mq;
 175         u8 promisc;
 176         u8 allmulti;
 177         __virtio16 vid;
 178         __virtio64 offloads;
 179 };
 180
 181 struct virtnet_info {
 182         struct virtio_device *vdev;
 183         struct virtqueue *cvq;
 184         struct net_device *dev;
 185         struct send_queue *sq;
 186         struct receive_queue *rq;
 187         unsigned int status;
 188
 189         /* Max # of queue pairs supported by the device */
 190         u16 max_queue_pairs;
 191
 192         /* # of queue pairs currently used by the driver */
 193         u16 curr_queue_pairs;
 194
 195         /* # of XDP queue pairs currently used by the driver */
 196         u16 xdp_queue_pairs;
 197
 198         /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
 199         bool xdp_enabled;
 200
 201         /* I like... big packets and I cannot lie! */
 202         bool big_packets;
 203
 204         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 205         bool mergeable_rx_bufs;
 206
 207         /* Has control virtqueue */
 208         bool has_cvq;
 209
 210         /* Host can handle any s/g split between our header and packet data */
 211         bool any_header_sg;
 212
 213         /* Packet virtio header size */
 214         u8 hdr_len;
 215
 216         /* Work struct for refilling if we run low on memory. */
 217         struct delayed_work refill;
 218
 219         /* Work struct for config space updates */
 220         struct work_struct config_work;
 221
 222         /* Does the affinity hint is set for virtqueues? */
 223         bool affinity_hint_set;
 224
 225         /* CPU hotplug instances for online & dead */
 226         struct hlist_node node;
 227         struct hlist_node node_dead;
 228
 229         struct control_buf *ctrl;
 230
 231         /* Ethtool settings */
 232         u8 duplex;
 233         u32 speed;
 234
 235         unsigned long guest_offloads;
 236         unsigned long guest_offloads_capable;
 237
 238         /* failover when STANDBY feature enabled */
 239         struct failover *failover;
 240 };
 241
 242 struct padded_vnet_hdr {
 243         struct virtio_net_hdr_mrg_rxbuf hdr;
 244         /*
 245          * hdr is in a separate sg buffer, and data sg buffer shares same page
 246          * with this header sg. This padding makes next sg 16 byte aligned
 247          * after the header.
 248          */
 249         char padding[4];
 250 };
 251
 252 static bool is_xdp_frame(void *ptr)
 253 {
 254         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 255 }
 256
 257 static void *xdp_to_ptr(struct xdp_frame *ptr)
 258 {
 259         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 260 }
 261
 262 static struct xdp_frame *ptr_to_xdp(void *ptr)
 263 {
 264         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 265 }
 266
 267 /* Converting between virtqueue no. and kernel tx/rx queue no.
 268  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 269  */
 270 static int vq2txq(struct virtqueue *vq)
 271 {
 272         return (vq->index - 1) / 2;
 273 }
 274
 275 static int txq2vq(int txq)
 276 {
 277         return txq * 2 + 1;
 278 }
 279
 280 static int vq2rxq(struct virtqueue *vq)
 281 {
 282         return vq->index / 2;
 283 }
 284
 285 static int rxq2vq(int rxq)
 286 {
 287         return rxq * 2;
 288 }
 289
 290 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 291 {
 292         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 293 }
 294
 295 /*
 296  * private is used to chain pages for big packets, put the whole
 297  * most recent used list in the beginning for reuse
 298  */
 299 static void give_pages(struct receive_queue *rq, struct page *page)
 300 {
 301         struct page *end;
 302
 303         /* Find end of list, sew whole thing into vi->rq.pages. */
 304         for (end = page; end->private; end = (struct page *)end->private);
 305         end->private = (unsigned long)rq->pages;
 306         rq->pages = page;
 307 }
 308
 309 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 310 {
 311         struct page *p = rq->pages;
 312
 313         if (p) {
 314                 rq->pages = (struct page *)p->private;
 315                 /* clear private here, it is used to chain pages */
 316                 p->private = 0;
 317         } else
 318                 p = alloc_page(gfp_mask);
 319         return p;
 320 }
 321
 322 static void virtqueue_napi_schedule(struct napi_struct *napi,
 323                                     struct virtqueue *vq)
 324 {
 325         if (napi_schedule_prep(napi)) {
 326                 virtqueue_disable_cb(vq);
 327                 __napi_schedule(napi);
 328         }
 329 }
 330
 331 static void virtqueue_napi_complete(struct napi_struct *napi,
 332                                     struct virtqueue *vq, int processed)
 333 {
 334         int opaque;
 335
 336         opaque = virtqueue_enable_cb_prepare(vq);
 337         if (napi_complete_done(napi, processed)) {
 338                 if (unlikely(virtqueue_poll(vq, opaque)))
 339                         virtqueue_napi_schedule(napi, vq);
 340         } else {
 341                 virtqueue_disable_cb(vq);
 342         }
 343 }
 344
 345 static void skb_xmit_done(struct virtqueue *vq)
 346 {
 347         struct virtnet_info *vi = vq->vdev->priv;
 348         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 349
 350         /* Suppress further interrupts. */
 351         virtqueue_disable_cb(vq);
 352
 353         if (napi->weight)
 354                 virtqueue_napi_schedule(napi, vq);
 355         else
 356                 /* We were probably waiting for more output buffers. */
 357                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 358 }
 359
 360 #define MRG_CTX_HEADER_SHIFT 22
 361 static void *mergeable_len_to_ctx(unsigned int truesize,
 362                                   unsigned int headroom)
 363 {
 364         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 365 }
 366
 367 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 368 {
 369         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 370 }
 371
 372 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 373 {
 374         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 375 }
 376
 377 /* Called from bottom half context */
 378 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 379                                    struct receive_queue *rq,
 380                                    struct page *page, unsigned int offset,
 381                                    unsigned int len, unsigned int truesize,
 382                                    bool hdr_valid, unsigned int metasize,
 383                                    unsigned int headroom)
 384 {
 385         struct sk_buff *skb;
 386         struct virtio_net_hdr_mrg_rxbuf *hdr;
 387         unsigned int copy, hdr_len, hdr_padded_len;
 388         struct page *page_to_free = NULL;
 389         int tailroom, shinfo_size;
 390         char *p, *hdr_p, *buf;
 391
 392         p = page_address(page) + offset;
 393         hdr_p = p;
 394
 395         hdr_len = vi->hdr_len;
 396         if (vi->mergeable_rx_bufs)
 397                 hdr_padded_len = sizeof(*hdr);
 398         else
 399                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 400
 401         /* If headroom is not 0, there is an offset between the beginning of the
 402          * data and the allocated space, otherwise the data and the allocated
 403          * space are aligned.
 404          *
 405          * Buffers with headroom use PAGE_SIZE as alloc size, see
 406          * add_recvbuf_mergeable() + get_mergeable_buf_len()
 407          */
 408         truesize = headroom ? PAGE_SIZE : truesize;
 409         tailroom = truesize - len - headroom - (hdr_padded_len - hdr_len);
 410         buf = p - headroom;
 411
 412         len -= hdr_len;
 413         offset += hdr_padded_len;
 414         p += hdr_padded_len;
 415
 416         shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 417
 418         /* copy small packet so we can reuse these pages */
 419         if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
 420                 skb = build_skb(buf, truesize);
 421                 if (unlikely(!skb))
 422                         return NULL;
 423
 424                 skb_reserve(skb, p - buf);
 425                 skb_put(skb, len);
 426                 goto ok;
 427         }
 428
 429         /* copy small packet so we can reuse these pages for small data */
 430         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 431         if (unlikely(!skb))
 432                 return NULL;
 433
 434         /* Copy all frame if it fits skb->head, otherwise
 435          * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
 436          */
 437         if (len <= skb_tailroom(skb))
 438                 copy = len;
 439         else
 440                 copy = ETH_HLEN + metasize;
 441         skb_put_data(skb, p, copy);
 442
 443         len -= copy;
 444         offset += copy;
 445
 446         if (vi->mergeable_rx_bufs) {
 447                 if (len)
 448                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 449                 else
 450                         page_to_free = page;
 451                 goto ok;
 452         }
 453
 454         /*
 455          * Verify that we can indeed put this data into a skb.
 456          * This is here to handle cases when the device erroneously
 457          * tries to receive more than is possible. This is usually
 458          * the case of a broken device.
 459          */
 460         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 461                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 462                 dev_kfree_skb(skb);
 463                 return NULL;
 464         }
 465         BUG_ON(offset >= PAGE_SIZE);
 466         while (len) {
 467                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 468                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 469                                 frag_size, truesize);
 470                 len -= frag_size;
 471                 page = (struct page *)page->private;
 472                 offset = 0;
 473         }
 474
 475         if (page)
 476                 give_pages(rq, page);
 477
 478 ok:
 479         /* hdr_valid means no XDP, so we can copy the vnet header */
 480         if (hdr_valid) {
 481                 hdr = skb_vnet_hdr(skb);
 482                 memcpy(hdr, hdr_p, hdr_len);
 483         }
 484         if (page_to_free)
 485                 put_page(page_to_free);
 486
 487         if (metasize) {
 488                 __skb_pull(skb, metasize);
 489                 skb_metadata_set(skb, metasize);
 490         }
 491
 492         return skb;
 493 }
 494
 495 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 496                                    struct send_queue *sq,
 497                                    struct xdp_frame *xdpf)
 498 {
 499         struct virtio_net_hdr_mrg_rxbuf *hdr;
 500         int err;
 501
 502         if (unlikely(xdpf->headroom < vi->hdr_len))
 503                 return -EOVERFLOW;
 504
 505         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 506         xdpf->data -= vi->hdr_len;
 507         /* Zero header and leave csum up to XDP layers */
 508         hdr = xdpf->data;
 509         memset(hdr, 0, vi->hdr_len);
 510         xdpf->len   += vi->hdr_len;
 511
 512         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 513
 514         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
 515                                    GFP_ATOMIC);
 516         if (unlikely(err))
 517                 return -ENOSPC; /* Caller handle free/refcnt */
 518
 519         return 0;
 520 }
 521
 522 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
 523  * the current cpu, so it does not need to be locked.
 524  *
 525  * Here we use marco instead of inline functions because we have to deal with
 526  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
 527  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
 528  * functions to perfectly solve these three problems at the same time.
 529  */
 530 #define virtnet_xdp_get_sq(vi) ({                                       \
 531         struct netdev_queue *txq;                                       \
 532         typeof(vi) v = (vi);                                            \
 533         unsigned int qp;                                                \
 534                                                                         \
 535         if (v->curr_queue_pairs > nr_cpu_ids) {                         \
 536                 qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
 537                 qp += smp_processor_id();                               \
 538                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 539                 __netif_tx_acquire(txq);                                \
 540         } else {                                                        \
 541                 qp = smp_processor_id() % v->curr_queue_pairs;          \
 542                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 543                 __netif_tx_lock(txq, raw_smp_processor_id());           \
 544         }                                                               \
 545         v->sq + qp;                                                     \
 546 })
 547
 548 #define virtnet_xdp_put_sq(vi, q) {                                     \
 549         struct netdev_queue *txq;                                       \
 550         typeof(vi) v = (vi);                                            \
 551                                                                         \
 552         txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
 553         if (v->curr_queue_pairs > nr_cpu_ids)                           \
 554                 __netif_tx_release(txq);                                \
 555         else                                                            \
 556                 __netif_tx_unlock(txq);                                 \
 557 }
 558
 559 static int virtnet_xdp_xmit(struct net_device *dev,
 560                             int n, struct xdp_frame **frames, u32 flags)
 561 {
 562         struct virtnet_info *vi = netdev_priv(dev);
 563         struct receive_queue *rq = vi->rq;
 564         struct bpf_prog *xdp_prog;
 565         struct send_queue *sq;
 566         unsigned int len;
 567         int packets = 0;
 568         int bytes = 0;
 569         int nxmit = 0;
 570         int kicks = 0;
 571         void *ptr;
 572         int ret;
 573         int i;
 574
 575         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 576          * indicate XDP resources have been successfully allocated.
 577          */
 578         xdp_prog = rcu_access_pointer(rq->xdp_prog);
 579         if (!xdp_prog)
 580                 return -ENXIO;
 581
 582         sq = virtnet_xdp_get_sq(vi);
 583
 584         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 585                 ret = -EINVAL;
 586                 goto out;
 587         }
 588
 589         /* Free up any pending old buffers before queueing new ones. */
 590         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 591                 if (likely(is_xdp_frame(ptr))) {
 592                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 593
 594                         bytes += frame->len;
 595                         xdp_return_frame(frame);
 596                 } else {
 597                         struct sk_buff *skb = ptr;
 598
 599                         bytes += skb->len;
 600                         napi_consume_skb(skb, false);
 601                 }
 602                 packets++;
 603         }
 604
 605         for (i = 0; i < n; i++) {
 606                 struct xdp_frame *xdpf = frames[i];
 607
 608                 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
 609                         break;
 610                 nxmit++;
 611         }
 612         ret = nxmit;
 613
 614         if (flags & XDP_XMIT_FLUSH) {
 615                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
 616                         kicks = 1;
 617         }
 618 out:
 619         u64_stats_update_begin(&sq->stats.syncp);
 620         sq->stats.bytes += bytes;
 621         sq->stats.packets += packets;
 622         sq->stats.xdp_tx += n;
 623         sq->stats.xdp_tx_drops += n - nxmit;
 624         sq->stats.kicks += kicks;
 625         u64_stats_update_end(&sq->stats.syncp);
 626
 627         virtnet_xdp_put_sq(vi, sq);
 628         return ret;
 629 }
 630
 631 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 632 {
 633         return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
 634 }
 635
 636 /* We copy the packet for XDP in the following cases:
 637  *
 638  * 1) Packet is scattered across multiple rx buffers.
 639  * 2) Headroom space is insufficient.
 640  *
 641  * This is inefficient but it's a temporary condition that
 642  * we hit right after XDP is enabled and until queue is refilled
 643  * with large buffers with sufficient headroom - so it should affect
 644  * at most queue size packets.
 645  * Afterwards, the conditions to enable
 646  * XDP should preclude the underlying device from sending packets
 647  * across multiple buffers (num_buf > 1), and we make sure buffers
 648  * have enough headroom.
 649  */
 650 static struct page *xdp_linearize_page(struct receive_queue *rq,
 651                                        u16 *num_buf,
 652                                        struct page *p,
 653                                        int offset,
 654                                        int page_off,
 655                                        unsigned int *len)
 656 {
 657         struct page *page = alloc_page(GFP_ATOMIC);
 658
 659         if (!page)
 660                 return NULL;
 661
 662         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 663         page_off += *len;
 664
 665         while (--*num_buf) {
 666                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 667                 unsigned int buflen;
 668                 void *buf;
 669                 int off;
 670
 671                 buf = virtqueue_get_buf(rq->vq, &buflen);
 672                 if (unlikely(!buf))
 673                         goto err_buf;
 674
 675                 p = virt_to_head_page(buf);
 676                 off = buf - page_address(p);
 677
 678                 /* guard against a misconfigured or uncooperative backend that
 679                  * is sending packet larger than the MTU.
 680                  */
 681                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 682                         put_page(p);
 683                         goto err_buf;
 684                 }
 685
 686                 memcpy(page_address(page) + page_off,
 687                        page_address(p) + off, buflen);
 688                 page_off += buflen;
 689                 put_page(p);
 690         }
 691
 692         /* Headroom does not contribute to packet length */
 693         *len = page_off - VIRTIO_XDP_HEADROOM;
 694         return page;
 695 err_buf:
 696         __free_pages(page, 0);
 697         return NULL;
 698 }
 699
 700 static struct sk_buff *receive_small(struct net_device *dev,
 701                                      struct virtnet_info *vi,
 702                                      struct receive_queue *rq,
 703                                      void *buf, void *ctx,
 704                                      unsigned int len,
 705                                      unsigned int *xdp_xmit,
 706                                      struct virtnet_rq_stats *stats)
 707 {
 708         struct sk_buff *skb;
 709         struct bpf_prog *xdp_prog;
 710         unsigned int xdp_headroom = (unsigned long)ctx;
 711         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 712         unsigned int headroom = vi->hdr_len + header_offset;
 713         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 714                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 715         struct page *page = virt_to_head_page(buf);
 716         unsigned int delta = 0;
 717         struct page *xdp_page;
 718         int err;
 719         unsigned int metasize = 0;
 720
 721         len -= vi->hdr_len;
 722         stats->bytes += len;
 723
 724         rcu_read_lock();
 725         xdp_prog = rcu_dereference(rq->xdp_prog);
 726         if (xdp_prog) {
 727                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 728                 struct xdp_frame *xdpf;
 729                 struct xdp_buff xdp;
 730                 void *orig_data;
 731                 u32 act;
 732
 733                 if (unlikely(hdr->hdr.gso_type))
 734                         goto err_xdp;
 735
 736                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 737                         int offset = buf - page_address(page) + header_offset;
 738                         unsigned int tlen = len + vi->hdr_len;
 739                         u16 num_buf = 1;
 740
 741                         xdp_headroom = virtnet_get_headroom(vi);
 742                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 743                         headroom = vi->hdr_len + header_offset;
 744                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 745                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 746                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 747                                                       offset, header_offset,
 748                                                       &tlen);
 749                         if (!xdp_page)
 750                                 goto err_xdp;
 751
 752                         buf = page_address(xdp_page);
 753                         put_page(page);
 754                         page = xdp_page;
 755                 }
 756
 757                 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
 758                 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
 759                                  xdp_headroom, len, true);
 760                 orig_data = xdp.data;
 761                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 762                 stats->xdp_packets++;
 763
 764                 switch (act) {
 765                 case XDP_PASS:
 766                         /* Recalculate length in case bpf program changed it */
 767                         delta = orig_data - xdp.data;
 768                         len = xdp.data_end - xdp.data;
 769                         metasize = xdp.data - xdp.data_meta;
 770                         break;
 771                 case XDP_TX:
 772                         stats->xdp_tx++;
 773                         xdpf = xdp_convert_buff_to_frame(&xdp);
 774                         if (unlikely(!xdpf))
 775                                 goto err_xdp;
 776                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 777                         if (unlikely(!err)) {
 778                                 xdp_return_frame_rx_napi(xdpf);
 779                         } else if (unlikely(err < 0)) {
 780                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 781                                 goto err_xdp;
 782                         }
 783                         *xdp_xmit |= VIRTIO_XDP_TX;
 784                         rcu_read_unlock();
 785                         goto xdp_xmit;
 786                 case XDP_REDIRECT:
 787                         stats->xdp_redirects++;
 788                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 789                         if (err)
 790                                 goto err_xdp;
 791                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 792                         rcu_read_unlock();
 793                         goto xdp_xmit;
 794                 default:
 795                         bpf_warn_invalid_xdp_action(act);
 796                         fallthrough;
 797                 case XDP_ABORTED:
 798                         trace_xdp_exception(vi->dev, xdp_prog, act);
 799                         goto err_xdp;
 800                 case XDP_DROP:
 801                         goto err_xdp;
 802                 }
 803         }
 804         rcu_read_unlock();
 805
 806         skb = build_skb(buf, buflen);
 807         if (!skb) {
 808                 put_page(page);
 809                 goto err;
 810         }
 811         skb_reserve(skb, headroom - delta);
 812         skb_put(skb, len);
 813         if (!xdp_prog) {
 814                 buf += header_offset;
 815                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 816         } /* keep zeroed vnet hdr since XDP is loaded */
 817
 818         if (metasize)
 819                 skb_metadata_set(skb, metasize);
 820
 821 err:
 822         return skb;
 823
 824 err_xdp:
 825         rcu_read_unlock();
 826         stats->xdp_drops++;
 827         stats->drops++;
 828         put_page(page);
 829 xdp_xmit:
 830         return NULL;
 831 }
 832
 833 static struct sk_buff *receive_big(struct net_device *dev,
 834                                    struct virtnet_info *vi,
 835                                    struct receive_queue *rq,
 836                                    void *buf,
 837                                    unsigned int len,
 838                                    struct virtnet_rq_stats *stats)
 839 {
 840         struct page *page = buf;
 841         struct sk_buff *skb =
 842                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
 843
 844         stats->bytes += len - vi->hdr_len;
 845         if (unlikely(!skb))
 846                 goto err;
 847
 848         return skb;
 849
 850 err:
 851         stats->drops++;
 852         give_pages(rq, page);
 853         return NULL;
 854 }
 855
 856 static struct sk_buff *receive_mergeable(struct net_device *dev,
 857                                          struct virtnet_info *vi,
 858                                          struct receive_queue *rq,
 859                                          void *buf,
 860                                          void *ctx,
 861                                          unsigned int len,
 862                                          unsigned int *xdp_xmit,
 863                                          struct virtnet_rq_stats *stats)
 864 {
 865         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 866         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 867         struct page *page = virt_to_head_page(buf);
 868         int offset = buf - page_address(page);
 869         struct sk_buff *head_skb, *curr_skb;
 870         struct bpf_prog *xdp_prog;
 871         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
 872         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 873         unsigned int metasize = 0;
 874         unsigned int frame_sz;
 875         int err;
 876
 877         head_skb = NULL;
 878         stats->bytes += len - vi->hdr_len;
 879
 880         rcu_read_lock();
 881         xdp_prog = rcu_dereference(rq->xdp_prog);
 882         if (xdp_prog) {
 883                 struct xdp_frame *xdpf;
 884                 struct page *xdp_page;
 885                 struct xdp_buff xdp;
 886                 void *data;
 887                 u32 act;
 888
 889                 /* Transient failure which in theory could occur if
 890                  * in-flight packets from before XDP was enabled reach
 891                  * the receive path after XDP is loaded.
 892                  */
 893                 if (unlikely(hdr->hdr.gso_type))
 894                         goto err_xdp;
 895
 896                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 897                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 898                  */
 899                 frame_sz = headroom ? PAGE_SIZE : truesize;
 900
 901                 /* This happens when rx buffer size is underestimated
 902                  * or headroom is not enough because of the buffer
 903                  * was refilled before XDP is set. This should only
 904                  * happen for the first several packets, so we don't
 905                  * care much about its performance.
 906                  */
 907                 if (unlikely(num_buf > 1 ||
 908                              headroom < virtnet_get_headroom(vi))) {
 909                         /* linearize data for XDP */
 910                         xdp_page = xdp_linearize_page(rq, &num_buf,
 911                                                       page, offset,
 912                                                       VIRTIO_XDP_HEADROOM,
 913                                                       &len);
 914                         frame_sz = PAGE_SIZE;
 915
 916                         if (!xdp_page)
 917                                 goto err_xdp;
 918                         offset = VIRTIO_XDP_HEADROOM;
 919                 } else {
 920                         xdp_page = page;
 921                 }
 922
 923                 /* Allow consuming headroom but reserve enough space to push
 924                  * the descriptor on if we get an XDP_TX return code.
 925                  */
 926                 data = page_address(xdp_page) + offset;
 927                 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
 928                 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
 929                                  VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
 930
 931                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 932                 stats->xdp_packets++;
 933
 934                 switch (act) {
 935                 case XDP_PASS:
 936                         metasize = xdp.data - xdp.data_meta;
 937
 938                         /* recalculate offset to account for any header
 939                          * adjustments and minus the metasize to copy the
 940                          * metadata in page_to_skb(). Note other cases do not
 941                          * build an skb and avoid using offset
 942                          */
 943                         offset = xdp.data - page_address(xdp_page) -
 944                                  vi->hdr_len - metasize;
 945
 946                         /* recalculate len if xdp.data, xdp.data_end or
 947                          * xdp.data_meta were adjusted
 948                          */
 949                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 950                         /* We can only create skb based on xdp_page. */
 951                         if (unlikely(xdp_page != page)) {
 952                                 rcu_read_unlock();
 953                                 put_page(page);
 954                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
 955                                                        len, PAGE_SIZE, false,
 956                                                        metasize,
 957                                                        VIRTIO_XDP_HEADROOM);
 958                                 return head_skb;
 959                         }
 960                         break;
 961                 case XDP_TX:
 962                         stats->xdp_tx++;
 963                         xdpf = xdp_convert_buff_to_frame(&xdp);
 964                         if (unlikely(!xdpf))
 965                                 goto err_xdp;
 966                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 967                         if (unlikely(!err)) {
 968                                 xdp_return_frame_rx_napi(xdpf);
 969                         } else if (unlikely(err < 0)) {
 970                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 971                                 if (unlikely(xdp_page != page))
 972                                         put_page(xdp_page);
 973                                 goto err_xdp;
 974                         }
 975                         *xdp_xmit |= VIRTIO_XDP_TX;
 976                         if (unlikely(xdp_page != page))
 977                                 put_page(page);
 978                         rcu_read_unlock();
 979                         goto xdp_xmit;
 980                 case XDP_REDIRECT:
 981                         stats->xdp_redirects++;
 982                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 983                         if (err) {
 984                                 if (unlikely(xdp_page != page))
 985                                         put_page(xdp_page);
 986                                 goto err_xdp;
 987                         }
 988                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 989                         if (unlikely(xdp_page != page))
 990                                 put_page(page);
 991                         rcu_read_unlock();
 992                         goto xdp_xmit;
 993                 default:
 994                         bpf_warn_invalid_xdp_action(act);
 995                         fallthrough;
 996                 case XDP_ABORTED:
 997                         trace_xdp_exception(vi->dev, xdp_prog, act);
 998                         fallthrough;
 999                 case XDP_DROP:
1000                         if (unlikely(xdp_page != page))
1001                                 __free_pages(xdp_page, 0);
1002                         goto err_xdp;
1003                 }
1004         }
1005         rcu_read_unlock();
1006
1007         if (unlikely(len > truesize)) {
1008                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1009                          dev->name, len, (unsigned long)ctx);
1010                 dev->stats.rx_length_errors++;
1011                 goto err_skb;
1012         }
1013
1014         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1015                                metasize, headroom);
1016         curr_skb = head_skb;
1017
1018         if (unlikely(!curr_skb))
1019                 goto err_skb;
1020         while (--num_buf) {
1021                 int num_skb_frags;
1022
1023                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1024                 if (unlikely(!buf)) {
1025                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1026                                  dev->name, num_buf,
1027                                  virtio16_to_cpu(vi->vdev,
1028                                                  hdr->num_buffers));
1029                         dev->stats.rx_length_errors++;
1030                         goto err_buf;
1031                 }
1032
1033                 stats->bytes += len;
1034                 page = virt_to_head_page(buf);
1035
1036                 truesize = mergeable_ctx_to_truesize(ctx);
1037                 if (unlikely(len > truesize)) {
1038                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1039                                  dev->name, len, (unsigned long)ctx);
1040                         dev->stats.rx_length_errors++;
1041                         goto err_skb;
1042                 }
1043
1044                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1045                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1046                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1047
1048                         if (unlikely(!nskb))
1049                                 goto err_skb;
1050                         if (curr_skb == head_skb)
1051                                 skb_shinfo(curr_skb)->frag_list = nskb;
1052                         else
1053                                 curr_skb->next = nskb;
1054                         curr_skb = nskb;
1055                         head_skb->truesize += nskb->truesize;
1056                         num_skb_frags = 0;
1057                 }
1058                 if (curr_skb != head_skb) {
1059                         head_skb->data_len += len;
1060                         head_skb->len += len;
1061                         head_skb->truesize += truesize;
1062                 }
1063                 offset = buf - page_address(page);
1064                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1065                         put_page(page);
1066                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1067                                              len, truesize);
1068                 } else {
1069                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1070                                         offset, len, truesize);
1071                 }
1072         }
1073
1074         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1075         return head_skb;
1076
1077 err_xdp:
1078         rcu_read_unlock();
1079         stats->xdp_drops++;
1080 err_skb:
1081         put_page(page);
1082         while (num_buf-- > 1) {
1083                 buf = virtqueue_get_buf(rq->vq, &len);
1084                 if (unlikely(!buf)) {
1085                         pr_debug("%s: rx error: %d buffers missing\n",
1086                                  dev->name, num_buf);
1087                         dev->stats.rx_length_errors++;
1088                         break;
1089                 }
1090                 stats->bytes += len;
1091                 page = virt_to_head_page(buf);
1092                 put_page(page);
1093         }
1094 err_buf:
1095         stats->drops++;
1096         dev_kfree_skb(head_skb);
1097 xdp_xmit:
1098         return NULL;
1099 }
1100
1101 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1102                         void *buf, unsigned int len, void **ctx,
1103                         unsigned int *xdp_xmit,
1104                         struct virtnet_rq_stats *stats)
1105 {
1106         struct net_device *dev = vi->dev;
1107         struct sk_buff *skb;
1108         struct virtio_net_hdr_mrg_rxbuf *hdr;
1109
1110         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1111                 pr_debug("%s: short packet %i\n", dev->name, len);
1112                 dev->stats.rx_length_errors++;
1113                 if (vi->mergeable_rx_bufs) {
1114                         put_page(virt_to_head_page(buf));
1115                 } else if (vi->big_packets) {
1116                         give_pages(rq, buf);
1117                 } else {
1118                         put_page(virt_to_head_page(buf));
1119                 }
1120                 return;
1121         }
1122
1123         if (vi->mergeable_rx_bufs)
1124                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1125                                         stats);
1126         else if (vi->big_packets)
1127                 skb = receive_big(dev, vi, rq, buf, len, stats);
1128         else
1129                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1130
1131         if (unlikely(!skb))
1132                 return;
1133
1134         hdr = skb_vnet_hdr(skb);
1135
1136         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1137                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1138
1139         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1140                                   virtio_is_little_endian(vi->vdev))) {
1141                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1142                                      dev->name, hdr->hdr.gso_type,
1143                                      hdr->hdr.gso_size);
1144                 goto frame_err;
1145         }
1146
1147         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1148         skb->protocol = eth_type_trans(skb, dev);
1149         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1150                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1151
1152         napi_gro_receive(&rq->napi, skb);
1153         return;
1154
1155 frame_err:
1156         dev->stats.rx_frame_errors++;
1157         dev_kfree_skb(skb);
1158 }
1159
1160 /* Unlike mergeable buffers, all buffers are allocated to the
1161  * same size, except for the headroom. For this reason we do
1162  * not need to use  mergeable_len_to_ctx here - it is enough
1163  * to store the headroom as the context ignoring the truesize.
1164  */
1165 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1166                              gfp_t gfp)
1167 {
1168         struct page_frag *alloc_frag = &rq->alloc_frag;
1169         char *buf;
1170         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1171         void *ctx = (void *)(unsigned long)xdp_headroom;
1172         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1173         int err;
1174
1175         len = SKB_DATA_ALIGN(len) +
1176               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1177         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1178                 return -ENOMEM;
1179
1180         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1181         get_page(alloc_frag->page);
1182         alloc_frag->offset += len;
1183         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1184                     vi->hdr_len + GOOD_PACKET_LEN);
1185         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1186         if (err < 0)
1187                 put_page(virt_to_head_page(buf));
1188         return err;
1189 }
1190
1191 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1192                            gfp_t gfp)
1193 {
1194         struct page *first, *list = NULL;
1195         char *p;
1196         int i, err, offset;
1197
1198         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1199
1200         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1201         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1202                 first = get_a_page(rq, gfp);
1203                 if (!first) {
1204                         if (list)
1205                                 give_pages(rq, list);
1206                         return -ENOMEM;
1207                 }
1208                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1209
1210                 /* chain new page in list head to match sg */
1211                 first->private = (unsigned long)list;
1212                 list = first;
1213         }
1214
1215         first = get_a_page(rq, gfp);
1216         if (!first) {
1217                 give_pages(rq, list);
1218                 return -ENOMEM;
1219         }
1220         p = page_address(first);
1221
1222         /* rq->sg[0], rq->sg[1] share the same page */
1223         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1224         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1225
1226         /* rq->sg[1] for data packet, from offset */
1227         offset = sizeof(struct padded_vnet_hdr);
1228         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1229
1230         /* chain first in list head */
1231         first->private = (unsigned long)list;
1232         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1233                                   first, gfp);
1234         if (err < 0)
1235                 give_pages(rq, first);
1236
1237         return err;
1238 }
1239
1240 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1241                                           struct ewma_pkt_len *avg_pkt_len,
1242                                           unsigned int room)
1243 {
1244         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1245         unsigned int len;
1246
1247         if (room)
1248                 return PAGE_SIZE - room;
1249
1250         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1251                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1252
1253         return ALIGN(len, L1_CACHE_BYTES);
1254 }
1255
1256 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1257                                  struct receive_queue *rq, gfp_t gfp)
1258 {
1259         struct page_frag *alloc_frag = &rq->alloc_frag;
1260         unsigned int headroom = virtnet_get_headroom(vi);
1261         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1262         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1263         char *buf;
1264         void *ctx;
1265         int err;
1266         unsigned int len, hole;
1267
1268         /* Extra tailroom is needed to satisfy XDP's assumption. This
1269          * means rx frags coalescing won't work, but consider we've
1270          * disabled GSO for XDP, it won't be a big issue.
1271          */
1272         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1273         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1274                 return -ENOMEM;
1275
1276         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1277         buf += headroom; /* advance address leaving hole at front of pkt */
1278         get_page(alloc_frag->page);
1279         alloc_frag->offset += len + room;
1280         hole = alloc_frag->size - alloc_frag->offset;
1281         if (hole < len + room) {
1282                 /* To avoid internal fragmentation, if there is very likely not
1283                  * enough space for another buffer, add the remaining space to
1284                  * the current buffer.
1285                  */
1286                 len += hole;
1287                 alloc_frag->offset += hole;
1288         }
1289
1290         sg_init_one(rq->sg, buf, len);
1291         ctx = mergeable_len_to_ctx(len, headroom);
1292         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1293         if (err < 0)
1294                 put_page(virt_to_head_page(buf));
1295
1296         return err;
1297 }
1298
1299 /*
1300  * Returns false if we couldn't fill entirely (OOM).
1301  *
1302  * Normally run in the receive path, but can also be run from ndo_open
1303  * before we're receiving packets, or from refill_work which is
1304  * careful to disable receiving (using napi_disable).
1305  */
1306 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1307                           gfp_t gfp)
1308 {
1309         int err;
1310         bool oom;
1311
1312         do {
1313                 if (vi->mergeable_rx_bufs)
1314                         err = add_recvbuf_mergeable(vi, rq, gfp);
1315                 else if (vi->big_packets)
1316                         err = add_recvbuf_big(vi, rq, gfp);
1317                 else
1318                         err = add_recvbuf_small(vi, rq, gfp);
1319
1320                 oom = err == -ENOMEM;
1321                 if (err)
1322                         break;
1323         } while (rq->vq->num_free);
1324         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1325                 unsigned long flags;
1326
1327                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1328                 rq->stats.kicks++;
1329                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1330         }
1331
1332         return !oom;
1333 }
1334
1335 static void skb_recv_done(struct virtqueue *rvq)
1336 {
1337         struct virtnet_info *vi = rvq->vdev->priv;
1338         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1339
1340         virtqueue_napi_schedule(&rq->napi, rvq);
1341 }
1342
1343 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1344 {
1345         napi_enable(napi);
1346
1347         /* If all buffers were filled by other side before we napi_enabled, we
1348          * won't get another interrupt, so process any outstanding packets now.
1349          * Call local_bh_enable after to trigger softIRQ processing.
1350          */
1351         local_bh_disable();
1352         virtqueue_napi_schedule(napi, vq);
1353         local_bh_enable();
1354 }
1355
1356 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1357                                    struct virtqueue *vq,
1358                                    struct napi_struct *napi)
1359 {
1360         if (!napi->weight)
1361                 return;
1362
1363         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1364          * enable the feature if this is likely affine with the transmit path.
1365          */
1366         if (!vi->affinity_hint_set) {
1367                 napi->weight = 0;
1368                 return;
1369         }
1370
1371         return virtnet_napi_enable(vq, napi);
1372 }
1373
1374 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1375 {
1376         if (napi->weight)
1377                 napi_disable(napi);
1378 }
1379
1380 static void refill_work(struct work_struct *work)
1381 {
1382         struct virtnet_info *vi =
1383                 container_of(work, struct virtnet_info, refill.work);
1384         bool still_empty;
1385         int i;
1386
1387         for (i = 0; i < vi->curr_queue_pairs; i++) {
1388                 struct receive_queue *rq = &vi->rq[i];
1389
1390                 napi_disable(&rq->napi);
1391                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1392                 virtnet_napi_enable(rq->vq, &rq->napi);
1393
1394                 /* In theory, this can happen: if we don't get any buffers in
1395                  * we will *never* try to fill again.
1396                  */
1397                 if (still_empty)
1398                         schedule_delayed_work(&vi->refill, HZ/2);
1399         }
1400 }
1401
1402 static int virtnet_receive(struct receive_queue *rq, int budget,
1403                            unsigned int *xdp_xmit)
1404 {
1405         struct virtnet_info *vi = rq->vq->vdev->priv;
1406         struct virtnet_rq_stats stats = {};
1407         unsigned int len;
1408         void *buf;
1409         int i;
1410
1411         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1412                 void *ctx;
1413
1414                 while (stats.packets < budget &&
1415                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1416                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1417                         stats.packets++;
1418                 }
1419         } else {
1420                 while (stats.packets < budget &&
1421                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1422                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1423                         stats.packets++;
1424                 }
1425         }
1426
1427         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1428                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1429                         schedule_delayed_work(&vi->refill, 0);
1430         }
1431
1432         u64_stats_update_begin(&rq->stats.syncp);
1433         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1434                 size_t offset = virtnet_rq_stats_desc[i].offset;
1435                 u64 *item;
1436
1437                 item = (u64 *)((u8 *)&rq->stats + offset);
1438                 *item += *(u64 *)((u8 *)&stats + offset);
1439         }
1440         u64_stats_update_end(&rq->stats.syncp);
1441
1442         return stats.packets;
1443 }
1444
1445 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1446 {
1447         unsigned int len;
1448         unsigned int packets = 0;
1449         unsigned int bytes = 0;
1450         void *ptr;
1451
1452         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1453                 if (likely(!is_xdp_frame(ptr))) {
1454                         struct sk_buff *skb = ptr;
1455
1456                         pr_debug("Sent skb %p\n", skb);
1457
1458                         bytes += skb->len;
1459                         napi_consume_skb(skb, in_napi);
1460                 } else {
1461                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1462
1463                         bytes += frame->len;
1464                         xdp_return_frame(frame);
1465                 }
1466                 packets++;
1467         }
1468
1469         /* Avoid overhead when no packets have been processed
1470          * happens when called speculatively from start_xmit.
1471          */
1472         if (!packets)
1473                 return;
1474
1475         u64_stats_update_begin(&sq->stats.syncp);
1476         sq->stats.bytes += bytes;
1477         sq->stats.packets += packets;
1478         u64_stats_update_end(&sq->stats.syncp);
1479 }
1480
1481 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1482 {
1483         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1484                 return false;
1485         else if (q < vi->curr_queue_pairs)
1486                 return true;
1487         else
1488                 return false;
1489 }
1490
1491 static void virtnet_poll_cleantx(struct receive_queue *rq)
1492 {
1493         struct virtnet_info *vi = rq->vq->vdev->priv;
1494         unsigned int index = vq2rxq(rq->vq);
1495         struct send_queue *sq = &vi->sq[index];
1496         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1497
1498         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1499                 return;
1500
1501         if (__netif_tx_trylock(txq)) {
1502                 free_old_xmit_skbs(sq, true);
1503                 __netif_tx_unlock(txq);
1504         }
1505
1506         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1507                 netif_tx_wake_queue(txq);
1508 }
1509
1510 static int virtnet_poll(struct napi_struct *napi, int budget)
1511 {
1512         struct receive_queue *rq =
1513                 container_of(napi, struct receive_queue, napi);
1514         struct virtnet_info *vi = rq->vq->vdev->priv;
1515         struct send_queue *sq;
1516         unsigned int received;
1517         unsigned int xdp_xmit = 0;
1518
1519         virtnet_poll_cleantx(rq);
1520
1521         received = virtnet_receive(rq, budget, &xdp_xmit);
1522
1523         /* Out of packets? */
1524         if (received < budget)
1525                 virtqueue_napi_complete(napi, rq->vq, received);
1526
1527         if (xdp_xmit & VIRTIO_XDP_REDIR)
1528                 xdp_do_flush();
1529
1530         if (xdp_xmit & VIRTIO_XDP_TX) {
1531                 sq = virtnet_xdp_get_sq(vi);
1532                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1533                         u64_stats_update_begin(&sq->stats.syncp);
1534                         sq->stats.kicks++;
1535                         u64_stats_update_end(&sq->stats.syncp);
1536                 }
1537                 virtnet_xdp_put_sq(vi, sq);
1538         }
1539
1540         return received;
1541 }
1542
1543 static int virtnet_open(struct net_device *dev)
1544 {
1545         struct virtnet_info *vi = netdev_priv(dev);
1546         int i, err;
1547
1548         for (i = 0; i < vi->max_queue_pairs; i++) {
1549                 if (i < vi->curr_queue_pairs)
1550                         /* Make sure we have some buffers: if oom use wq. */
1551                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1552                                 schedule_delayed_work(&vi->refill, 0);
1553
1554                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1555                 if (err < 0)
1556                         return err;
1557
1558                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1559                                                  MEM_TYPE_PAGE_SHARED, NULL);
1560                 if (err < 0) {
1561                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1562                         return err;
1563                 }
1564
1565                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1566                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1567         }
1568
1569         return 0;
1570 }
1571
1572 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1573 {
1574         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1575         struct virtnet_info *vi = sq->vq->vdev->priv;
1576         unsigned int index = vq2txq(sq->vq);
1577         struct netdev_queue *txq;
1578
1579         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1580                 /* We don't need to enable cb for XDP */
1581                 napi_complete_done(napi, 0);
1582                 return 0;
1583         }
1584
1585         txq = netdev_get_tx_queue(vi->dev, index);
1586         __netif_tx_lock(txq, raw_smp_processor_id());
1587         free_old_xmit_skbs(sq, true);
1588         __netif_tx_unlock(txq);
1589
1590         virtqueue_napi_complete(napi, sq->vq, 0);
1591
1592         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1593                 netif_tx_wake_queue(txq);
1594
1595         return 0;
1596 }
1597
1598 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1599 {
1600         struct virtio_net_hdr_mrg_rxbuf *hdr;
1601         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1602         struct virtnet_info *vi = sq->vq->vdev->priv;
1603         int num_sg;
1604         unsigned hdr_len = vi->hdr_len;
1605         bool can_push;
1606
1607         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1608
1609         can_push = vi->any_header_sg &&
1610                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1611                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1612         /* Even if we can, don't push here yet as this would skew
1613          * csum_start offset below. */
1614         if (can_push)
1615                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1616         else
1617                 hdr = skb_vnet_hdr(skb);
1618
1619         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1620                                     virtio_is_little_endian(vi->vdev), false,
1621                                     0))
1622                 BUG();
1623
1624         if (vi->mergeable_rx_bufs)
1625                 hdr->num_buffers = 0;
1626
1627         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1628         if (can_push) {
1629                 __skb_push(skb, hdr_len);
1630                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1631                 if (unlikely(num_sg < 0))
1632                         return num_sg;
1633                 /* Pull header back to avoid skew in tx bytes calculations. */
1634                 __skb_pull(skb, hdr_len);
1635         } else {
1636                 sg_set_buf(sq->sg, hdr, hdr_len);
1637                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1638                 if (unlikely(num_sg < 0))
1639                         return num_sg;
1640                 num_sg++;
1641         }
1642         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1643 }
1644
1645 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1646 {
1647         struct virtnet_info *vi = netdev_priv(dev);
1648         int qnum = skb_get_queue_mapping(skb);
1649         struct send_queue *sq = &vi->sq[qnum];
1650         int err;
1651         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1652         bool kick = !netdev_xmit_more();
1653         bool use_napi = sq->napi.weight;
1654
1655         /* Free up any pending old buffers before queueing new ones. */
1656         free_old_xmit_skbs(sq, false);
1657
1658         if (use_napi && kick)
1659                 virtqueue_enable_cb_delayed(sq->vq);
1660
1661         /* timestamp packet in software */
1662         skb_tx_timestamp(skb);
1663
1664         /* Try to transmit */
1665         err = xmit_skb(sq, skb);
1666
1667         /* This should not happen! */
1668         if (unlikely(err)) {
1669                 dev->stats.tx_fifo_errors++;
1670                 if (net_ratelimit())
1671                         dev_warn(&dev->dev,
1672                                  "Unexpected TXQ (%d) queue failure: %d\n",
1673                                  qnum, err);
1674                 dev->stats.tx_dropped++;
1675                 dev_kfree_skb_any(skb);
1676                 return NETDEV_TX_OK;
1677         }
1678
1679         /* Don't wait up for transmitted skbs to be freed. */
1680         if (!use_napi) {
1681                 skb_orphan(skb);
1682                 nf_reset_ct(skb);
1683         }
1684
1685         /* If running out of space, stop queue to avoid getting packets that we
1686          * are then unable to transmit.
1687          * An alternative would be to force queuing layer to requeue the skb by
1688          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1689          * returned in a normal path of operation: it means that driver is not
1690          * maintaining the TX queue stop/start state properly, and causes
1691          * the stack to do a non-trivial amount of useless work.
1692          * Since most packets only take 1 or 2 ring slots, stopping the queue
1693          * early means 16 slots are typically wasted.
1694          */
1695         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1696                 netif_stop_subqueue(dev, qnum);
1697                 if (!use_napi &&
1698                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1699                         /* More just got used, free them then recheck. */
1700                         free_old_xmit_skbs(sq, false);
1701                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1702                                 netif_start_subqueue(dev, qnum);
1703                                 virtqueue_disable_cb(sq->vq);
1704                         }
1705                 }
1706         }
1707
1708         if (kick || netif_xmit_stopped(txq)) {
1709                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1710                         u64_stats_update_begin(&sq->stats.syncp);
1711                         sq->stats.kicks++;
1712                         u64_stats_update_end(&sq->stats.syncp);
1713                 }
1714         }
1715
1716         return NETDEV_TX_OK;
1717 }
1718
1719 /*
1720  * Send command via the control virtqueue and check status.  Commands
1721  * supported by the hypervisor, as indicated by feature bits, should
1722  * never fail unless improperly formatted.
1723  */
1724 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1725                                  struct scatterlist *out)
1726 {
1727         struct scatterlist *sgs[4], hdr, stat;
1728         unsigned out_num = 0, tmp;
1729
1730         /* Caller should know better */
1731         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1732
1733         vi->ctrl->status = ~0;
1734         vi->ctrl->hdr.class = class;
1735         vi->ctrl->hdr.cmd = cmd;
1736         /* Add header */
1737         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1738         sgs[out_num++] = &hdr;
1739
1740         if (out)
1741                 sgs[out_num++] = out;
1742
1743         /* Add return status. */
1744         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1745         sgs[out_num] = &stat;
1746
1747         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1748         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1749
1750         if (unlikely(!virtqueue_kick(vi->cvq)))
1751                 return vi->ctrl->status == VIRTIO_NET_OK;
1752
1753         /* Spin for a response, the kick causes an ioport write, trapping
1754          * into the hypervisor, so the request should be handled immediately.
1755          */
1756         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1757                !virtqueue_is_broken(vi->cvq))
1758                 cpu_relax();
1759
1760         return vi->ctrl->status == VIRTIO_NET_OK;
1761 }
1762
1763 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1764 {
1765         struct virtnet_info *vi = netdev_priv(dev);
1766         struct virtio_device *vdev = vi->vdev;
1767         int ret;
1768         struct sockaddr *addr;
1769         struct scatterlist sg;
1770
1771         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1772                 return -EOPNOTSUPP;
1773
1774         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1775         if (!addr)
1776                 return -ENOMEM;
1777
1778         ret = eth_prepare_mac_addr_change(dev, addr);
1779         if (ret)
1780                 goto out;
1781
1782         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1783                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1784                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1785                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1786                         dev_warn(&vdev->dev,
1787                                  "Failed to set mac address by vq command.\n");
1788                         ret = -EINVAL;
1789                         goto out;
1790                 }
1791         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1792                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1793                 unsigned int i;
1794
1795                 /* Naturally, this has an atomicity problem. */
1796                 for (i = 0; i < dev->addr_len; i++)
1797                         virtio_cwrite8(vdev,
1798                                        offsetof(struct virtio_net_config, mac) +
1799                                        i, addr->sa_data[i]);
1800         }
1801
1802         eth_commit_mac_addr_change(dev, p);
1803         ret = 0;
1804
1805 out:
1806         kfree(addr);
1807         return ret;
1808 }
1809
1810 static void virtnet_stats(struct net_device *dev,
1811                           struct rtnl_link_stats64 *tot)
1812 {
1813         struct virtnet_info *vi = netdev_priv(dev);
1814         unsigned int start;
1815         int i;
1816
1817         for (i = 0; i < vi->max_queue_pairs; i++) {
1818                 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1819                 struct receive_queue *rq = &vi->rq[i];
1820                 struct send_queue *sq = &vi->sq[i];
1821
1822                 do {
1823                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1824                         tpackets = sq->stats.packets;
1825                         tbytes   = sq->stats.bytes;
1826                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1827
1828                 do {
1829                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1830                         rpackets = rq->stats.packets;
1831                         rbytes   = rq->stats.bytes;
1832                         rdrops   = rq->stats.drops;
1833                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1834
1835                 tot->rx_packets += rpackets;
1836                 tot->tx_packets += tpackets;
1837                 tot->rx_bytes   += rbytes;
1838                 tot->tx_bytes   += tbytes;
1839                 tot->rx_dropped += rdrops;
1840         }
1841
1842         tot->tx_dropped = dev->stats.tx_dropped;
1843         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1844         tot->rx_length_errors = dev->stats.rx_length_errors;
1845         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1846 }
1847
1848 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1849 {
1850         rtnl_lock();
1851         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1852                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1853                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1854         rtnl_unlock();
1855 }
1856
1857 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1858 {
1859         struct scatterlist sg;
1860         struct net_device *dev = vi->dev;
1861
1862         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1863                 return 0;
1864
1865         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1866         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1867
1868         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1869                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1870                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1871                          queue_pairs);
1872                 return -EINVAL;
1873         } else {
1874                 vi->curr_queue_pairs = queue_pairs;
1875                 /* virtnet_open() will refill when device is going to up. */
1876                 if (dev->flags & IFF_UP)
1877                         schedule_delayed_work(&vi->refill, 0);
1878         }
1879
1880         return 0;
1881 }
1882
1883 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1884 {
1885         int err;
1886
1887         rtnl_lock();
1888         err = _virtnet_set_queues(vi, queue_pairs);
1889         rtnl_unlock();
1890         return err;
1891 }
1892
1893 static int virtnet_close(struct net_device *dev)
1894 {
1895         struct virtnet_info *vi = netdev_priv(dev);
1896         int i;
1897
1898         /* Make sure refill_work doesn't re-enable napi! */
1899         cancel_delayed_work_sync(&vi->refill);
1900
1901         for (i = 0; i < vi->max_queue_pairs; i++) {
1902                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1903                 napi_disable(&vi->rq[i].napi);
1904                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1905         }
1906
1907         return 0;
1908 }
1909
1910 static void virtnet_set_rx_mode(struct net_device *dev)
1911 {
1912         struct virtnet_info *vi = netdev_priv(dev);
1913         struct scatterlist sg[2];
1914         struct virtio_net_ctrl_mac *mac_data;
1915         struct netdev_hw_addr *ha;
1916         int uc_count;
1917         int mc_count;
1918         void *buf;
1919         int i;
1920
1921         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1922         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1923                 return;
1924
1925         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1926         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1927
1928         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1929
1930         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1931                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1932                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1933                          vi->ctrl->promisc ? "en" : "dis");
1934
1935         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1936
1937         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1938                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1939                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1940                          vi->ctrl->allmulti ? "en" : "dis");
1941
1942         uc_count = netdev_uc_count(dev);
1943         mc_count = netdev_mc_count(dev);
1944         /* MAC filter - use one buffer for both lists */
1945         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1946                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1947         mac_data = buf;
1948         if (!buf)
1949                 return;
1950
1951         sg_init_table(sg, 2);
1952
1953         /* Store the unicast list and count in the front of the buffer */
1954         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1955         i = 0;
1956         netdev_for_each_uc_addr(ha, dev)
1957                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1958
1959         sg_set_buf(&sg[0], mac_data,
1960                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1961
1962         /* multicast list and count fill the end */
1963         mac_data = (void *)&mac_data->macs[uc_count][0];
1964
1965         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1966         i = 0;
1967         netdev_for_each_mc_addr(ha, dev)
1968                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1969
1970         sg_set_buf(&sg[1], mac_data,
1971                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1972
1973         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1974                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1975                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1976
1977         kfree(buf);
1978 }
1979
1980 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1981                                    __be16 proto, u16 vid)
1982 {
1983         struct virtnet_info *vi = netdev_priv(dev);
1984         struct scatterlist sg;
1985
1986         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1987         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1988
1989         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1990                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1991                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1992         return 0;
1993 }
1994
1995 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1996                                     __be16 proto, u16 vid)
1997 {
1998         struct virtnet_info *vi = netdev_priv(dev);
1999         struct scatterlist sg;
2000
2001         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2002         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2003
2004         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2005                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2006                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2007         return 0;
2008 }
2009
2010 static void virtnet_clean_affinity(struct virtnet_info *vi)
2011 {
2012         int i;
2013
2014         if (vi->affinity_hint_set) {
2015                 for (i = 0; i < vi->max_queue_pairs; i++) {
2016                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
2017                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
2018                 }
2019
2020                 vi->affinity_hint_set = false;
2021         }
2022 }
2023
2024 static void virtnet_set_affinity(struct virtnet_info *vi)
2025 {
2026         cpumask_var_t mask;
2027         int stragglers;
2028         int group_size;
2029         int i, j, cpu;
2030         int num_cpu;
2031         int stride;
2032
2033         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2034                 virtnet_clean_affinity(vi);
2035                 return;
2036         }
2037
2038         num_cpu = num_online_cpus();
2039         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2040         stragglers = num_cpu >= vi->curr_queue_pairs ?
2041                         num_cpu % vi->curr_queue_pairs :
2042                         0;
2043         cpu = cpumask_next(-1, cpu_online_mask);
2044
2045         for (i = 0; i < vi->curr_queue_pairs; i++) {
2046                 group_size = stride + (i < stragglers ? 1 : 0);
2047
2048                 for (j = 0; j < group_size; j++) {
2049                         cpumask_set_cpu(cpu, mask);
2050                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2051                                                 nr_cpu_ids, false);
2052                 }
2053                 virtqueue_set_affinity(vi->rq[i].vq, mask);
2054                 virtqueue_set_affinity(vi->sq[i].vq, mask);
2055                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2056                 cpumask_clear(mask);
2057         }
2058
2059         vi->affinity_hint_set = true;
2060         free_cpumask_var(mask);
2061 }
2062
2063 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2064 {
2065         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2066                                                    node);
2067         virtnet_set_affinity(vi);
2068         return 0;
2069 }
2070
2071 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2072 {
2073         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2074                                                    node_dead);
2075         virtnet_set_affinity(vi);
2076         return 0;
2077 }
2078
2079 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2080 {
2081         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2082                                                    node);
2083
2084         virtnet_clean_affinity(vi);
2085         return 0;
2086 }
2087
2088 static enum cpuhp_state virtionet_online;
2089
2090 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2091 {
2092         int ret;
2093
2094         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2095         if (ret)
2096                 return ret;
2097         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2098                                                &vi->node_dead);
2099         if (!ret)
2100                 return ret;
2101         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2102         return ret;
2103 }
2104
2105 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2106 {
2107         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2108         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2109                                             &vi->node_dead);
2110 }
2111
2112 static void virtnet_get_ringparam(struct net_device *dev,
2113                                 struct ethtool_ringparam *ring)
2114 {
2115         struct virtnet_info *vi = netdev_priv(dev);
2116
2117         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2118         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2119         ring->rx_pending = ring->rx_max_pending;
2120         ring->tx_pending = ring->tx_max_pending;
2121 }
2122
2123
2124 static void virtnet_get_drvinfo(struct net_device *dev,
2125                                 struct ethtool_drvinfo *info)
2126 {
2127         struct virtnet_info *vi = netdev_priv(dev);
2128         struct virtio_device *vdev = vi->vdev;
2129
2130         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2131         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2132         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2133
2134 }
2135
2136 /* TODO: Eliminate OOO packets during switching */
2137 static int virtnet_set_channels(struct net_device *dev,
2138                                 struct ethtool_channels *channels)
2139 {
2140         struct virtnet_info *vi = netdev_priv(dev);
2141         u16 queue_pairs = channels->combined_count;
2142         int err;
2143
2144         /* We don't support separate rx/tx channels.
2145          * We don't allow setting 'other' channels.
2146          */
2147         if (channels->rx_count || channels->tx_count || channels->other_count)
2148                 return -EINVAL;
2149
2150         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2151                 return -EINVAL;
2152
2153         /* For now we don't support modifying channels while XDP is loaded
2154          * also when XDP is loaded all RX queues have XDP programs so we only
2155          * need to check a single RX queue.
2156          */
2157         if (vi->rq[0].xdp_prog)
2158                 return -EINVAL;
2159
2160         get_online_cpus();
2161         err = _virtnet_set_queues(vi, queue_pairs);
2162         if (err) {
2163                 put_online_cpus();
2164                 goto err;
2165         }
2166         virtnet_set_affinity(vi);
2167         put_online_cpus();
2168
2169         netif_set_real_num_tx_queues(dev, queue_pairs);
2170         netif_set_real_num_rx_queues(dev, queue_pairs);
2171  err:
2172         return err;
2173 }
2174
2175 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2176 {
2177         struct virtnet_info *vi = netdev_priv(dev);
2178         unsigned int i, j;
2179         u8 *p = data;
2180
2181         switch (stringset) {
2182         case ETH_SS_STATS:
2183                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2184                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2185                                 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2186                                                 virtnet_rq_stats_desc[j].desc);
2187                 }
2188
2189                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2190                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2191                                 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2192                                                 virtnet_sq_stats_desc[j].desc);
2193                 }
2194                 break;
2195         }
2196 }
2197
2198 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2199 {
2200         struct virtnet_info *vi = netdev_priv(dev);
2201
2202         switch (sset) {
2203         case ETH_SS_STATS:
2204                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2205                                                VIRTNET_SQ_STATS_LEN);
2206         default:
2207                 return -EOPNOTSUPP;
2208         }
2209 }
2210
2211 static void virtnet_get_ethtool_stats(struct net_device *dev,
2212                                       struct ethtool_stats *stats, u64 *data)
2213 {
2214         struct virtnet_info *vi = netdev_priv(dev);
2215         unsigned int idx = 0, start, i, j;
2216         const u8 *stats_base;
2217         size_t offset;
2218
2219         for (i = 0; i < vi->curr_queue_pairs; i++) {
2220                 struct receive_queue *rq = &vi->rq[i];
2221
2222                 stats_base = (u8 *)&rq->stats;
2223                 do {
2224                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2225                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2226                                 offset = virtnet_rq_stats_desc[j].offset;
2227                                 data[idx + j] = *(u64 *)(stats_base + offset);
2228                         }
2229                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2230                 idx += VIRTNET_RQ_STATS_LEN;
2231         }
2232
2233         for (i = 0; i < vi->curr_queue_pairs; i++) {
2234                 struct send_queue *sq = &vi->sq[i];
2235
2236                 stats_base = (u8 *)&sq->stats;
2237                 do {
2238                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2239                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2240                                 offset = virtnet_sq_stats_desc[j].offset;
2241                                 data[idx + j] = *(u64 *)(stats_base + offset);
2242                         }
2243                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2244                 idx += VIRTNET_SQ_STATS_LEN;
2245         }
2246 }
2247
2248 static void virtnet_get_channels(struct net_device *dev,
2249                                  struct ethtool_channels *channels)
2250 {
2251         struct virtnet_info *vi = netdev_priv(dev);
2252
2253         channels->combined_count = vi->curr_queue_pairs;
2254         channels->max_combined = vi->max_queue_pairs;
2255         channels->max_other = 0;
2256         channels->rx_count = 0;
2257         channels->tx_count = 0;
2258         channels->other_count = 0;
2259 }
2260
2261 static int virtnet_set_link_ksettings(struct net_device *dev,
2262                                       const struct ethtool_link_ksettings *cmd)
2263 {
2264         struct virtnet_info *vi = netdev_priv(dev);
2265
2266         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2267                                                   &vi->speed, &vi->duplex);
2268 }
2269
2270 static int virtnet_get_link_ksettings(struct net_device *dev,
2271                                       struct ethtool_link_ksettings *cmd)
2272 {
2273         struct virtnet_info *vi = netdev_priv(dev);
2274
2275         cmd->base.speed = vi->speed;
2276         cmd->base.duplex = vi->duplex;
2277         cmd->base.port = PORT_OTHER;
2278
2279         return 0;
2280 }
2281
2282 static int virtnet_set_coalesce(struct net_device *dev,
2283                                 struct ethtool_coalesce *ec)
2284 {
2285         struct virtnet_info *vi = netdev_priv(dev);
2286         int i, napi_weight;
2287
2288         if (ec->tx_max_coalesced_frames > 1 ||
2289             ec->rx_max_coalesced_frames != 1)
2290                 return -EINVAL;
2291
2292         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2293         if (napi_weight ^ vi->sq[0].napi.weight) {
2294                 if (dev->flags & IFF_UP)
2295                         return -EBUSY;
2296                 for (i = 0; i < vi->max_queue_pairs; i++)
2297                         vi->sq[i].napi.weight = napi_weight;
2298         }
2299
2300         return 0;
2301 }
2302
2303 static int virtnet_get_coalesce(struct net_device *dev,
2304                                 struct ethtool_coalesce *ec)
2305 {
2306         struct ethtool_coalesce ec_default = {
2307                 .cmd = ETHTOOL_GCOALESCE,
2308                 .rx_max_coalesced_frames = 1,
2309         };
2310         struct virtnet_info *vi = netdev_priv(dev);
2311
2312         memcpy(ec, &ec_default, sizeof(ec_default));
2313
2314         if (vi->sq[0].napi.weight)
2315                 ec->tx_max_coalesced_frames = 1;
2316
2317         return 0;
2318 }
2319
2320 static void virtnet_init_settings(struct net_device *dev)
2321 {
2322         struct virtnet_info *vi = netdev_priv(dev);
2323
2324         vi->speed = SPEED_UNKNOWN;
2325         vi->duplex = DUPLEX_UNKNOWN;
2326 }
2327
2328 static void virtnet_update_settings(struct virtnet_info *vi)
2329 {
2330         u32 speed;
2331         u8 duplex;
2332
2333         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2334                 return;
2335
2336         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2337
2338         if (ethtool_validate_speed(speed))
2339                 vi->speed = speed;
2340
2341         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2342
2343         if (ethtool_validate_duplex(duplex))
2344                 vi->duplex = duplex;
2345 }
2346
2347 static const struct ethtool_ops virtnet_ethtool_ops = {
2348         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2349         .get_drvinfo = virtnet_get_drvinfo,
2350         .get_link = ethtool_op_get_link,
2351         .get_ringparam = virtnet_get_ringparam,
2352         .get_strings = virtnet_get_strings,
2353         .get_sset_count = virtnet_get_sset_count,
2354         .get_ethtool_stats = virtnet_get_ethtool_stats,
2355         .set_channels = virtnet_set_channels,
2356         .get_channels = virtnet_get_channels,
2357         .get_ts_info = ethtool_op_get_ts_info,
2358         .get_link_ksettings = virtnet_get_link_ksettings,
2359         .set_link_ksettings = virtnet_set_link_ksettings,
2360         .set_coalesce = virtnet_set_coalesce,
2361         .get_coalesce = virtnet_get_coalesce,
2362 };
2363
2364 static void virtnet_freeze_down(struct virtio_device *vdev)
2365 {
2366         struct virtnet_info *vi = vdev->priv;
2367         int i;
2368
2369         /* Make sure no work handler is accessing the device */
2370         flush_work(&vi->config_work);
2371
2372         netif_tx_lock_bh(vi->dev);
2373         netif_device_detach(vi->dev);
2374         netif_tx_unlock_bh(vi->dev);
2375         cancel_delayed_work_sync(&vi->refill);
2376
2377         if (netif_running(vi->dev)) {
2378                 for (i = 0; i < vi->max_queue_pairs; i++) {
2379                         napi_disable(&vi->rq[i].napi);
2380                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2381                 }
2382         }
2383 }
2384
2385 static int init_vqs(struct virtnet_info *vi);
2386
2387 static int virtnet_restore_up(struct virtio_device *vdev)
2388 {
2389         struct virtnet_info *vi = vdev->priv;
2390         int err, i;
2391
2392         err = init_vqs(vi);
2393         if (err)
2394                 return err;
2395
2396         virtio_device_ready(vdev);
2397
2398         if (netif_running(vi->dev)) {
2399                 for (i = 0; i < vi->curr_queue_pairs; i++)
2400                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2401                                 schedule_delayed_work(&vi->refill, 0);
2402
2403                 for (i = 0; i < vi->max_queue_pairs; i++) {
2404                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2405                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2406                                                &vi->sq[i].napi);
2407                 }
2408         }
2409
2410         netif_tx_lock_bh(vi->dev);
2411         netif_device_attach(vi->dev);
2412         netif_tx_unlock_bh(vi->dev);
2413         return err;
2414 }
2415
2416 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2417 {
2418         struct scatterlist sg;
2419         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2420
2421         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2422
2423         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2424                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2425                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2426                 return -EINVAL;
2427         }
2428
2429         return 0;
2430 }
2431
2432 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2433 {
2434         u64 offloads = 0;
2435
2436         if (!vi->guest_offloads)
2437                 return 0;
2438
2439         return virtnet_set_guest_offloads(vi, offloads);
2440 }
2441
2442 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2443 {
2444         u64 offloads = vi->guest_offloads;
2445
2446         if (!vi->guest_offloads)
2447                 return 0;
2448
2449         return virtnet_set_guest_offloads(vi, offloads);
2450 }
2451
2452 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2453                            struct netlink_ext_ack *extack)
2454 {
2455         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2456         struct virtnet_info *vi = netdev_priv(dev);
2457         struct bpf_prog *old_prog;
2458         u16 xdp_qp = 0, curr_qp;
2459         int i, err;
2460
2461         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2462             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2463                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2464                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2465                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2466                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2467                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2468                 return -EOPNOTSUPP;
2469         }
2470
2471         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2472                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2473                 return -EINVAL;
2474         }
2475
2476         if (dev->mtu > max_sz) {
2477                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2478                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2479                 return -EINVAL;
2480         }
2481
2482         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2483         if (prog)
2484                 xdp_qp = nr_cpu_ids;
2485
2486         /* XDP requires extra queues for XDP_TX */
2487         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2488                 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",
2489                             curr_qp + xdp_qp, vi->max_queue_pairs);
2490                 xdp_qp = 0;
2491         }
2492
2493         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2494         if (!prog && !old_prog)
2495                 return 0;
2496
2497         if (prog)
2498                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2499
2500         /* Make sure NAPI is not using any XDP TX queues for RX. */
2501         if (netif_running(dev)) {
2502                 for (i = 0; i < vi->max_queue_pairs; i++) {
2503                         napi_disable(&vi->rq[i].napi);
2504                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2505                 }
2506         }
2507
2508         if (!prog) {
2509                 for (i = 0; i < vi->max_queue_pairs; i++) {
2510                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2511                         if (i == 0)
2512                                 virtnet_restore_guest_offloads(vi);
2513                 }
2514                 synchronize_net();
2515         }
2516
2517         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2518         if (err)
2519                 goto err;
2520         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2521         vi->xdp_queue_pairs = xdp_qp;
2522
2523         if (prog) {
2524                 vi->xdp_enabled = true;
2525                 for (i = 0; i < vi->max_queue_pairs; i++) {
2526                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2527                         if (i == 0 && !old_prog)
2528                                 virtnet_clear_guest_offloads(vi);
2529                 }
2530         } else {
2531                 vi->xdp_enabled = false;
2532         }
2533
2534         for (i = 0; i < vi->max_queue_pairs; i++) {
2535                 if (old_prog)
2536                         bpf_prog_put(old_prog);
2537                 if (netif_running(dev)) {
2538                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2539                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2540                                                &vi->sq[i].napi);
2541                 }
2542         }
2543
2544         return 0;
2545
2546 err:
2547         if (!prog) {
2548                 virtnet_clear_guest_offloads(vi);
2549                 for (i = 0; i < vi->max_queue_pairs; i++)
2550                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2551         }
2552
2553         if (netif_running(dev)) {
2554                 for (i = 0; i < vi->max_queue_pairs; i++) {
2555                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2556                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2557                                                &vi->sq[i].napi);
2558                 }
2559         }
2560         if (prog)
2561                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2562         return err;
2563 }
2564
2565 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2566 {
2567         switch (xdp->command) {
2568         case XDP_SETUP_PROG:
2569                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2570         default:
2571                 return -EINVAL;
2572         }
2573 }
2574
2575 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2576                                       size_t len)
2577 {
2578         struct virtnet_info *vi = netdev_priv(dev);
2579         int ret;
2580
2581         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2582                 return -EOPNOTSUPP;
2583
2584         ret = snprintf(buf, len, "sby");
2585         if (ret >= len)
2586                 return -EOPNOTSUPP;
2587
2588         return 0;
2589 }
2590
2591 static int virtnet_set_features(struct net_device *dev,
2592                                 netdev_features_t features)
2593 {
2594         struct virtnet_info *vi = netdev_priv(dev);
2595         u64 offloads;
2596         int err;
2597
2598         if ((dev->features ^ features) & NETIF_F_LRO) {
2599                 if (vi->xdp_enabled)
2600                         return -EBUSY;
2601
2602                 if (features & NETIF_F_LRO)
2603                         offloads = vi->guest_offloads_capable;
2604                 else
2605                         offloads = vi->guest_offloads_capable &
2606                                    ~GUEST_OFFLOAD_LRO_MASK;
2607
2608                 err = virtnet_set_guest_offloads(vi, offloads);
2609                 if (err)
2610                         return err;
2611                 vi->guest_offloads = offloads;
2612         }
2613
2614         return 0;
2615 }
2616
2617 static const struct net_device_ops virtnet_netdev = {
2618         .ndo_open            = virtnet_open,
2619         .ndo_stop            = virtnet_close,
2620         .ndo_start_xmit      = start_xmit,
2621         .ndo_validate_addr   = eth_validate_addr,
2622         .ndo_set_mac_address = virtnet_set_mac_address,
2623         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2624         .ndo_get_stats64     = virtnet_stats,
2625         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2626         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2627         .ndo_bpf                = virtnet_xdp,
2628         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2629         .ndo_features_check     = passthru_features_check,
2630         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2631         .ndo_set_features       = virtnet_set_features,
2632 };
2633
2634 static void virtnet_config_changed_work(struct work_struct *work)
2635 {
2636         struct virtnet_info *vi =
2637                 container_of(work, struct virtnet_info, config_work);
2638         u16 v;
2639
2640         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2641                                  struct virtio_net_config, status, &v) < 0)
2642                 return;
2643
2644         if (v & VIRTIO_NET_S_ANNOUNCE) {
2645                 netdev_notify_peers(vi->dev);
2646                 virtnet_ack_link_announce(vi);
2647         }
2648
2649         /* Ignore unknown (future) status bits */
2650         v &= VIRTIO_NET_S_LINK_UP;
2651
2652         if (vi->status == v)
2653                 return;
2654
2655         vi->status = v;
2656
2657         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2658                 virtnet_update_settings(vi);
2659                 netif_carrier_on(vi->dev);
2660                 netif_tx_wake_all_queues(vi->dev);
2661         } else {
2662                 netif_carrier_off(vi->dev);
2663                 netif_tx_stop_all_queues(vi->dev);
2664         }
2665 }
2666
2667 static void virtnet_config_changed(struct virtio_device *vdev)
2668 {
2669         struct virtnet_info *vi = vdev->priv;
2670
2671         schedule_work(&vi->config_work);
2672 }
2673
2674 static void virtnet_free_queues(struct virtnet_info *vi)
2675 {
2676         int i;
2677
2678         for (i = 0; i < vi->max_queue_pairs; i++) {
2679                 __netif_napi_del(&vi->rq[i].napi);
2680                 __netif_napi_del(&vi->sq[i].napi);
2681         }
2682
2683         /* We called __netif_napi_del(),
2684          * we need to respect an RCU grace period before freeing vi->rq
2685          */
2686         synchronize_net();
2687
2688         kfree(vi->rq);
2689         kfree(vi->sq);
2690         kfree(vi->ctrl);
2691 }
2692
2693 static void _free_receive_bufs(struct virtnet_info *vi)
2694 {
2695         struct bpf_prog *old_prog;
2696         int i;
2697
2698         for (i = 0; i < vi->max_queue_pairs; i++) {
2699                 while (vi->rq[i].pages)
2700                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2701
2702                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2703                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2704                 if (old_prog)
2705                         bpf_prog_put(old_prog);
2706         }
2707 }
2708
2709 static void free_receive_bufs(struct virtnet_info *vi)
2710 {
2711         rtnl_lock();
2712         _free_receive_bufs(vi);
2713         rtnl_unlock();
2714 }
2715
2716 static void free_receive_page_frags(struct virtnet_info *vi)
2717 {
2718         int i;
2719         for (i = 0; i < vi->max_queue_pairs; i++)
2720                 if (vi->rq[i].alloc_frag.page)
2721                         put_page(vi->rq[i].alloc_frag.page);
2722 }
2723
2724 static void free_unused_bufs(struct virtnet_info *vi)
2725 {
2726         void *buf;
2727         int i;
2728
2729         for (i = 0; i < vi->max_queue_pairs; i++) {
2730                 struct virtqueue *vq = vi->sq[i].vq;
2731                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2732                         if (!is_xdp_frame(buf))
2733                                 dev_kfree_skb(buf);
2734                         else
2735                                 xdp_return_frame(ptr_to_xdp(buf));
2736                 }
2737         }
2738
2739         for (i = 0; i < vi->max_queue_pairs; i++) {
2740                 struct virtqueue *vq = vi->rq[i].vq;
2741
2742                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2743                         if (vi->mergeable_rx_bufs) {
2744                                 put_page(virt_to_head_page(buf));
2745                         } else if (vi->big_packets) {
2746                                 give_pages(&vi->rq[i], buf);
2747                         } else {
2748                                 put_page(virt_to_head_page(buf));
2749                         }
2750                 }
2751         }
2752 }
2753
2754 static void virtnet_del_vqs(struct virtnet_info *vi)
2755 {
2756         struct virtio_device *vdev = vi->vdev;
2757
2758         virtnet_clean_affinity(vi);
2759
2760         vdev->config->del_vqs(vdev);
2761
2762         virtnet_free_queues(vi);
2763 }
2764
2765 /* How large should a single buffer be so a queue full of these can fit at
2766  * least one full packet?
2767  * Logic below assumes the mergeable buffer header is used.
2768  */
2769 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2770 {
2771         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2772         unsigned int rq_size = virtqueue_get_vring_size(vq);
2773         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2774         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2775         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2776
2777         return max(max(min_buf_len, hdr_len) - hdr_len,
2778                    (unsigned int)GOOD_PACKET_LEN);
2779 }
2780
2781 static int virtnet_find_vqs(struct virtnet_info *vi)
2782 {
2783         vq_callback_t **callbacks;
2784         struct virtqueue **vqs;
2785         int ret = -ENOMEM;
2786         int i, total_vqs;
2787         const char **names;
2788         bool *ctx;
2789
2790         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2791          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2792          * possible control vq.
2793          */
2794         total_vqs = vi->max_queue_pairs * 2 +
2795                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2796
2797         /* Allocate space for find_vqs parameters */
2798         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2799         if (!vqs)
2800                 goto err_vq;
2801         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2802         if (!callbacks)
2803                 goto err_callback;
2804         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2805         if (!names)
2806                 goto err_names;
2807         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2808                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2809                 if (!ctx)
2810                         goto err_ctx;
2811         } else {
2812                 ctx = NULL;
2813         }
2814
2815         /* Parameters for control virtqueue, if any */
2816         if (vi->has_cvq) {
2817                 callbacks[total_vqs - 1] = NULL;
2818                 names[total_vqs - 1] = "control";
2819         }
2820
2821         /* Allocate/initialize parameters for send/receive virtqueues */
2822         for (i = 0; i < vi->max_queue_pairs; i++) {
2823                 callbacks[rxq2vq(i)] = skb_recv_done;
2824                 callbacks[txq2vq(i)] = skb_xmit_done;
2825                 sprintf(vi->rq[i].name, "input.%d", i);
2826                 sprintf(vi->sq[i].name, "output.%d", i);
2827                 names[rxq2vq(i)] = vi->rq[i].name;
2828                 names[txq2vq(i)] = vi->sq[i].name;
2829                 if (ctx)
2830                         ctx[rxq2vq(i)] = true;
2831         }
2832
2833         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2834                                          names, ctx, NULL);
2835         if (ret)
2836                 goto err_find;
2837
2838         if (vi->has_cvq) {
2839                 vi->cvq = vqs[total_vqs - 1];
2840                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2841                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2842         }
2843
2844         for (i = 0; i < vi->max_queue_pairs; i++) {
2845                 vi->rq[i].vq = vqs[rxq2vq(i)];
2846                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2847                 vi->sq[i].vq = vqs[txq2vq(i)];
2848         }
2849
2850         /* run here: ret == 0. */
2851
2852
2853 err_find:
2854         kfree(ctx);
2855 err_ctx:
2856         kfree(names);
2857 err_names:
2858         kfree(callbacks);
2859 err_callback:
2860         kfree(vqs);
2861 err_vq:
2862         return ret;
2863 }
2864
2865 static int virtnet_alloc_queues(struct virtnet_info *vi)
2866 {
2867         int i;
2868
2869         if (vi->has_cvq) {
2870                 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2871                 if (!vi->ctrl)
2872                         goto err_ctrl;
2873         } else {
2874                 vi->ctrl = NULL;
2875         }
2876         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2877         if (!vi->sq)
2878                 goto err_sq;
2879         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2880         if (!vi->rq)
2881                 goto err_rq;
2882
2883         INIT_DELAYED_WORK(&vi->refill, refill_work);
2884         for (i = 0; i < vi->max_queue_pairs; i++) {
2885                 vi->rq[i].pages = NULL;
2886                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2887                                napi_weight);
2888                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2889                                   napi_tx ? napi_weight : 0);
2890
2891                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2892                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2893                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2894
2895                 u64_stats_init(&vi->rq[i].stats.syncp);
2896                 u64_stats_init(&vi->sq[i].stats.syncp);
2897         }
2898
2899         return 0;
2900
2901 err_rq:
2902         kfree(vi->sq);
2903 err_sq:
2904         kfree(vi->ctrl);
2905 err_ctrl:
2906         return -ENOMEM;
2907 }
2908
2909 static int init_vqs(struct virtnet_info *vi)
2910 {
2911         int ret;
2912
2913         /* Allocate send & receive queues */
2914         ret = virtnet_alloc_queues(vi);
2915         if (ret)
2916                 goto err;
2917
2918         ret = virtnet_find_vqs(vi);
2919         if (ret)
2920                 goto err_free;
2921
2922         get_online_cpus();
2923         virtnet_set_affinity(vi);
2924         put_online_cpus();
2925
2926         return 0;
2927
2928 err_free:
2929         virtnet_free_queues(vi);
2930 err:
2931         return ret;
2932 }
2933
2934 #ifdef CONFIG_SYSFS
2935 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2936                 char *buf)
2937 {
2938         struct virtnet_info *vi = netdev_priv(queue->dev);
2939         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2940         unsigned int headroom = virtnet_get_headroom(vi);
2941         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2942         struct ewma_pkt_len *avg;
2943
2944         BUG_ON(queue_index >= vi->max_queue_pairs);
2945         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2946         return sprintf(buf, "%u\n",
2947                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
2948                                        SKB_DATA_ALIGN(headroom + tailroom)));
2949 }
2950
2951 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2952         __ATTR_RO(mergeable_rx_buffer_size);
2953
2954 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2955         &mergeable_rx_buffer_size_attribute.attr,
2956         NULL
2957 };
2958
2959 static const struct attribute_group virtio_net_mrg_rx_group = {
2960         .name = "virtio_net",
2961         .attrs = virtio_net_mrg_rx_attrs
2962 };
2963 #endif
2964
2965 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2966                                     unsigned int fbit,
2967                                     const char *fname, const char *dname)
2968 {
2969         if (!virtio_has_feature(vdev, fbit))
2970                 return false;
2971
2972         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2973                 fname, dname);
2974
2975         return true;
2976 }
2977
2978 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2979         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2980
2981 static bool virtnet_validate_features(struct virtio_device *vdev)
2982 {
2983         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2984             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2985                              "VIRTIO_NET_F_CTRL_VQ") ||
2986              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2987                              "VIRTIO_NET_F_CTRL_VQ") ||
2988              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2989                              "VIRTIO_NET_F_CTRL_VQ") ||
2990              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2991              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2992                              "VIRTIO_NET_F_CTRL_VQ"))) {
2993                 return false;
2994         }
2995
2996         return true;
2997 }
2998
2999 #define MIN_MTU ETH_MIN_MTU
3000 #define MAX_MTU ETH_MAX_MTU
3001
3002 static int virtnet_validate(struct virtio_device *vdev)
3003 {
3004         if (!vdev->config->get) {
3005                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3006                         __func__);
3007                 return -EINVAL;
3008         }
3009
3010         if (!virtnet_validate_features(vdev))
3011                 return -EINVAL;
3012
3013         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3014                 int mtu = virtio_cread16(vdev,
3015                                          offsetof(struct virtio_net_config,
3016                                                   mtu));
3017                 if (mtu < MIN_MTU)
3018                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3019         }
3020
3021         return 0;
3022 }
3023
3024 static int virtnet_probe(struct virtio_device *vdev)
3025 {
3026         int i, err = -ENOMEM;
3027         struct net_device *dev;
3028         struct virtnet_info *vi;
3029         u16 max_queue_pairs;
3030         int mtu;
3031
3032         /* Find if host supports multiqueue virtio_net device */
3033         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3034                                    struct virtio_net_config,
3035                                    max_virtqueue_pairs, &max_queue_pairs);
3036
3037         /* We need at least 2 queue's */
3038         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3039             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3040             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3041                 max_queue_pairs = 1;
3042
3043         /* Allocate ourselves a network device with room for our info */
3044         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3045         if (!dev)
3046                 return -ENOMEM;
3047
3048         /* Set up network device as normal. */
3049         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3050                            IFF_TX_SKB_NO_LINEAR;
3051         dev->netdev_ops = &virtnet_netdev;
3052         dev->features = NETIF_F_HIGHDMA;
3053
3054         dev->ethtool_ops = &virtnet_ethtool_ops;
3055         SET_NETDEV_DEV(dev, &vdev->dev);
3056
3057         /* Do we support "hardware" checksums? */
3058         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3059                 /* This opens up the world of extra features. */
3060                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3061                 if (csum)
3062                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3063
3064                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3065                         dev->hw_features |= NETIF_F_TSO
3066                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3067                 }
3068                 /* Individual feature bits: what can host handle? */
3069                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3070                         dev->hw_features |= NETIF_F_TSO;
3071                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3072                         dev->hw_features |= NETIF_F_TSO6;
3073                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3074                         dev->hw_features |= NETIF_F_TSO_ECN;
3075
3076                 dev->features |= NETIF_F_GSO_ROBUST;
3077
3078                 if (gso)
3079                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3080                 /* (!csum && gso) case will be fixed by register_netdev() */
3081         }
3082         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3083                 dev->features |= NETIF_F_RXCSUM;
3084         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3085             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3086                 dev->features |= NETIF_F_LRO;
3087         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3088                 dev->hw_features |= NETIF_F_LRO;
3089
3090         dev->vlan_features = dev->features;
3091
3092         /* MTU range: 68 - 65535 */
3093         dev->min_mtu = MIN_MTU;
3094         dev->max_mtu = MAX_MTU;
3095
3096         /* Configuration may specify what MAC to use.  Otherwise random. */
3097         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3098                 virtio_cread_bytes(vdev,
3099                                    offsetof(struct virtio_net_config, mac),
3100                                    dev->dev_addr, dev->addr_len);
3101         else
3102                 eth_hw_addr_random(dev);
3103
3104         /* Set up our device-specific information */
3105         vi = netdev_priv(dev);
3106         vi->dev = dev;
3107         vi->vdev = vdev;
3108         vdev->priv = vi;
3109
3110         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3111
3112         /* If we can receive ANY GSO packets, we must allocate large ones. */
3113         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3114             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3115             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3116             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3117                 vi->big_packets = true;
3118
3119         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3120                 vi->mergeable_rx_bufs = true;
3121
3122         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3123             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3124                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3125         else
3126                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3127
3128         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3129             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3130                 vi->any_header_sg = true;
3131
3132         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3133                 vi->has_cvq = true;
3134
3135         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3136                 mtu = virtio_cread16(vdev,
3137                                      offsetof(struct virtio_net_config,
3138                                               mtu));
3139                 if (mtu < dev->min_mtu) {
3140                         /* Should never trigger: MTU was previously validated
3141                          * in virtnet_validate.
3142                          */
3143                         dev_err(&vdev->dev,
3144                                 "device MTU appears to have changed it is now %d < %d",
3145                                 mtu, dev->min_mtu);
3146                         err = -EINVAL;
3147                         goto free;
3148                 }
3149
3150                 dev->mtu = mtu;
3151                 dev->max_mtu = mtu;
3152
3153                 /* TODO: size buffers correctly in this case. */
3154                 if (dev->mtu > ETH_DATA_LEN)
3155                         vi->big_packets = true;
3156         }
3157
3158         if (vi->any_header_sg)
3159                 dev->needed_headroom = vi->hdr_len;
3160
3161         /* Enable multiqueue by default */
3162         if (num_online_cpus() >= max_queue_pairs)
3163                 vi->curr_queue_pairs = max_queue_pairs;
3164         else
3165                 vi->curr_queue_pairs = num_online_cpus();
3166         vi->max_queue_pairs = max_queue_pairs;
3167
3168         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3169         err = init_vqs(vi);
3170         if (err)
3171                 goto free;
3172
3173 #ifdef CONFIG_SYSFS
3174         if (vi->mergeable_rx_bufs)
3175                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3176 #endif
3177         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3178         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3179
3180         virtnet_init_settings(dev);
3181
3182         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3183                 vi->failover = net_failover_create(vi->dev);
3184                 if (IS_ERR(vi->failover)) {
3185                         err = PTR_ERR(vi->failover);
3186                         goto free_vqs;
3187                 }
3188         }
3189
3190         err = register_netdev(dev);
3191         if (err) {
3192                 pr_debug("virtio_net: registering device failed\n");
3193                 goto free_failover;
3194         }
3195
3196         virtio_device_ready(vdev);
3197
3198         err = virtnet_cpu_notif_add(vi);
3199         if (err) {
3200                 pr_debug("virtio_net: registering cpu notifier failed\n");
3201                 goto free_unregister_netdev;
3202         }
3203
3204         virtnet_set_queues(vi, vi->curr_queue_pairs);
3205
3206         /* Assume link up if device can't report link status,
3207            otherwise get link status from config. */
3208         netif_carrier_off(dev);
3209         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3210                 schedule_work(&vi->config_work);
3211         } else {
3212                 vi->status = VIRTIO_NET_S_LINK_UP;
3213                 virtnet_update_settings(vi);
3214                 netif_carrier_on(dev);
3215         }
3216
3217         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3218                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3219                         set_bit(guest_offloads[i], &vi->guest_offloads);
3220         vi->guest_offloads_capable = vi->guest_offloads;
3221
3222         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3223                  dev->name, max_queue_pairs);
3224
3225         return 0;
3226
3227 free_unregister_netdev:
3228         vi->vdev->config->reset(vdev);
3229
3230         unregister_netdev(dev);
3231 free_failover:
3232         net_failover_destroy(vi->failover);
3233 free_vqs:
3234         cancel_delayed_work_sync(&vi->refill);
3235         free_receive_page_frags(vi);
3236         virtnet_del_vqs(vi);
3237 free:
3238         free_netdev(dev);
3239         return err;
3240 }
3241
3242 static void remove_vq_common(struct virtnet_info *vi)
3243 {
3244         vi->vdev->config->reset(vi->vdev);
3245
3246         /* Free unused buffers in both send and recv, if any. */
3247         free_unused_bufs(vi);
3248
3249         free_receive_bufs(vi);
3250
3251         free_receive_page_frags(vi);
3252
3253         virtnet_del_vqs(vi);
3254 }
3255
3256 static void virtnet_remove(struct virtio_device *vdev)
3257 {
3258         struct virtnet_info *vi = vdev->priv;
3259
3260         virtnet_cpu_notif_remove(vi);
3261
3262         /* Make sure no work handler is accessing the device. */
3263         flush_work(&vi->config_work);
3264
3265         unregister_netdev(vi->dev);
3266
3267         net_failover_destroy(vi->failover);
3268
3269         remove_vq_common(vi);
3270
3271         free_netdev(vi->dev);
3272 }
3273
3274 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3275 {
3276         struct virtnet_info *vi = vdev->priv;
3277
3278         virtnet_cpu_notif_remove(vi);
3279         virtnet_freeze_down(vdev);
3280         remove_vq_common(vi);
3281
3282         return 0;
3283 }
3284
3285 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3286 {
3287         struct virtnet_info *vi = vdev->priv;
3288         int err;
3289
3290         err = virtnet_restore_up(vdev);
3291         if (err)
3292                 return err;
3293         virtnet_set_queues(vi, vi->curr_queue_pairs);
3294
3295         err = virtnet_cpu_notif_add(vi);
3296         if (err)
3297                 return err;
3298
3299         return 0;
3300 }
3301
3302 static struct virtio_device_id id_table[] = {
3303         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3304         { 0 },
3305 };
3306
3307 #define VIRTNET_FEATURES \
3308         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3309         VIRTIO_NET_F_MAC, \
3310         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3311         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3312         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3313         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3314         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3315         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3316         VIRTIO_NET_F_CTRL_MAC_ADDR, \
3317         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3318         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3319
3320 static unsigned int features[] = {
3321         VIRTNET_FEATURES,
3322 };
3323
3324 static unsigned int features_legacy[] = {
3325         VIRTNET_FEATURES,
3326         VIRTIO_NET_F_GSO,
3327         VIRTIO_F_ANY_LAYOUT,
3328 };
3329
3330 static struct virtio_driver virtio_net_driver = {
3331         .feature_table = features,
3332         .feature_table_size = ARRAY_SIZE(features),
3333         .feature_table_legacy = features_legacy,
3334         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3335         .driver.name =  KBUILD_MODNAME,
3336         .driver.owner = THIS_MODULE,
3337         .id_table =     id_table,
3338         .validate =     virtnet_validate,
3339         .probe =        virtnet_probe,
3340         .remove =       virtnet_remove,
3341         .config_changed = virtnet_config_changed,
3342 #ifdef CONFIG_PM_SLEEP
3343         .freeze =       virtnet_freeze,
3344         .restore =      virtnet_restore,
3345 #endif
3346 };
3347
3348 static __init int virtio_net_driver_init(void)
3349 {
3350         int ret;
3351
3352         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3353                                       virtnet_cpu_online,
3354                                       virtnet_cpu_down_prep);
3355         if (ret < 0)
3356                 goto out;
3357         virtionet_online = ret;
3358         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3359                                       NULL, virtnet_cpu_dead);
3360         if (ret)
3361                 goto err_dead;
3362
3363         ret = register_virtio_driver(&virtio_net_driver);
3364         if (ret)
3365                 goto err_virtio;
3366         return 0;
3367 err_virtio:
3368         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3369 err_dead:
3370         cpuhp_remove_multi_state(virtionet_online);
3371 out:
3372         return ret;
3373 }
3374 module_init(virtio_net_driver_init);
3375
3376 static __exit void virtio_net_driver_exit(void)
3377 {
3378         unregister_virtio_driver(&virtio_net_driver);
3379         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3380         cpuhp_remove_multi_state(virtionet_online);
3381 }
3382 module_exit(virtio_net_driver_exit);
3383
3384 MODULE_DEVICE_TABLE(virtio, id_table);
3385 MODULE_DESCRIPTION("Virtio network driver");
3386 MODULE_LICENSE("GPL");