drivers/net/virtio_net.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* A network driver using virtio.
   3  *
   4  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   5  */
   6 //#define DEBUG
   7 #include <linux/netdevice.h>
   8 #include <linux/etherdevice.h>
   9 #include <linux/ethtool.h>
  10 #include <linux/module.h>
  11 #include <linux/virtio.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/bpf.h>
  14 #include <linux/bpf_trace.h>
  15 #include <linux/scatterlist.h>
  16 #include <linux/if_vlan.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/average.h>
  20 #include <linux/filter.h>
  21 #include <linux/kernel.h>
  22 #include <net/route.h>
  23 #include <net/xdp.h>
  24 #include <net/net_failover.h>
  25
  26 static int napi_weight = NAPI_POLL_WEIGHT;
  27 module_param(napi_weight, int, 0444);
  28
  29 static bool csum = true, gso = true, napi_tx = true;
  30 module_param(csum, bool, 0444);
  31 module_param(gso, bool, 0444);
  32 module_param(napi_tx, bool, 0644);
  33
  34 /* FIXME: MTU in config. */
  35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  36 #define GOOD_COPY_LEN   128
  37
  38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  39
  40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  41 #define VIRTIO_XDP_HEADROOM 256
  42
  43 /* Separating two types of XDP xmit */
  44 #define VIRTIO_XDP_TX           BIT(0)
  45 #define VIRTIO_XDP_REDIR        BIT(1)
  46
  47 #define VIRTIO_XDP_FLAG BIT(0)
  48
  49 /* RX packet size EWMA. The average packet size is used to determine the packet
  50  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  51  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  52  * term, transient changes in packet size.
  53  */
  54 DECLARE_EWMA(pkt_len, 0, 64)
  55
  56 #define VIRTNET_DRIVER_VERSION "1.0.0"
  57
  58 static const unsigned long guest_offloads[] = {
  59         VIRTIO_NET_F_GUEST_TSO4,
  60         VIRTIO_NET_F_GUEST_TSO6,
  61         VIRTIO_NET_F_GUEST_ECN,
  62         VIRTIO_NET_F_GUEST_UFO,
  63         VIRTIO_NET_F_GUEST_CSUM
  64 };
  65
  66 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
  67                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
  68                                 (1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
  69                                 (1ULL << VIRTIO_NET_F_GUEST_UFO))
  70
  71 struct virtnet_stat_desc {
  72         char desc[ETH_GSTRING_LEN];
  73         size_t offset;
  74 };
  75
  76 struct virtnet_sq_stats {
  77         struct u64_stats_sync syncp;
  78         u64 packets;
  79         u64 bytes;
  80         u64 xdp_tx;
  81         u64 xdp_tx_drops;
  82         u64 kicks;
  83         u64 tx_timeouts;
  84 };
  85
  86 struct virtnet_rq_stats {
  87         struct u64_stats_sync syncp;
  88         u64 packets;
  89         u64 bytes;
  90         u64 drops;
  91         u64 xdp_packets;
  92         u64 xdp_tx;
  93         u64 xdp_redirects;
  94         u64 xdp_drops;
  95         u64 kicks;
  96 };
  97
  98 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
  99 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
 100
 101 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
 102         { "packets",            VIRTNET_SQ_STAT(packets) },
 103         { "bytes",              VIRTNET_SQ_STAT(bytes) },
 104         { "xdp_tx",             VIRTNET_SQ_STAT(xdp_tx) },
 105         { "xdp_tx_drops",       VIRTNET_SQ_STAT(xdp_tx_drops) },
 106         { "kicks",              VIRTNET_SQ_STAT(kicks) },
 107         { "tx_timeouts",        VIRTNET_SQ_STAT(tx_timeouts) },
 108 };
 109
 110 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
 111         { "packets",            VIRTNET_RQ_STAT(packets) },
 112         { "bytes",              VIRTNET_RQ_STAT(bytes) },
 113         { "drops",              VIRTNET_RQ_STAT(drops) },
 114         { "xdp_packets",        VIRTNET_RQ_STAT(xdp_packets) },
 115         { "xdp_tx",             VIRTNET_RQ_STAT(xdp_tx) },
 116         { "xdp_redirects",      VIRTNET_RQ_STAT(xdp_redirects) },
 117         { "xdp_drops",          VIRTNET_RQ_STAT(xdp_drops) },
 118         { "kicks",              VIRTNET_RQ_STAT(kicks) },
 119 };
 120
 121 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 122 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 123
 124 /* Internal representation of a send virtqueue */
 125 struct send_queue {
 126         /* Virtqueue associated with this send _queue */
 127         struct virtqueue *vq;
 128
 129         /* TX: fragments + linear part + virtio header */
 130         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 131
 132         /* Name of the send queue: output.$index */
 133         char name[40];
 134
 135         struct virtnet_sq_stats stats;
 136
 137         struct napi_struct napi;
 138 };
 139
 140 /* Internal representation of a receive virtqueue */
 141 struct receive_queue {
 142         /* Virtqueue associated with this receive_queue */
 143         struct virtqueue *vq;
 144
 145         struct napi_struct napi;
 146
 147         struct bpf_prog __rcu *xdp_prog;
 148
 149         struct virtnet_rq_stats stats;
 150
 151         /* Chain pages by the private ptr. */
 152         struct page *pages;
 153
 154         /* Average packet length for mergeable receive buffers. */
 155         struct ewma_pkt_len mrg_avg_pkt_len;
 156
 157         /* Page frag for packet buffer allocation. */
 158         struct page_frag alloc_frag;
 159
 160         /* RX: fragments + linear part + virtio header */
 161         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 162
 163         /* Min single buffer size for mergeable buffers case. */
 164         unsigned int min_buf_len;
 165
 166         /* Name of this receive queue: input.$index */
 167         char name[40];
 168
 169         struct xdp_rxq_info xdp_rxq;
 170 };
 171
 172 /* Control VQ buffers: protected by the rtnl lock */
 173 struct control_buf {
 174         struct virtio_net_ctrl_hdr hdr;
 175         virtio_net_ctrl_ack status;
 176         struct virtio_net_ctrl_mq mq;
 177         u8 promisc;
 178         u8 allmulti;
 179         __virtio16 vid;
 180         __virtio64 offloads;
 181 };
 182
 183 struct virtnet_info {
 184         struct virtio_device *vdev;
 185         struct virtqueue *cvq;
 186         struct net_device *dev;
 187         struct send_queue *sq;
 188         struct receive_queue *rq;
 189         unsigned int status;
 190
 191         /* Max # of queue pairs supported by the device */
 192         u16 max_queue_pairs;
 193
 194         /* # of queue pairs currently used by the driver */
 195         u16 curr_queue_pairs;
 196
 197         /* # of XDP queue pairs currently used by the driver */
 198         u16 xdp_queue_pairs;
 199
 200         /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
 201         bool xdp_enabled;
 202
 203         /* I like... big packets and I cannot lie! */
 204         bool big_packets;
 205
 206         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 207         bool mergeable_rx_bufs;
 208
 209         /* Has control virtqueue */
 210         bool has_cvq;
 211
 212         /* Host can handle any s/g split between our header and packet data */
 213         bool any_header_sg;
 214
 215         /* Packet virtio header size */
 216         u8 hdr_len;
 217
 218         /* Work struct for refilling if we run low on memory. */
 219         struct delayed_work refill;
 220
 221         /* Work struct for config space updates */
 222         struct work_struct config_work;
 223
 224         /* Does the affinity hint is set for virtqueues? */
 225         bool affinity_hint_set;
 226
 227         /* CPU hotplug instances for online & dead */
 228         struct hlist_node node;
 229         struct hlist_node node_dead;
 230
 231         struct control_buf *ctrl;
 232
 233         /* Ethtool settings */
 234         u8 duplex;
 235         u32 speed;
 236
 237         unsigned long guest_offloads;
 238         unsigned long guest_offloads_capable;
 239
 240         /* failover when STANDBY feature enabled */
 241         struct failover *failover;
 242 };
 243
 244 struct padded_vnet_hdr {
 245         struct virtio_net_hdr_mrg_rxbuf hdr;
 246         /*
 247          * hdr is in a separate sg buffer, and data sg buffer shares same page
 248          * with this header sg. This padding makes next sg 16 byte aligned
 249          * after the header.
 250          */
 251         char padding[4];
 252 };
 253
 254 static bool is_xdp_frame(void *ptr)
 255 {
 256         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 257 }
 258
 259 static void *xdp_to_ptr(struct xdp_frame *ptr)
 260 {
 261         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 262 }
 263
 264 static struct xdp_frame *ptr_to_xdp(void *ptr)
 265 {
 266         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 267 }
 268
 269 /* Converting between virtqueue no. and kernel tx/rx queue no.
 270  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 271  */
 272 static int vq2txq(struct virtqueue *vq)
 273 {
 274         return (vq->index - 1) / 2;
 275 }
 276
 277 static int txq2vq(int txq)
 278 {
 279         return txq * 2 + 1;
 280 }
 281
 282 static int vq2rxq(struct virtqueue *vq)
 283 {
 284         return vq->index / 2;
 285 }
 286
 287 static int rxq2vq(int rxq)
 288 {
 289         return rxq * 2;
 290 }
 291
 292 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 293 {
 294         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 295 }
 296
 297 /*
 298  * private is used to chain pages for big packets, put the whole
 299  * most recent used list in the beginning for reuse
 300  */
 301 static void give_pages(struct receive_queue *rq, struct page *page)
 302 {
 303         struct page *end;
 304
 305         /* Find end of list, sew whole thing into vi->rq.pages. */
 306         for (end = page; end->private; end = (struct page *)end->private);
 307         end->private = (unsigned long)rq->pages;
 308         rq->pages = page;
 309 }
 310
 311 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 312 {
 313         struct page *p = rq->pages;
 314
 315         if (p) {
 316                 rq->pages = (struct page *)p->private;
 317                 /* clear private here, it is used to chain pages */
 318                 p->private = 0;
 319         } else
 320                 p = alloc_page(gfp_mask);
 321         return p;
 322 }
 323
 324 static void virtqueue_napi_schedule(struct napi_struct *napi,
 325                                     struct virtqueue *vq)
 326 {
 327         if (napi_schedule_prep(napi)) {
 328                 virtqueue_disable_cb(vq);
 329                 __napi_schedule(napi);
 330         }
 331 }
 332
 333 static void virtqueue_napi_complete(struct napi_struct *napi,
 334                                     struct virtqueue *vq, int processed)
 335 {
 336         int opaque;
 337
 338         opaque = virtqueue_enable_cb_prepare(vq);
 339         if (napi_complete_done(napi, processed)) {
 340                 if (unlikely(virtqueue_poll(vq, opaque)))
 341                         virtqueue_napi_schedule(napi, vq);
 342         } else {
 343                 virtqueue_disable_cb(vq);
 344         }
 345 }
 346
 347 static void skb_xmit_done(struct virtqueue *vq)
 348 {
 349         struct virtnet_info *vi = vq->vdev->priv;
 350         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 351
 352         /* Suppress further interrupts. */
 353         virtqueue_disable_cb(vq);
 354
 355         if (napi->weight)
 356                 virtqueue_napi_schedule(napi, vq);
 357         else
 358                 /* We were probably waiting for more output buffers. */
 359                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 360 }
 361
 362 #define MRG_CTX_HEADER_SHIFT 22
 363 static void *mergeable_len_to_ctx(unsigned int truesize,
 364                                   unsigned int headroom)
 365 {
 366         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 367 }
 368
 369 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 370 {
 371         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 372 }
 373
 374 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 375 {
 376         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 377 }
 378
 379 /* Called from bottom half context */
 380 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 381                                    struct receive_queue *rq,
 382                                    struct page *page, unsigned int offset,
 383                                    unsigned int len, unsigned int truesize,
 384                                    bool hdr_valid, unsigned int metasize,
 385                                    unsigned int headroom)
 386 {
 387         struct sk_buff *skb;
 388         struct virtio_net_hdr_mrg_rxbuf *hdr;
 389         unsigned int copy, hdr_len, hdr_padded_len;
 390         struct page *page_to_free = NULL;
 391         int tailroom, shinfo_size;
 392         char *p, *hdr_p, *buf;
 393
 394         p = page_address(page) + offset;
 395         hdr_p = p;
 396
 397         hdr_len = vi->hdr_len;
 398         if (vi->mergeable_rx_bufs)
 399                 hdr_padded_len = sizeof(*hdr);
 400         else
 401                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 402
 403         /* If headroom is not 0, there is an offset between the beginning of the
 404          * data and the allocated space, otherwise the data and the allocated
 405          * space are aligned.
 406          *
 407          * Buffers with headroom use PAGE_SIZE as alloc size, see
 408          * add_recvbuf_mergeable() + get_mergeable_buf_len()
 409          */
 410         truesize = headroom ? PAGE_SIZE : truesize;
 411         tailroom = truesize - headroom;
 412         buf = p - headroom;
 413
 414         len -= hdr_len;
 415         offset += hdr_padded_len;
 416         p += hdr_padded_len;
 417         tailroom -= hdr_padded_len + len;
 418
 419         shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 420
 421         /* copy small packet so we can reuse these pages */
 422         if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
 423                 skb = build_skb(buf, truesize);
 424                 if (unlikely(!skb))
 425                         return NULL;
 426
 427                 skb_reserve(skb, p - buf);
 428                 skb_put(skb, len);
 429
 430                 page = (struct page *)page->private;
 431                 if (page)
 432                         give_pages(rq, page);
 433                 goto ok;
 434         }
 435
 436         /* copy small packet so we can reuse these pages for small data */
 437         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 438         if (unlikely(!skb))
 439                 return NULL;
 440
 441         /* Copy all frame if it fits skb->head, otherwise
 442          * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
 443          */
 444         if (len <= skb_tailroom(skb))
 445                 copy = len;
 446         else
 447                 copy = ETH_HLEN + metasize;
 448         skb_put_data(skb, p, copy);
 449
 450         len -= copy;
 451         offset += copy;
 452
 453         if (vi->mergeable_rx_bufs) {
 454                 if (len)
 455                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 456                 else
 457                         page_to_free = page;
 458                 goto ok;
 459         }
 460
 461         /*
 462          * Verify that we can indeed put this data into a skb.
 463          * This is here to handle cases when the device erroneously
 464          * tries to receive more than is possible. This is usually
 465          * the case of a broken device.
 466          */
 467         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 468                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 469                 dev_kfree_skb(skb);
 470                 return NULL;
 471         }
 472         BUG_ON(offset >= PAGE_SIZE);
 473         while (len) {
 474                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 475                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 476                                 frag_size, truesize);
 477                 len -= frag_size;
 478                 page = (struct page *)page->private;
 479                 offset = 0;
 480         }
 481
 482         if (page)
 483                 give_pages(rq, page);
 484
 485 ok:
 486         /* hdr_valid means no XDP, so we can copy the vnet header */
 487         if (hdr_valid) {
 488                 hdr = skb_vnet_hdr(skb);
 489                 memcpy(hdr, hdr_p, hdr_len);
 490         }
 491         if (page_to_free)
 492                 put_page(page_to_free);
 493
 494         if (metasize) {
 495                 __skb_pull(skb, metasize);
 496                 skb_metadata_set(skb, metasize);
 497         }
 498
 499         return skb;
 500 }
 501
 502 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 503                                    struct send_queue *sq,
 504                                    struct xdp_frame *xdpf)
 505 {
 506         struct virtio_net_hdr_mrg_rxbuf *hdr;
 507         int err;
 508
 509         if (unlikely(xdpf->headroom < vi->hdr_len))
 510                 return -EOVERFLOW;
 511
 512         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 513         xdpf->data -= vi->hdr_len;
 514         /* Zero header and leave csum up to XDP layers */
 515         hdr = xdpf->data;
 516         memset(hdr, 0, vi->hdr_len);
 517         xdpf->len   += vi->hdr_len;
 518
 519         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 520
 521         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
 522                                    GFP_ATOMIC);
 523         if (unlikely(err))
 524                 return -ENOSPC; /* Caller handle free/refcnt */
 525
 526         return 0;
 527 }
 528
 529 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
 530  * the current cpu, so it does not need to be locked.
 531  *
 532  * Here we use marco instead of inline functions because we have to deal with
 533  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
 534  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
 535  * functions to perfectly solve these three problems at the same time.
 536  */
 537 #define virtnet_xdp_get_sq(vi) ({                                       \
 538         int cpu = smp_processor_id();                                   \
 539         struct netdev_queue *txq;                                       \
 540         typeof(vi) v = (vi);                                            \
 541         unsigned int qp;                                                \
 542                                                                         \
 543         if (v->curr_queue_pairs > nr_cpu_ids) {                         \
 544                 qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
 545                 qp += cpu;                                              \
 546                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 547                 __netif_tx_acquire(txq);                                \
 548         } else {                                                        \
 549                 qp = cpu % v->curr_queue_pairs;                         \
 550                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 551                 __netif_tx_lock(txq, cpu);                              \
 552         }                                                               \
 553         v->sq + qp;                                                     \
 554 })
 555
 556 #define virtnet_xdp_put_sq(vi, q) {                                     \
 557         struct netdev_queue *txq;                                       \
 558         typeof(vi) v = (vi);                                            \
 559                                                                         \
 560         txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
 561         if (v->curr_queue_pairs > nr_cpu_ids)                           \
 562                 __netif_tx_release(txq);                                \
 563         else                                                            \
 564                 __netif_tx_unlock(txq);                                 \
 565 }
 566
 567 static int virtnet_xdp_xmit(struct net_device *dev,
 568                             int n, struct xdp_frame **frames, u32 flags)
 569 {
 570         struct virtnet_info *vi = netdev_priv(dev);
 571         struct receive_queue *rq = vi->rq;
 572         struct bpf_prog *xdp_prog;
 573         struct send_queue *sq;
 574         unsigned int len;
 575         int packets = 0;
 576         int bytes = 0;
 577         int nxmit = 0;
 578         int kicks = 0;
 579         void *ptr;
 580         int ret;
 581         int i;
 582
 583         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 584          * indicate XDP resources have been successfully allocated.
 585          */
 586         xdp_prog = rcu_access_pointer(rq->xdp_prog);
 587         if (!xdp_prog)
 588                 return -ENXIO;
 589
 590         sq = virtnet_xdp_get_sq(vi);
 591
 592         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 593                 ret = -EINVAL;
 594                 goto out;
 595         }
 596
 597         /* Free up any pending old buffers before queueing new ones. */
 598         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 599                 if (likely(is_xdp_frame(ptr))) {
 600                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 601
 602                         bytes += frame->len;
 603                         xdp_return_frame(frame);
 604                 } else {
 605                         struct sk_buff *skb = ptr;
 606
 607                         bytes += skb->len;
 608                         napi_consume_skb(skb, false);
 609                 }
 610                 packets++;
 611         }
 612
 613         for (i = 0; i < n; i++) {
 614                 struct xdp_frame *xdpf = frames[i];
 615
 616                 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
 617                         break;
 618                 nxmit++;
 619         }
 620         ret = nxmit;
 621
 622         if (flags & XDP_XMIT_FLUSH) {
 623                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
 624                         kicks = 1;
 625         }
 626 out:
 627         u64_stats_update_begin(&sq->stats.syncp);
 628         sq->stats.bytes += bytes;
 629         sq->stats.packets += packets;
 630         sq->stats.xdp_tx += n;
 631         sq->stats.xdp_tx_drops += n - nxmit;
 632         sq->stats.kicks += kicks;
 633         u64_stats_update_end(&sq->stats.syncp);
 634
 635         virtnet_xdp_put_sq(vi, sq);
 636         return ret;
 637 }
 638
 639 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 640 {
 641         return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
 642 }
 643
 644 /* We copy the packet for XDP in the following cases:
 645  *
 646  * 1) Packet is scattered across multiple rx buffers.
 647  * 2) Headroom space is insufficient.
 648  *
 649  * This is inefficient but it's a temporary condition that
 650  * we hit right after XDP is enabled and until queue is refilled
 651  * with large buffers with sufficient headroom - so it should affect
 652  * at most queue size packets.
 653  * Afterwards, the conditions to enable
 654  * XDP should preclude the underlying device from sending packets
 655  * across multiple buffers (num_buf > 1), and we make sure buffers
 656  * have enough headroom.
 657  */
 658 static struct page *xdp_linearize_page(struct receive_queue *rq,
 659                                        u16 *num_buf,
 660                                        struct page *p,
 661                                        int offset,
 662                                        int page_off,
 663                                        unsigned int *len)
 664 {
 665         struct page *page = alloc_page(GFP_ATOMIC);
 666
 667         if (!page)
 668                 return NULL;
 669
 670         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 671         page_off += *len;
 672
 673         while (--*num_buf) {
 674                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 675                 unsigned int buflen;
 676                 void *buf;
 677                 int off;
 678
 679                 buf = virtqueue_get_buf(rq->vq, &buflen);
 680                 if (unlikely(!buf))
 681                         goto err_buf;
 682
 683                 p = virt_to_head_page(buf);
 684                 off = buf - page_address(p);
 685
 686                 /* guard against a misconfigured or uncooperative backend that
 687                  * is sending packet larger than the MTU.
 688                  */
 689                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 690                         put_page(p);
 691                         goto err_buf;
 692                 }
 693
 694                 memcpy(page_address(page) + page_off,
 695                        page_address(p) + off, buflen);
 696                 page_off += buflen;
 697                 put_page(p);
 698         }
 699
 700         /* Headroom does not contribute to packet length */
 701         *len = page_off - VIRTIO_XDP_HEADROOM;
 702         return page;
 703 err_buf:
 704         __free_pages(page, 0);
 705         return NULL;
 706 }
 707
 708 static struct sk_buff *receive_small(struct net_device *dev,
 709                                      struct virtnet_info *vi,
 710                                      struct receive_queue *rq,
 711                                      void *buf, void *ctx,
 712                                      unsigned int len,
 713                                      unsigned int *xdp_xmit,
 714                                      struct virtnet_rq_stats *stats)
 715 {
 716         struct sk_buff *skb;
 717         struct bpf_prog *xdp_prog;
 718         unsigned int xdp_headroom = (unsigned long)ctx;
 719         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 720         unsigned int headroom = vi->hdr_len + header_offset;
 721         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 722                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 723         struct page *page = virt_to_head_page(buf);
 724         unsigned int delta = 0;
 725         struct page *xdp_page;
 726         int err;
 727         unsigned int metasize = 0;
 728
 729         len -= vi->hdr_len;
 730         stats->bytes += len;
 731
 732         if (unlikely(len > GOOD_PACKET_LEN)) {
 733                 pr_debug("%s: rx error: len %u exceeds max size %d\n",
 734                          dev->name, len, GOOD_PACKET_LEN);
 735                 dev->stats.rx_length_errors++;
 736                 goto err;
 737         }
 738
 739         if (likely(!vi->xdp_enabled)) {
 740                 xdp_prog = NULL;
 741                 goto skip_xdp;
 742         }
 743
 744         rcu_read_lock();
 745         xdp_prog = rcu_dereference(rq->xdp_prog);
 746         if (xdp_prog) {
 747                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 748                 struct xdp_frame *xdpf;
 749                 struct xdp_buff xdp;
 750                 void *orig_data;
 751                 u32 act;
 752
 753                 if (unlikely(hdr->hdr.gso_type))
 754                         goto err_xdp;
 755
 756                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 757                         int offset = buf - page_address(page) + header_offset;
 758                         unsigned int tlen = len + vi->hdr_len;
 759                         u16 num_buf = 1;
 760
 761                         xdp_headroom = virtnet_get_headroom(vi);
 762                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 763                         headroom = vi->hdr_len + header_offset;
 764                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 765                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 766                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 767                                                       offset, header_offset,
 768                                                       &tlen);
 769                         if (!xdp_page)
 770                                 goto err_xdp;
 771
 772                         buf = page_address(xdp_page);
 773                         put_page(page);
 774                         page = xdp_page;
 775                 }
 776
 777                 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
 778                 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
 779                                  xdp_headroom, len, true);
 780                 orig_data = xdp.data;
 781                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 782                 stats->xdp_packets++;
 783
 784                 switch (act) {
 785                 case XDP_PASS:
 786                         /* Recalculate length in case bpf program changed it */
 787                         delta = orig_data - xdp.data;
 788                         len = xdp.data_end - xdp.data;
 789                         metasize = xdp.data - xdp.data_meta;
 790                         break;
 791                 case XDP_TX:
 792                         stats->xdp_tx++;
 793                         xdpf = xdp_convert_buff_to_frame(&xdp);
 794                         if (unlikely(!xdpf))
 795                                 goto err_xdp;
 796                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 797                         if (unlikely(!err)) {
 798                                 xdp_return_frame_rx_napi(xdpf);
 799                         } else if (unlikely(err < 0)) {
 800                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 801                                 goto err_xdp;
 802                         }
 803                         *xdp_xmit |= VIRTIO_XDP_TX;
 804                         rcu_read_unlock();
 805                         goto xdp_xmit;
 806                 case XDP_REDIRECT:
 807                         stats->xdp_redirects++;
 808                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 809                         if (err)
 810                                 goto err_xdp;
 811                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 812                         rcu_read_unlock();
 813                         goto xdp_xmit;
 814                 default:
 815                         bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act);
 816                         fallthrough;
 817                 case XDP_ABORTED:
 818                         trace_xdp_exception(vi->dev, xdp_prog, act);
 819                         goto err_xdp;
 820                 case XDP_DROP:
 821                         goto err_xdp;
 822                 }
 823         }
 824         rcu_read_unlock();
 825
 826 skip_xdp:
 827         skb = build_skb(buf, buflen);
 828         if (!skb)
 829                 goto err;
 830         skb_reserve(skb, headroom - delta);
 831         skb_put(skb, len);
 832         if (!xdp_prog) {
 833                 buf += header_offset;
 834                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 835         } /* keep zeroed vnet hdr since XDP is loaded */
 836
 837         if (metasize)
 838                 skb_metadata_set(skb, metasize);
 839
 840         return skb;
 841
 842 err_xdp:
 843         rcu_read_unlock();
 844         stats->xdp_drops++;
 845 err:
 846         stats->drops++;
 847         put_page(page);
 848 xdp_xmit:
 849         return NULL;
 850 }
 851
 852 static struct sk_buff *receive_big(struct net_device *dev,
 853                                    struct virtnet_info *vi,
 854                                    struct receive_queue *rq,
 855                                    void *buf,
 856                                    unsigned int len,
 857                                    struct virtnet_rq_stats *stats)
 858 {
 859         struct page *page = buf;
 860         struct sk_buff *skb =
 861                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
 862
 863         stats->bytes += len - vi->hdr_len;
 864         if (unlikely(!skb))
 865                 goto err;
 866
 867         return skb;
 868
 869 err:
 870         stats->drops++;
 871         give_pages(rq, page);
 872         return NULL;
 873 }
 874
 875 static struct sk_buff *receive_mergeable(struct net_device *dev,
 876                                          struct virtnet_info *vi,
 877                                          struct receive_queue *rq,
 878                                          void *buf,
 879                                          void *ctx,
 880                                          unsigned int len,
 881                                          unsigned int *xdp_xmit,
 882                                          struct virtnet_rq_stats *stats)
 883 {
 884         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 885         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 886         struct page *page = virt_to_head_page(buf);
 887         int offset = buf - page_address(page);
 888         struct sk_buff *head_skb, *curr_skb;
 889         struct bpf_prog *xdp_prog;
 890         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
 891         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 892         unsigned int metasize = 0;
 893         unsigned int frame_sz;
 894         int err;
 895
 896         head_skb = NULL;
 897         stats->bytes += len - vi->hdr_len;
 898
 899         if (unlikely(len > truesize)) {
 900                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 901                          dev->name, len, (unsigned long)ctx);
 902                 dev->stats.rx_length_errors++;
 903                 goto err_skb;
 904         }
 905
 906         if (likely(!vi->xdp_enabled)) {
 907                 xdp_prog = NULL;
 908                 goto skip_xdp;
 909         }
 910
 911         rcu_read_lock();
 912         xdp_prog = rcu_dereference(rq->xdp_prog);
 913         if (xdp_prog) {
 914                 struct xdp_frame *xdpf;
 915                 struct page *xdp_page;
 916                 struct xdp_buff xdp;
 917                 void *data;
 918                 u32 act;
 919
 920                 /* Transient failure which in theory could occur if
 921                  * in-flight packets from before XDP was enabled reach
 922                  * the receive path after XDP is loaded.
 923                  */
 924                 if (unlikely(hdr->hdr.gso_type))
 925                         goto err_xdp;
 926
 927                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 928                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 929                  */
 930                 frame_sz = headroom ? PAGE_SIZE : truesize;
 931
 932                 /* This happens when rx buffer size is underestimated
 933                  * or headroom is not enough because of the buffer
 934                  * was refilled before XDP is set. This should only
 935                  * happen for the first several packets, so we don't
 936                  * care much about its performance.
 937                  */
 938                 if (unlikely(num_buf > 1 ||
 939                              headroom < virtnet_get_headroom(vi))) {
 940                         /* linearize data for XDP */
 941                         xdp_page = xdp_linearize_page(rq, &num_buf,
 942                                                       page, offset,
 943                                                       VIRTIO_XDP_HEADROOM,
 944                                                       &len);
 945                         frame_sz = PAGE_SIZE;
 946
 947                         if (!xdp_page)
 948                                 goto err_xdp;
 949                         offset = VIRTIO_XDP_HEADROOM;
 950                 } else {
 951                         xdp_page = page;
 952                 }
 953
 954                 /* Allow consuming headroom but reserve enough space to push
 955                  * the descriptor on if we get an XDP_TX return code.
 956                  */
 957                 data = page_address(xdp_page) + offset;
 958                 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
 959                 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
 960                                  VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
 961
 962                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 963                 stats->xdp_packets++;
 964
 965                 switch (act) {
 966                 case XDP_PASS:
 967                         metasize = xdp.data - xdp.data_meta;
 968
 969                         /* recalculate offset to account for any header
 970                          * adjustments and minus the metasize to copy the
 971                          * metadata in page_to_skb(). Note other cases do not
 972                          * build an skb and avoid using offset
 973                          */
 974                         offset = xdp.data - page_address(xdp_page) -
 975                                  vi->hdr_len - metasize;
 976
 977                         /* recalculate len if xdp.data, xdp.data_end or
 978                          * xdp.data_meta were adjusted
 979                          */
 980                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 981                         /* We can only create skb based on xdp_page. */
 982                         if (unlikely(xdp_page != page)) {
 983                                 rcu_read_unlock();
 984                                 put_page(page);
 985                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
 986                                                        len, PAGE_SIZE, false,
 987                                                        metasize,
 988                                                        VIRTIO_XDP_HEADROOM);
 989                                 return head_skb;
 990                         }
 991                         break;
 992                 case XDP_TX:
 993                         stats->xdp_tx++;
 994                         xdpf = xdp_convert_buff_to_frame(&xdp);
 995                         if (unlikely(!xdpf))
 996                                 goto err_xdp;
 997                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 998                         if (unlikely(!err)) {
 999                                 xdp_return_frame_rx_napi(xdpf);
1000                         } else if (unlikely(err < 0)) {
1001                                 trace_xdp_exception(vi->dev, xdp_prog, act);
1002                                 if (unlikely(xdp_page != page))
1003                                         put_page(xdp_page);
1004                                 goto err_xdp;
1005                         }
1006                         *xdp_xmit |= VIRTIO_XDP_TX;
1007                         if (unlikely(xdp_page != page))
1008                                 put_page(page);
1009                         rcu_read_unlock();
1010                         goto xdp_xmit;
1011                 case XDP_REDIRECT:
1012                         stats->xdp_redirects++;
1013                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
1014                         if (err) {
1015                                 if (unlikely(xdp_page != page))
1016                                         put_page(xdp_page);
1017                                 goto err_xdp;
1018                         }
1019                         *xdp_xmit |= VIRTIO_XDP_REDIR;
1020                         if (unlikely(xdp_page != page))
1021                                 put_page(page);
1022                         rcu_read_unlock();
1023                         goto xdp_xmit;
1024                 default:
1025                         bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act);
1026                         fallthrough;
1027                 case XDP_ABORTED:
1028                         trace_xdp_exception(vi->dev, xdp_prog, act);
1029                         fallthrough;
1030                 case XDP_DROP:
1031                         if (unlikely(xdp_page != page))
1032                                 __free_pages(xdp_page, 0);
1033                         goto err_xdp;
1034                 }
1035         }
1036         rcu_read_unlock();
1037
1038 skip_xdp:
1039         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1040                                metasize, headroom);
1041         curr_skb = head_skb;
1042
1043         if (unlikely(!curr_skb))
1044                 goto err_skb;
1045         while (--num_buf) {
1046                 int num_skb_frags;
1047
1048                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1049                 if (unlikely(!buf)) {
1050                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1051                                  dev->name, num_buf,
1052                                  virtio16_to_cpu(vi->vdev,
1053                                                  hdr->num_buffers));
1054                         dev->stats.rx_length_errors++;
1055                         goto err_buf;
1056                 }
1057
1058                 stats->bytes += len;
1059                 page = virt_to_head_page(buf);
1060
1061                 truesize = mergeable_ctx_to_truesize(ctx);
1062                 if (unlikely(len > truesize)) {
1063                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1064                                  dev->name, len, (unsigned long)ctx);
1065                         dev->stats.rx_length_errors++;
1066                         goto err_skb;
1067                 }
1068
1069                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1070                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1071                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1072
1073                         if (unlikely(!nskb))
1074                                 goto err_skb;
1075                         if (curr_skb == head_skb)
1076                                 skb_shinfo(curr_skb)->frag_list = nskb;
1077                         else
1078                                 curr_skb->next = nskb;
1079                         curr_skb = nskb;
1080                         head_skb->truesize += nskb->truesize;
1081                         num_skb_frags = 0;
1082                 }
1083                 if (curr_skb != head_skb) {
1084                         head_skb->data_len += len;
1085                         head_skb->len += len;
1086                         head_skb->truesize += truesize;
1087                 }
1088                 offset = buf - page_address(page);
1089                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1090                         put_page(page);
1091                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1092                                              len, truesize);
1093                 } else {
1094                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1095                                         offset, len, truesize);
1096                 }
1097         }
1098
1099         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1100         return head_skb;
1101
1102 err_xdp:
1103         rcu_read_unlock();
1104         stats->xdp_drops++;
1105 err_skb:
1106         put_page(page);
1107         while (num_buf-- > 1) {
1108                 buf = virtqueue_get_buf(rq->vq, &len);
1109                 if (unlikely(!buf)) {
1110                         pr_debug("%s: rx error: %d buffers missing\n",
1111                                  dev->name, num_buf);
1112                         dev->stats.rx_length_errors++;
1113                         break;
1114                 }
1115                 stats->bytes += len;
1116                 page = virt_to_head_page(buf);
1117                 put_page(page);
1118         }
1119 err_buf:
1120         stats->drops++;
1121         dev_kfree_skb(head_skb);
1122 xdp_xmit:
1123         return NULL;
1124 }
1125
1126 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1127                         void *buf, unsigned int len, void **ctx,
1128                         unsigned int *xdp_xmit,
1129                         struct virtnet_rq_stats *stats)
1130 {
1131         struct net_device *dev = vi->dev;
1132         struct sk_buff *skb;
1133         struct virtio_net_hdr_mrg_rxbuf *hdr;
1134
1135         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1136                 pr_debug("%s: short packet %i\n", dev->name, len);
1137                 dev->stats.rx_length_errors++;
1138                 if (vi->mergeable_rx_bufs) {
1139                         put_page(virt_to_head_page(buf));
1140                 } else if (vi->big_packets) {
1141                         give_pages(rq, buf);
1142                 } else {
1143                         put_page(virt_to_head_page(buf));
1144                 }
1145                 return;
1146         }
1147
1148         if (vi->mergeable_rx_bufs)
1149                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1150                                         stats);
1151         else if (vi->big_packets)
1152                 skb = receive_big(dev, vi, rq, buf, len, stats);
1153         else
1154                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1155
1156         if (unlikely(!skb))
1157                 return;
1158
1159         hdr = skb_vnet_hdr(skb);
1160
1161         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1162                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1163
1164         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1165                                   virtio_is_little_endian(vi->vdev))) {
1166                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1167                                      dev->name, hdr->hdr.gso_type,
1168                                      hdr->hdr.gso_size);
1169                 goto frame_err;
1170         }
1171
1172         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1173         skb->protocol = eth_type_trans(skb, dev);
1174         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1175                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1176
1177         napi_gro_receive(&rq->napi, skb);
1178         return;
1179
1180 frame_err:
1181         dev->stats.rx_frame_errors++;
1182         dev_kfree_skb(skb);
1183 }
1184
1185 /* Unlike mergeable buffers, all buffers are allocated to the
1186  * same size, except for the headroom. For this reason we do
1187  * not need to use  mergeable_len_to_ctx here - it is enough
1188  * to store the headroom as the context ignoring the truesize.
1189  */
1190 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1191                              gfp_t gfp)
1192 {
1193         struct page_frag *alloc_frag = &rq->alloc_frag;
1194         char *buf;
1195         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1196         void *ctx = (void *)(unsigned long)xdp_headroom;
1197         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1198         int err;
1199
1200         len = SKB_DATA_ALIGN(len) +
1201               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1202         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1203                 return -ENOMEM;
1204
1205         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1206         get_page(alloc_frag->page);
1207         alloc_frag->offset += len;
1208         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1209                     vi->hdr_len + GOOD_PACKET_LEN);
1210         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1211         if (err < 0)
1212                 put_page(virt_to_head_page(buf));
1213         return err;
1214 }
1215
1216 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1217                            gfp_t gfp)
1218 {
1219         struct page *first, *list = NULL;
1220         char *p;
1221         int i, err, offset;
1222
1223         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1224
1225         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1226         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1227                 first = get_a_page(rq, gfp);
1228                 if (!first) {
1229                         if (list)
1230                                 give_pages(rq, list);
1231                         return -ENOMEM;
1232                 }
1233                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1234
1235                 /* chain new page in list head to match sg */
1236                 first->private = (unsigned long)list;
1237                 list = first;
1238         }
1239
1240         first = get_a_page(rq, gfp);
1241         if (!first) {
1242                 give_pages(rq, list);
1243                 return -ENOMEM;
1244         }
1245         p = page_address(first);
1246
1247         /* rq->sg[0], rq->sg[1] share the same page */
1248         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1249         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1250
1251         /* rq->sg[1] for data packet, from offset */
1252         offset = sizeof(struct padded_vnet_hdr);
1253         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1254
1255         /* chain first in list head */
1256         first->private = (unsigned long)list;
1257         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1258                                   first, gfp);
1259         if (err < 0)
1260                 give_pages(rq, first);
1261
1262         return err;
1263 }
1264
1265 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1266                                           struct ewma_pkt_len *avg_pkt_len,
1267                                           unsigned int room)
1268 {
1269         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1270         unsigned int len;
1271
1272         if (room)
1273                 return PAGE_SIZE - room;
1274
1275         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1276                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1277
1278         return ALIGN(len, L1_CACHE_BYTES);
1279 }
1280
1281 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1282                                  struct receive_queue *rq, gfp_t gfp)
1283 {
1284         struct page_frag *alloc_frag = &rq->alloc_frag;
1285         unsigned int headroom = virtnet_get_headroom(vi);
1286         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1287         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1288         char *buf;
1289         void *ctx;
1290         int err;
1291         unsigned int len, hole;
1292
1293         /* Extra tailroom is needed to satisfy XDP's assumption. This
1294          * means rx frags coalescing won't work, but consider we've
1295          * disabled GSO for XDP, it won't be a big issue.
1296          */
1297         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1298         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1299                 return -ENOMEM;
1300
1301         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1302         buf += headroom; /* advance address leaving hole at front of pkt */
1303         get_page(alloc_frag->page);
1304         alloc_frag->offset += len + room;
1305         hole = alloc_frag->size - alloc_frag->offset;
1306         if (hole < len + room) {
1307                 /* To avoid internal fragmentation, if there is very likely not
1308                  * enough space for another buffer, add the remaining space to
1309                  * the current buffer.
1310                  */
1311                 len += hole;
1312                 alloc_frag->offset += hole;
1313         }
1314
1315         sg_init_one(rq->sg, buf, len);
1316         ctx = mergeable_len_to_ctx(len, headroom);
1317         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1318         if (err < 0)
1319                 put_page(virt_to_head_page(buf));
1320
1321         return err;
1322 }
1323
1324 /*
1325  * Returns false if we couldn't fill entirely (OOM).
1326  *
1327  * Normally run in the receive path, but can also be run from ndo_open
1328  * before we're receiving packets, or from refill_work which is
1329  * careful to disable receiving (using napi_disable).
1330  */
1331 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1332                           gfp_t gfp)
1333 {
1334         int err;
1335         bool oom;
1336
1337         do {
1338                 if (vi->mergeable_rx_bufs)
1339                         err = add_recvbuf_mergeable(vi, rq, gfp);
1340                 else if (vi->big_packets)
1341                         err = add_recvbuf_big(vi, rq, gfp);
1342                 else
1343                         err = add_recvbuf_small(vi, rq, gfp);
1344
1345                 oom = err == -ENOMEM;
1346                 if (err)
1347                         break;
1348         } while (rq->vq->num_free);
1349         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1350                 unsigned long flags;
1351
1352                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1353                 rq->stats.kicks++;
1354                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1355         }
1356
1357         return !oom;
1358 }
1359
1360 static void skb_recv_done(struct virtqueue *rvq)
1361 {
1362         struct virtnet_info *vi = rvq->vdev->priv;
1363         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1364
1365         virtqueue_napi_schedule(&rq->napi, rvq);
1366 }
1367
1368 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1369 {
1370         napi_enable(napi);
1371
1372         /* If all buffers were filled by other side before we napi_enabled, we
1373          * won't get another interrupt, so process any outstanding packets now.
1374          * Call local_bh_enable after to trigger softIRQ processing.
1375          */
1376         local_bh_disable();
1377         virtqueue_napi_schedule(napi, vq);
1378         local_bh_enable();
1379 }
1380
1381 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1382                                    struct virtqueue *vq,
1383                                    struct napi_struct *napi)
1384 {
1385         if (!napi->weight)
1386                 return;
1387
1388         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1389          * enable the feature if this is likely affine with the transmit path.
1390          */
1391         if (!vi->affinity_hint_set) {
1392                 napi->weight = 0;
1393                 return;
1394         }
1395
1396         return virtnet_napi_enable(vq, napi);
1397 }
1398
1399 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1400 {
1401         if (napi->weight)
1402                 napi_disable(napi);
1403 }
1404
1405 static void refill_work(struct work_struct *work)
1406 {
1407         struct virtnet_info *vi =
1408                 container_of(work, struct virtnet_info, refill.work);
1409         bool still_empty;
1410         int i;
1411
1412         for (i = 0; i < vi->curr_queue_pairs; i++) {
1413                 struct receive_queue *rq = &vi->rq[i];
1414
1415                 napi_disable(&rq->napi);
1416                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1417                 virtnet_napi_enable(rq->vq, &rq->napi);
1418
1419                 /* In theory, this can happen: if we don't get any buffers in
1420                  * we will *never* try to fill again.
1421                  */
1422                 if (still_empty)
1423                         schedule_delayed_work(&vi->refill, HZ/2);
1424         }
1425 }
1426
1427 static int virtnet_receive(struct receive_queue *rq, int budget,
1428                            unsigned int *xdp_xmit)
1429 {
1430         struct virtnet_info *vi = rq->vq->vdev->priv;
1431         struct virtnet_rq_stats stats = {};
1432         unsigned int len;
1433         void *buf;
1434         int i;
1435
1436         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1437                 void *ctx;
1438
1439                 while (stats.packets < budget &&
1440                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1441                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1442                         stats.packets++;
1443                 }
1444         } else {
1445                 while (stats.packets < budget &&
1446                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1447                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1448                         stats.packets++;
1449                 }
1450         }
1451
1452         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1453                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1454                         schedule_delayed_work(&vi->refill, 0);
1455         }
1456
1457         u64_stats_update_begin(&rq->stats.syncp);
1458         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1459                 size_t offset = virtnet_rq_stats_desc[i].offset;
1460                 u64 *item;
1461
1462                 item = (u64 *)((u8 *)&rq->stats + offset);
1463                 *item += *(u64 *)((u8 *)&stats + offset);
1464         }
1465         u64_stats_update_end(&rq->stats.syncp);
1466
1467         return stats.packets;
1468 }
1469
1470 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1471 {
1472         unsigned int len;
1473         unsigned int packets = 0;
1474         unsigned int bytes = 0;
1475         void *ptr;
1476
1477         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1478                 if (likely(!is_xdp_frame(ptr))) {
1479                         struct sk_buff *skb = ptr;
1480
1481                         pr_debug("Sent skb %p\n", skb);
1482
1483                         bytes += skb->len;
1484                         napi_consume_skb(skb, in_napi);
1485                 } else {
1486                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1487
1488                         bytes += frame->len;
1489                         xdp_return_frame(frame);
1490                 }
1491                 packets++;
1492         }
1493
1494         /* Avoid overhead when no packets have been processed
1495          * happens when called speculatively from start_xmit.
1496          */
1497         if (!packets)
1498                 return;
1499
1500         u64_stats_update_begin(&sq->stats.syncp);
1501         sq->stats.bytes += bytes;
1502         sq->stats.packets += packets;
1503         u64_stats_update_end(&sq->stats.syncp);
1504 }
1505
1506 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1507 {
1508         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1509                 return false;
1510         else if (q < vi->curr_queue_pairs)
1511                 return true;
1512         else
1513                 return false;
1514 }
1515
1516 static void virtnet_poll_cleantx(struct receive_queue *rq)
1517 {
1518         struct virtnet_info *vi = rq->vq->vdev->priv;
1519         unsigned int index = vq2rxq(rq->vq);
1520         struct send_queue *sq = &vi->sq[index];
1521         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1522
1523         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1524                 return;
1525
1526         if (__netif_tx_trylock(txq)) {
1527                 do {
1528                         virtqueue_disable_cb(sq->vq);
1529                         free_old_xmit_skbs(sq, true);
1530                 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1531
1532                 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1533                         netif_tx_wake_queue(txq);
1534
1535                 __netif_tx_unlock(txq);
1536         }
1537 }
1538
1539 static int virtnet_poll(struct napi_struct *napi, int budget)
1540 {
1541         struct receive_queue *rq =
1542                 container_of(napi, struct receive_queue, napi);
1543         struct virtnet_info *vi = rq->vq->vdev->priv;
1544         struct send_queue *sq;
1545         unsigned int received;
1546         unsigned int xdp_xmit = 0;
1547
1548         virtnet_poll_cleantx(rq);
1549
1550         received = virtnet_receive(rq, budget, &xdp_xmit);
1551
1552         /* Out of packets? */
1553         if (received < budget)
1554                 virtqueue_napi_complete(napi, rq->vq, received);
1555
1556         if (xdp_xmit & VIRTIO_XDP_REDIR)
1557                 xdp_do_flush();
1558
1559         if (xdp_xmit & VIRTIO_XDP_TX) {
1560                 sq = virtnet_xdp_get_sq(vi);
1561                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1562                         u64_stats_update_begin(&sq->stats.syncp);
1563                         sq->stats.kicks++;
1564                         u64_stats_update_end(&sq->stats.syncp);
1565                 }
1566                 virtnet_xdp_put_sq(vi, sq);
1567         }
1568
1569         return received;
1570 }
1571
1572 static int virtnet_open(struct net_device *dev)
1573 {
1574         struct virtnet_info *vi = netdev_priv(dev);
1575         int i, err;
1576
1577         for (i = 0; i < vi->max_queue_pairs; i++) {
1578                 if (i < vi->curr_queue_pairs)
1579                         /* Make sure we have some buffers: if oom use wq. */
1580                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1581                                 schedule_delayed_work(&vi->refill, 0);
1582
1583                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1584                 if (err < 0)
1585                         return err;
1586
1587                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1588                                                  MEM_TYPE_PAGE_SHARED, NULL);
1589                 if (err < 0) {
1590                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1591                         return err;
1592                 }
1593
1594                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1595                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1596         }
1597
1598         return 0;
1599 }
1600
1601 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1602 {
1603         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1604         struct virtnet_info *vi = sq->vq->vdev->priv;
1605         unsigned int index = vq2txq(sq->vq);
1606         struct netdev_queue *txq;
1607         int opaque;
1608         bool done;
1609
1610         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1611                 /* We don't need to enable cb for XDP */
1612                 napi_complete_done(napi, 0);
1613                 return 0;
1614         }
1615
1616         txq = netdev_get_tx_queue(vi->dev, index);
1617         __netif_tx_lock(txq, raw_smp_processor_id());
1618         virtqueue_disable_cb(sq->vq);
1619         free_old_xmit_skbs(sq, true);
1620
1621         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1622                 netif_tx_wake_queue(txq);
1623
1624         opaque = virtqueue_enable_cb_prepare(sq->vq);
1625
1626         done = napi_complete_done(napi, 0);
1627
1628         if (!done)
1629                 virtqueue_disable_cb(sq->vq);
1630
1631         __netif_tx_unlock(txq);
1632
1633         if (done) {
1634                 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1635                         if (napi_schedule_prep(napi)) {
1636                                 __netif_tx_lock(txq, raw_smp_processor_id());
1637                                 virtqueue_disable_cb(sq->vq);
1638                                 __netif_tx_unlock(txq);
1639                                 __napi_schedule(napi);
1640                         }
1641                 }
1642         }
1643
1644         return 0;
1645 }
1646
1647 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1648 {
1649         struct virtio_net_hdr_mrg_rxbuf *hdr;
1650         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1651         struct virtnet_info *vi = sq->vq->vdev->priv;
1652         int num_sg;
1653         unsigned hdr_len = vi->hdr_len;
1654         bool can_push;
1655
1656         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1657
1658         can_push = vi->any_header_sg &&
1659                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1660                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1661         /* Even if we can, don't push here yet as this would skew
1662          * csum_start offset below. */
1663         if (can_push)
1664                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1665         else
1666                 hdr = skb_vnet_hdr(skb);
1667
1668         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1669                                     virtio_is_little_endian(vi->vdev), false,
1670                                     0))
1671                 return -EPROTO;
1672
1673         if (vi->mergeable_rx_bufs)
1674                 hdr->num_buffers = 0;
1675
1676         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1677         if (can_push) {
1678                 __skb_push(skb, hdr_len);
1679                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1680                 if (unlikely(num_sg < 0))
1681                         return num_sg;
1682                 /* Pull header back to avoid skew in tx bytes calculations. */
1683                 __skb_pull(skb, hdr_len);
1684         } else {
1685                 sg_set_buf(sq->sg, hdr, hdr_len);
1686                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1687                 if (unlikely(num_sg < 0))
1688                         return num_sg;
1689                 num_sg++;
1690         }
1691         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1692 }
1693
1694 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1695 {
1696         struct virtnet_info *vi = netdev_priv(dev);
1697         int qnum = skb_get_queue_mapping(skb);
1698         struct send_queue *sq = &vi->sq[qnum];
1699         int err;
1700         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1701         bool kick = !netdev_xmit_more();
1702         bool use_napi = sq->napi.weight;
1703
1704         /* Free up any pending old buffers before queueing new ones. */
1705         do {
1706                 if (use_napi)
1707                         virtqueue_disable_cb(sq->vq);
1708
1709                 free_old_xmit_skbs(sq, false);
1710
1711         } while (use_napi && kick &&
1712                unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1713
1714         /* timestamp packet in software */
1715         skb_tx_timestamp(skb);
1716
1717         /* Try to transmit */
1718         err = xmit_skb(sq, skb);
1719
1720         /* This should not happen! */
1721         if (unlikely(err)) {
1722                 dev->stats.tx_fifo_errors++;
1723                 if (net_ratelimit())
1724                         dev_warn(&dev->dev,
1725                                  "Unexpected TXQ (%d) queue failure: %d\n",
1726                                  qnum, err);
1727                 dev->stats.tx_dropped++;
1728                 dev_kfree_skb_any(skb);
1729                 return NETDEV_TX_OK;
1730         }
1731
1732         /* Don't wait up for transmitted skbs to be freed. */
1733         if (!use_napi) {
1734                 skb_orphan(skb);
1735                 nf_reset_ct(skb);
1736         }
1737
1738         /* If running out of space, stop queue to avoid getting packets that we
1739          * are then unable to transmit.
1740          * An alternative would be to force queuing layer to requeue the skb by
1741          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1742          * returned in a normal path of operation: it means that driver is not
1743          * maintaining the TX queue stop/start state properly, and causes
1744          * the stack to do a non-trivial amount of useless work.
1745          * Since most packets only take 1 or 2 ring slots, stopping the queue
1746          * early means 16 slots are typically wasted.
1747          */
1748         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1749                 netif_stop_subqueue(dev, qnum);
1750                 if (!use_napi &&
1751                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1752                         /* More just got used, free them then recheck. */
1753                         free_old_xmit_skbs(sq, false);
1754                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1755                                 netif_start_subqueue(dev, qnum);
1756                                 virtqueue_disable_cb(sq->vq);
1757                         }
1758                 }
1759         }
1760
1761         if (kick || netif_xmit_stopped(txq)) {
1762                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1763                         u64_stats_update_begin(&sq->stats.syncp);
1764                         sq->stats.kicks++;
1765                         u64_stats_update_end(&sq->stats.syncp);
1766                 }
1767         }
1768
1769         return NETDEV_TX_OK;
1770 }
1771
1772 /*
1773  * Send command via the control virtqueue and check status.  Commands
1774  * supported by the hypervisor, as indicated by feature bits, should
1775  * never fail unless improperly formatted.
1776  */
1777 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1778                                  struct scatterlist *out)
1779 {
1780         struct scatterlist *sgs[4], hdr, stat;
1781         unsigned out_num = 0, tmp;
1782         int ret;
1783
1784         /* Caller should know better */
1785         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1786
1787         vi->ctrl->status = ~0;
1788         vi->ctrl->hdr.class = class;
1789         vi->ctrl->hdr.cmd = cmd;
1790         /* Add header */
1791         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1792         sgs[out_num++] = &hdr;
1793
1794         if (out)
1795                 sgs[out_num++] = out;
1796
1797         /* Add return status. */
1798         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1799         sgs[out_num] = &stat;
1800
1801         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1802         ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1803         if (ret < 0) {
1804                 dev_warn(&vi->vdev->dev,
1805                          "Failed to add sgs for command vq: %d\n.", ret);
1806                 return false;
1807         }
1808
1809         if (unlikely(!virtqueue_kick(vi->cvq)))
1810                 return vi->ctrl->status == VIRTIO_NET_OK;
1811
1812         /* Spin for a response, the kick causes an ioport write, trapping
1813          * into the hypervisor, so the request should be handled immediately.
1814          */
1815         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1816                !virtqueue_is_broken(vi->cvq))
1817                 cpu_relax();
1818
1819         return vi->ctrl->status == VIRTIO_NET_OK;
1820 }
1821
1822 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1823 {
1824         struct virtnet_info *vi = netdev_priv(dev);
1825         struct virtio_device *vdev = vi->vdev;
1826         int ret;
1827         struct sockaddr *addr;
1828         struct scatterlist sg;
1829
1830         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1831                 return -EOPNOTSUPP;
1832
1833         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1834         if (!addr)
1835                 return -ENOMEM;
1836
1837         ret = eth_prepare_mac_addr_change(dev, addr);
1838         if (ret)
1839                 goto out;
1840
1841         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1842                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1843                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1844                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1845                         dev_warn(&vdev->dev,
1846                                  "Failed to set mac address by vq command.\n");
1847                         ret = -EINVAL;
1848                         goto out;
1849                 }
1850         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1851                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1852                 unsigned int i;
1853
1854                 /* Naturally, this has an atomicity problem. */
1855                 for (i = 0; i < dev->addr_len; i++)
1856                         virtio_cwrite8(vdev,
1857                                        offsetof(struct virtio_net_config, mac) +
1858                                        i, addr->sa_data[i]);
1859         }
1860
1861         eth_commit_mac_addr_change(dev, p);
1862         ret = 0;
1863
1864 out:
1865         kfree(addr);
1866         return ret;
1867 }
1868
1869 static void virtnet_stats(struct net_device *dev,
1870                           struct rtnl_link_stats64 *tot)
1871 {
1872         struct virtnet_info *vi = netdev_priv(dev);
1873         unsigned int start;
1874         int i;
1875
1876         for (i = 0; i < vi->max_queue_pairs; i++) {
1877                 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
1878                 struct receive_queue *rq = &vi->rq[i];
1879                 struct send_queue *sq = &vi->sq[i];
1880
1881                 do {
1882                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1883                         tpackets = sq->stats.packets;
1884                         tbytes   = sq->stats.bytes;
1885                         terrors  = sq->stats.tx_timeouts;
1886                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1887
1888                 do {
1889                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1890                         rpackets = rq->stats.packets;
1891                         rbytes   = rq->stats.bytes;
1892                         rdrops   = rq->stats.drops;
1893                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1894
1895                 tot->rx_packets += rpackets;
1896                 tot->tx_packets += tpackets;
1897                 tot->rx_bytes   += rbytes;
1898                 tot->tx_bytes   += tbytes;
1899                 tot->rx_dropped += rdrops;
1900                 tot->tx_errors  += terrors;
1901         }
1902
1903         tot->tx_dropped = dev->stats.tx_dropped;
1904         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1905         tot->rx_length_errors = dev->stats.rx_length_errors;
1906         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1907 }
1908
1909 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1910 {
1911         rtnl_lock();
1912         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1913                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1914                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1915         rtnl_unlock();
1916 }
1917
1918 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1919 {
1920         struct scatterlist sg;
1921         struct net_device *dev = vi->dev;
1922
1923         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1924                 return 0;
1925
1926         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1927         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1928
1929         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1930                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1931                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1932                          queue_pairs);
1933                 return -EINVAL;
1934         } else {
1935                 vi->curr_queue_pairs = queue_pairs;
1936                 /* virtnet_open() will refill when device is going to up. */
1937                 if (dev->flags & IFF_UP)
1938                         schedule_delayed_work(&vi->refill, 0);
1939         }
1940
1941         return 0;
1942 }
1943
1944 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1945 {
1946         int err;
1947
1948         rtnl_lock();
1949         err = _virtnet_set_queues(vi, queue_pairs);
1950         rtnl_unlock();
1951         return err;
1952 }
1953
1954 static int virtnet_close(struct net_device *dev)
1955 {
1956         struct virtnet_info *vi = netdev_priv(dev);
1957         int i;
1958
1959         /* Make sure refill_work doesn't re-enable napi! */
1960         cancel_delayed_work_sync(&vi->refill);
1961
1962         for (i = 0; i < vi->max_queue_pairs; i++) {
1963                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1964                 napi_disable(&vi->rq[i].napi);
1965                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1966         }
1967
1968         return 0;
1969 }
1970
1971 static void virtnet_set_rx_mode(struct net_device *dev)
1972 {
1973         struct virtnet_info *vi = netdev_priv(dev);
1974         struct scatterlist sg[2];
1975         struct virtio_net_ctrl_mac *mac_data;
1976         struct netdev_hw_addr *ha;
1977         int uc_count;
1978         int mc_count;
1979         void *buf;
1980         int i;
1981
1982         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1983         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1984                 return;
1985
1986         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1987         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1988
1989         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1990
1991         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1992                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1993                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1994                          vi->ctrl->promisc ? "en" : "dis");
1995
1996         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1997
1998         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1999                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2000                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2001                          vi->ctrl->allmulti ? "en" : "dis");
2002
2003         uc_count = netdev_uc_count(dev);
2004         mc_count = netdev_mc_count(dev);
2005         /* MAC filter - use one buffer for both lists */
2006         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2007                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2008         mac_data = buf;
2009         if (!buf)
2010                 return;
2011
2012         sg_init_table(sg, 2);
2013
2014         /* Store the unicast list and count in the front of the buffer */
2015         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2016         i = 0;
2017         netdev_for_each_uc_addr(ha, dev)
2018                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2019
2020         sg_set_buf(&sg[0], mac_data,
2021                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2022
2023         /* multicast list and count fill the end */
2024         mac_data = (void *)&mac_data->macs[uc_count][0];
2025
2026         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2027         i = 0;
2028         netdev_for_each_mc_addr(ha, dev)
2029                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2030
2031         sg_set_buf(&sg[1], mac_data,
2032                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2033
2034         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2035                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2036                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2037
2038         kfree(buf);
2039 }
2040
2041 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2042                                    __be16 proto, u16 vid)
2043 {
2044         struct virtnet_info *vi = netdev_priv(dev);
2045         struct scatterlist sg;
2046
2047         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2048         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2049
2050         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2051                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2052                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2053         return 0;
2054 }
2055
2056 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2057                                     __be16 proto, u16 vid)
2058 {
2059         struct virtnet_info *vi = netdev_priv(dev);
2060         struct scatterlist sg;
2061
2062         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2063         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2064
2065         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2066                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2067                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2068         return 0;
2069 }
2070
2071 static void virtnet_clean_affinity(struct virtnet_info *vi)
2072 {
2073         int i;
2074
2075         if (vi->affinity_hint_set) {
2076                 for (i = 0; i < vi->max_queue_pairs; i++) {
2077                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
2078                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
2079                 }
2080
2081                 vi->affinity_hint_set = false;
2082         }
2083 }
2084
2085 static void virtnet_set_affinity(struct virtnet_info *vi)
2086 {
2087         cpumask_var_t mask;
2088         int stragglers;
2089         int group_size;
2090         int i, j, cpu;
2091         int num_cpu;
2092         int stride;
2093
2094         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2095                 virtnet_clean_affinity(vi);
2096                 return;
2097         }
2098
2099         num_cpu = num_online_cpus();
2100         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2101         stragglers = num_cpu >= vi->curr_queue_pairs ?
2102                         num_cpu % vi->curr_queue_pairs :
2103                         0;
2104         cpu = cpumask_first(cpu_online_mask);
2105
2106         for (i = 0; i < vi->curr_queue_pairs; i++) {
2107                 group_size = stride + (i < stragglers ? 1 : 0);
2108
2109                 for (j = 0; j < group_size; j++) {
2110                         cpumask_set_cpu(cpu, mask);
2111                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2112                                                 nr_cpu_ids, false);
2113                 }
2114                 virtqueue_set_affinity(vi->rq[i].vq, mask);
2115                 virtqueue_set_affinity(vi->sq[i].vq, mask);
2116                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2117                 cpumask_clear(mask);
2118         }
2119
2120         vi->affinity_hint_set = true;
2121         free_cpumask_var(mask);
2122 }
2123
2124 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2125 {
2126         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2127                                                    node);
2128         virtnet_set_affinity(vi);
2129         return 0;
2130 }
2131
2132 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2133 {
2134         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2135                                                    node_dead);
2136         virtnet_set_affinity(vi);
2137         return 0;
2138 }
2139
2140 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2141 {
2142         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2143                                                    node);
2144
2145         virtnet_clean_affinity(vi);
2146         return 0;
2147 }
2148
2149 static enum cpuhp_state virtionet_online;
2150
2151 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2152 {
2153         int ret;
2154
2155         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2156         if (ret)
2157                 return ret;
2158         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2159                                                &vi->node_dead);
2160         if (!ret)
2161                 return ret;
2162         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2163         return ret;
2164 }
2165
2166 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2167 {
2168         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2169         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2170                                             &vi->node_dead);
2171 }
2172
2173 static void virtnet_get_ringparam(struct net_device *dev,
2174                                   struct ethtool_ringparam *ring,
2175                                   struct kernel_ethtool_ringparam *kernel_ring,
2176                                   struct netlink_ext_ack *extack)
2177 {
2178         struct virtnet_info *vi = netdev_priv(dev);
2179
2180         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2181         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2182         ring->rx_pending = ring->rx_max_pending;
2183         ring->tx_pending = ring->tx_max_pending;
2184 }
2185
2186
2187 static void virtnet_get_drvinfo(struct net_device *dev,
2188                                 struct ethtool_drvinfo *info)
2189 {
2190         struct virtnet_info *vi = netdev_priv(dev);
2191         struct virtio_device *vdev = vi->vdev;
2192
2193         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2194         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2195         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2196
2197 }
2198
2199 /* TODO: Eliminate OOO packets during switching */
2200 static int virtnet_set_channels(struct net_device *dev,
2201                                 struct ethtool_channels *channels)
2202 {
2203         struct virtnet_info *vi = netdev_priv(dev);
2204         u16 queue_pairs = channels->combined_count;
2205         int err;
2206
2207         /* We don't support separate rx/tx channels.
2208          * We don't allow setting 'other' channels.
2209          */
2210         if (channels->rx_count || channels->tx_count || channels->other_count)
2211                 return -EINVAL;
2212
2213         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2214                 return -EINVAL;
2215
2216         /* For now we don't support modifying channels while XDP is loaded
2217          * also when XDP is loaded all RX queues have XDP programs so we only
2218          * need to check a single RX queue.
2219          */
2220         if (vi->rq[0].xdp_prog)
2221                 return -EINVAL;
2222
2223         cpus_read_lock();
2224         err = _virtnet_set_queues(vi, queue_pairs);
2225         if (err) {
2226                 cpus_read_unlock();
2227                 goto err;
2228         }
2229         virtnet_set_affinity(vi);
2230         cpus_read_unlock();
2231
2232         netif_set_real_num_tx_queues(dev, queue_pairs);
2233         netif_set_real_num_rx_queues(dev, queue_pairs);
2234  err:
2235         return err;
2236 }
2237
2238 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2239 {
2240         struct virtnet_info *vi = netdev_priv(dev);
2241         unsigned int i, j;
2242         u8 *p = data;
2243
2244         switch (stringset) {
2245         case ETH_SS_STATS:
2246                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2247                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2248                                 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2249                                                 virtnet_rq_stats_desc[j].desc);
2250                 }
2251
2252                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2253                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2254                                 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2255                                                 virtnet_sq_stats_desc[j].desc);
2256                 }
2257                 break;
2258         }
2259 }
2260
2261 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2262 {
2263         struct virtnet_info *vi = netdev_priv(dev);
2264
2265         switch (sset) {
2266         case ETH_SS_STATS:
2267                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2268                                                VIRTNET_SQ_STATS_LEN);
2269         default:
2270                 return -EOPNOTSUPP;
2271         }
2272 }
2273
2274 static void virtnet_get_ethtool_stats(struct net_device *dev,
2275                                       struct ethtool_stats *stats, u64 *data)
2276 {
2277         struct virtnet_info *vi = netdev_priv(dev);
2278         unsigned int idx = 0, start, i, j;
2279         const u8 *stats_base;
2280         size_t offset;
2281
2282         for (i = 0; i < vi->curr_queue_pairs; i++) {
2283                 struct receive_queue *rq = &vi->rq[i];
2284
2285                 stats_base = (u8 *)&rq->stats;
2286                 do {
2287                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2288                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2289                                 offset = virtnet_rq_stats_desc[j].offset;
2290                                 data[idx + j] = *(u64 *)(stats_base + offset);
2291                         }
2292                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2293                 idx += VIRTNET_RQ_STATS_LEN;
2294         }
2295
2296         for (i = 0; i < vi->curr_queue_pairs; i++) {
2297                 struct send_queue *sq = &vi->sq[i];
2298
2299                 stats_base = (u8 *)&sq->stats;
2300                 do {
2301                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2302                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2303                                 offset = virtnet_sq_stats_desc[j].offset;
2304                                 data[idx + j] = *(u64 *)(stats_base + offset);
2305                         }
2306                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2307                 idx += VIRTNET_SQ_STATS_LEN;
2308         }
2309 }
2310
2311 static void virtnet_get_channels(struct net_device *dev,
2312                                  struct ethtool_channels *channels)
2313 {
2314         struct virtnet_info *vi = netdev_priv(dev);
2315
2316         channels->combined_count = vi->curr_queue_pairs;
2317         channels->max_combined = vi->max_queue_pairs;
2318         channels->max_other = 0;
2319         channels->rx_count = 0;
2320         channels->tx_count = 0;
2321         channels->other_count = 0;
2322 }
2323
2324 static int virtnet_set_link_ksettings(struct net_device *dev,
2325                                       const struct ethtool_link_ksettings *cmd)
2326 {
2327         struct virtnet_info *vi = netdev_priv(dev);
2328
2329         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2330                                                   &vi->speed, &vi->duplex);
2331 }
2332
2333 static int virtnet_get_link_ksettings(struct net_device *dev,
2334                                       struct ethtool_link_ksettings *cmd)
2335 {
2336         struct virtnet_info *vi = netdev_priv(dev);
2337
2338         cmd->base.speed = vi->speed;
2339         cmd->base.duplex = vi->duplex;
2340         cmd->base.port = PORT_OTHER;
2341
2342         return 0;
2343 }
2344
2345 static int virtnet_set_coalesce(struct net_device *dev,
2346                                 struct ethtool_coalesce *ec,
2347                                 struct kernel_ethtool_coalesce *kernel_coal,
2348                                 struct netlink_ext_ack *extack)
2349 {
2350         struct virtnet_info *vi = netdev_priv(dev);
2351         int i, napi_weight;
2352
2353         if (ec->tx_max_coalesced_frames > 1 ||
2354             ec->rx_max_coalesced_frames != 1)
2355                 return -EINVAL;
2356
2357         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2358         if (napi_weight ^ vi->sq[0].napi.weight) {
2359                 if (dev->flags & IFF_UP)
2360                         return -EBUSY;
2361                 for (i = 0; i < vi->max_queue_pairs; i++)
2362                         vi->sq[i].napi.weight = napi_weight;
2363         }
2364
2365         return 0;
2366 }
2367
2368 static int virtnet_get_coalesce(struct net_device *dev,
2369                                 struct ethtool_coalesce *ec,
2370                                 struct kernel_ethtool_coalesce *kernel_coal,
2371                                 struct netlink_ext_ack *extack)
2372 {
2373         struct ethtool_coalesce ec_default = {
2374                 .cmd = ETHTOOL_GCOALESCE,
2375                 .rx_max_coalesced_frames = 1,
2376         };
2377         struct virtnet_info *vi = netdev_priv(dev);
2378
2379         memcpy(ec, &ec_default, sizeof(ec_default));
2380
2381         if (vi->sq[0].napi.weight)
2382                 ec->tx_max_coalesced_frames = 1;
2383
2384         return 0;
2385 }
2386
2387 static void virtnet_init_settings(struct net_device *dev)
2388 {
2389         struct virtnet_info *vi = netdev_priv(dev);
2390
2391         vi->speed = SPEED_UNKNOWN;
2392         vi->duplex = DUPLEX_UNKNOWN;
2393 }
2394
2395 static void virtnet_update_settings(struct virtnet_info *vi)
2396 {
2397         u32 speed;
2398         u8 duplex;
2399
2400         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2401                 return;
2402
2403         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2404
2405         if (ethtool_validate_speed(speed))
2406                 vi->speed = speed;
2407
2408         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2409
2410         if (ethtool_validate_duplex(duplex))
2411                 vi->duplex = duplex;
2412 }
2413
2414 static const struct ethtool_ops virtnet_ethtool_ops = {
2415         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2416         .get_drvinfo = virtnet_get_drvinfo,
2417         .get_link = ethtool_op_get_link,
2418         .get_ringparam = virtnet_get_ringparam,
2419         .get_strings = virtnet_get_strings,
2420         .get_sset_count = virtnet_get_sset_count,
2421         .get_ethtool_stats = virtnet_get_ethtool_stats,
2422         .set_channels = virtnet_set_channels,
2423         .get_channels = virtnet_get_channels,
2424         .get_ts_info = ethtool_op_get_ts_info,
2425         .get_link_ksettings = virtnet_get_link_ksettings,
2426         .set_link_ksettings = virtnet_set_link_ksettings,
2427         .set_coalesce = virtnet_set_coalesce,
2428         .get_coalesce = virtnet_get_coalesce,
2429 };
2430
2431 static void virtnet_freeze_down(struct virtio_device *vdev)
2432 {
2433         struct virtnet_info *vi = vdev->priv;
2434         int i;
2435
2436         /* Make sure no work handler is accessing the device */
2437         flush_work(&vi->config_work);
2438
2439         netif_tx_lock_bh(vi->dev);
2440         netif_device_detach(vi->dev);
2441         netif_tx_unlock_bh(vi->dev);
2442         cancel_delayed_work_sync(&vi->refill);
2443
2444         if (netif_running(vi->dev)) {
2445                 for (i = 0; i < vi->max_queue_pairs; i++) {
2446                         napi_disable(&vi->rq[i].napi);
2447                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2448                 }
2449         }
2450 }
2451
2452 static int init_vqs(struct virtnet_info *vi);
2453
2454 static int virtnet_restore_up(struct virtio_device *vdev)
2455 {
2456         struct virtnet_info *vi = vdev->priv;
2457         int err, i;
2458
2459         err = init_vqs(vi);
2460         if (err)
2461                 return err;
2462
2463         virtio_device_ready(vdev);
2464
2465         if (netif_running(vi->dev)) {
2466                 for (i = 0; i < vi->curr_queue_pairs; i++)
2467                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2468                                 schedule_delayed_work(&vi->refill, 0);
2469
2470                 for (i = 0; i < vi->max_queue_pairs; i++) {
2471                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2472                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2473                                                &vi->sq[i].napi);
2474                 }
2475         }
2476
2477         netif_tx_lock_bh(vi->dev);
2478         netif_device_attach(vi->dev);
2479         netif_tx_unlock_bh(vi->dev);
2480         return err;
2481 }
2482
2483 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2484 {
2485         struct scatterlist sg;
2486         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2487
2488         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2489
2490         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2491                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2492                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2493                 return -EINVAL;
2494         }
2495
2496         return 0;
2497 }
2498
2499 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2500 {
2501         u64 offloads = 0;
2502
2503         if (!vi->guest_offloads)
2504                 return 0;
2505
2506         return virtnet_set_guest_offloads(vi, offloads);
2507 }
2508
2509 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2510 {
2511         u64 offloads = vi->guest_offloads;
2512
2513         if (!vi->guest_offloads)
2514                 return 0;
2515
2516         return virtnet_set_guest_offloads(vi, offloads);
2517 }
2518
2519 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2520                            struct netlink_ext_ack *extack)
2521 {
2522         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2523         struct virtnet_info *vi = netdev_priv(dev);
2524         struct bpf_prog *old_prog;
2525         u16 xdp_qp = 0, curr_qp;
2526         int i, err;
2527
2528         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2529             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2530                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2531                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2532                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2533                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2534                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
2535                 return -EOPNOTSUPP;
2536         }
2537
2538         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2539                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2540                 return -EINVAL;
2541         }
2542
2543         if (dev->mtu > max_sz) {
2544                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2545                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2546                 return -EINVAL;
2547         }
2548
2549         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2550         if (prog)
2551                 xdp_qp = nr_cpu_ids;
2552
2553         /* XDP requires extra queues for XDP_TX */
2554         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2555                 netdev_warn_once(dev, "XDP request %i queues but max is %i. XDP_TX and XDP_REDIRECT will operate in a slower locked tx mode.\n",
2556                                  curr_qp + xdp_qp, vi->max_queue_pairs);
2557                 xdp_qp = 0;
2558         }
2559
2560         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2561         if (!prog && !old_prog)
2562                 return 0;
2563
2564         if (prog)
2565                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2566
2567         /* Make sure NAPI is not using any XDP TX queues for RX. */
2568         if (netif_running(dev)) {
2569                 for (i = 0; i < vi->max_queue_pairs; i++) {
2570                         napi_disable(&vi->rq[i].napi);
2571                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2572                 }
2573         }
2574
2575         if (!prog) {
2576                 for (i = 0; i < vi->max_queue_pairs; i++) {
2577                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2578                         if (i == 0)
2579                                 virtnet_restore_guest_offloads(vi);
2580                 }
2581                 synchronize_net();
2582         }
2583
2584         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2585         if (err)
2586                 goto err;
2587         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2588         vi->xdp_queue_pairs = xdp_qp;
2589
2590         if (prog) {
2591                 vi->xdp_enabled = true;
2592                 for (i = 0; i < vi->max_queue_pairs; i++) {
2593                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2594                         if (i == 0 && !old_prog)
2595                                 virtnet_clear_guest_offloads(vi);
2596                 }
2597         } else {
2598                 vi->xdp_enabled = false;
2599         }
2600
2601         for (i = 0; i < vi->max_queue_pairs; i++) {
2602                 if (old_prog)
2603                         bpf_prog_put(old_prog);
2604                 if (netif_running(dev)) {
2605                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2606                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2607                                                &vi->sq[i].napi);
2608                 }
2609         }
2610
2611         return 0;
2612
2613 err:
2614         if (!prog) {
2615                 virtnet_clear_guest_offloads(vi);
2616                 for (i = 0; i < vi->max_queue_pairs; i++)
2617                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2618         }
2619
2620         if (netif_running(dev)) {
2621                 for (i = 0; i < vi->max_queue_pairs; i++) {
2622                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2623                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2624                                                &vi->sq[i].napi);
2625                 }
2626         }
2627         if (prog)
2628                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2629         return err;
2630 }
2631
2632 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2633 {
2634         switch (xdp->command) {
2635         case XDP_SETUP_PROG:
2636                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2637         default:
2638                 return -EINVAL;
2639         }
2640 }
2641
2642 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2643                                       size_t len)
2644 {
2645         struct virtnet_info *vi = netdev_priv(dev);
2646         int ret;
2647
2648         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2649                 return -EOPNOTSUPP;
2650
2651         ret = snprintf(buf, len, "sby");
2652         if (ret >= len)
2653                 return -EOPNOTSUPP;
2654
2655         return 0;
2656 }
2657
2658 static int virtnet_set_features(struct net_device *dev,
2659                                 netdev_features_t features)
2660 {
2661         struct virtnet_info *vi = netdev_priv(dev);
2662         u64 offloads;
2663         int err;
2664
2665         if ((dev->features ^ features) & NETIF_F_GRO_HW) {
2666                 if (vi->xdp_enabled)
2667                         return -EBUSY;
2668
2669                 if (features & NETIF_F_GRO_HW)
2670                         offloads = vi->guest_offloads_capable;
2671                 else
2672                         offloads = vi->guest_offloads_capable &
2673                                    ~GUEST_OFFLOAD_GRO_HW_MASK;
2674
2675                 err = virtnet_set_guest_offloads(vi, offloads);
2676                 if (err)
2677                         return err;
2678                 vi->guest_offloads = offloads;
2679         }
2680
2681         return 0;
2682 }
2683
2684 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
2685 {
2686         struct virtnet_info *priv = netdev_priv(dev);
2687         struct send_queue *sq = &priv->sq[txqueue];
2688         struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
2689
2690         u64_stats_update_begin(&sq->stats.syncp);
2691         sq->stats.tx_timeouts++;
2692         u64_stats_update_end(&sq->stats.syncp);
2693
2694         netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
2695                    txqueue, sq->name, sq->vq->index, sq->vq->name,
2696                    jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
2697 }
2698
2699 static const struct net_device_ops virtnet_netdev = {
2700         .ndo_open            = virtnet_open,
2701         .ndo_stop            = virtnet_close,
2702         .ndo_start_xmit      = start_xmit,
2703         .ndo_validate_addr   = eth_validate_addr,
2704         .ndo_set_mac_address = virtnet_set_mac_address,
2705         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2706         .ndo_get_stats64     = virtnet_stats,
2707         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2708         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2709         .ndo_bpf                = virtnet_xdp,
2710         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2711         .ndo_features_check     = passthru_features_check,
2712         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2713         .ndo_set_features       = virtnet_set_features,
2714         .ndo_tx_timeout         = virtnet_tx_timeout,
2715 };
2716
2717 static void virtnet_config_changed_work(struct work_struct *work)
2718 {
2719         struct virtnet_info *vi =
2720                 container_of(work, struct virtnet_info, config_work);
2721         u16 v;
2722
2723         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2724                                  struct virtio_net_config, status, &v) < 0)
2725                 return;
2726
2727         if (v & VIRTIO_NET_S_ANNOUNCE) {
2728                 netdev_notify_peers(vi->dev);
2729                 virtnet_ack_link_announce(vi);
2730         }
2731
2732         /* Ignore unknown (future) status bits */
2733         v &= VIRTIO_NET_S_LINK_UP;
2734
2735         if (vi->status == v)
2736                 return;
2737
2738         vi->status = v;
2739
2740         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2741                 virtnet_update_settings(vi);
2742                 netif_carrier_on(vi->dev);
2743                 netif_tx_wake_all_queues(vi->dev);
2744         } else {
2745                 netif_carrier_off(vi->dev);
2746                 netif_tx_stop_all_queues(vi->dev);
2747         }
2748 }
2749
2750 static void virtnet_config_changed(struct virtio_device *vdev)
2751 {
2752         struct virtnet_info *vi = vdev->priv;
2753
2754         schedule_work(&vi->config_work);
2755 }
2756
2757 static void virtnet_free_queues(struct virtnet_info *vi)
2758 {
2759         int i;
2760
2761         for (i = 0; i < vi->max_queue_pairs; i++) {
2762                 __netif_napi_del(&vi->rq[i].napi);
2763                 __netif_napi_del(&vi->sq[i].napi);
2764         }
2765
2766         /* We called __netif_napi_del(),
2767          * we need to respect an RCU grace period before freeing vi->rq
2768          */
2769         synchronize_net();
2770
2771         kfree(vi->rq);
2772         kfree(vi->sq);
2773         kfree(vi->ctrl);
2774 }
2775
2776 static void _free_receive_bufs(struct virtnet_info *vi)
2777 {
2778         struct bpf_prog *old_prog;
2779         int i;
2780
2781         for (i = 0; i < vi->max_queue_pairs; i++) {
2782                 while (vi->rq[i].pages)
2783                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2784
2785                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2786                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2787                 if (old_prog)
2788                         bpf_prog_put(old_prog);
2789         }
2790 }
2791
2792 static void free_receive_bufs(struct virtnet_info *vi)
2793 {
2794         rtnl_lock();
2795         _free_receive_bufs(vi);
2796         rtnl_unlock();
2797 }
2798
2799 static void free_receive_page_frags(struct virtnet_info *vi)
2800 {
2801         int i;
2802         for (i = 0; i < vi->max_queue_pairs; i++)
2803                 if (vi->rq[i].alloc_frag.page)
2804                         put_page(vi->rq[i].alloc_frag.page);
2805 }
2806
2807 static void free_unused_bufs(struct virtnet_info *vi)
2808 {
2809         void *buf;
2810         int i;
2811
2812         for (i = 0; i < vi->max_queue_pairs; i++) {
2813                 struct virtqueue *vq = vi->sq[i].vq;
2814                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2815                         if (!is_xdp_frame(buf))
2816                                 dev_kfree_skb(buf);
2817                         else
2818                                 xdp_return_frame(ptr_to_xdp(buf));
2819                 }
2820         }
2821
2822         for (i = 0; i < vi->max_queue_pairs; i++) {
2823                 struct virtqueue *vq = vi->rq[i].vq;
2824
2825                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2826                         if (vi->mergeable_rx_bufs) {
2827                                 put_page(virt_to_head_page(buf));
2828                         } else if (vi->big_packets) {
2829                                 give_pages(&vi->rq[i], buf);
2830                         } else {
2831                                 put_page(virt_to_head_page(buf));
2832                         }
2833                 }
2834         }
2835 }
2836
2837 static void virtnet_del_vqs(struct virtnet_info *vi)
2838 {
2839         struct virtio_device *vdev = vi->vdev;
2840
2841         virtnet_clean_affinity(vi);
2842
2843         vdev->config->del_vqs(vdev);
2844
2845         virtnet_free_queues(vi);
2846 }
2847
2848 /* How large should a single buffer be so a queue full of these can fit at
2849  * least one full packet?
2850  * Logic below assumes the mergeable buffer header is used.
2851  */
2852 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2853 {
2854         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2855         unsigned int rq_size = virtqueue_get_vring_size(vq);
2856         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2857         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2858         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2859
2860         return max(max(min_buf_len, hdr_len) - hdr_len,
2861                    (unsigned int)GOOD_PACKET_LEN);
2862 }
2863
2864 static int virtnet_find_vqs(struct virtnet_info *vi)
2865 {
2866         vq_callback_t **callbacks;
2867         struct virtqueue **vqs;
2868         int ret = -ENOMEM;
2869         int i, total_vqs;
2870         const char **names;
2871         bool *ctx;
2872
2873         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2874          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2875          * possible control vq.
2876          */
2877         total_vqs = vi->max_queue_pairs * 2 +
2878                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2879
2880         /* Allocate space for find_vqs parameters */
2881         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2882         if (!vqs)
2883                 goto err_vq;
2884         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2885         if (!callbacks)
2886                 goto err_callback;
2887         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2888         if (!names)
2889                 goto err_names;
2890         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2891                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2892                 if (!ctx)
2893                         goto err_ctx;
2894         } else {
2895                 ctx = NULL;
2896         }
2897
2898         /* Parameters for control virtqueue, if any */
2899         if (vi->has_cvq) {
2900                 callbacks[total_vqs - 1] = NULL;
2901                 names[total_vqs - 1] = "control";
2902         }
2903
2904         /* Allocate/initialize parameters for send/receive virtqueues */
2905         for (i = 0; i < vi->max_queue_pairs; i++) {
2906                 callbacks[rxq2vq(i)] = skb_recv_done;
2907                 callbacks[txq2vq(i)] = skb_xmit_done;
2908                 sprintf(vi->rq[i].name, "input.%d", i);
2909                 sprintf(vi->sq[i].name, "output.%d", i);
2910                 names[rxq2vq(i)] = vi->rq[i].name;
2911                 names[txq2vq(i)] = vi->sq[i].name;
2912                 if (ctx)
2913                         ctx[rxq2vq(i)] = true;
2914         }
2915
2916         ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
2917                                   names, ctx, NULL);
2918         if (ret)
2919                 goto err_find;
2920
2921         if (vi->has_cvq) {
2922                 vi->cvq = vqs[total_vqs - 1];
2923                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2924                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2925         }
2926
2927         for (i = 0; i < vi->max_queue_pairs; i++) {
2928                 vi->rq[i].vq = vqs[rxq2vq(i)];
2929                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2930                 vi->sq[i].vq = vqs[txq2vq(i)];
2931         }
2932
2933         /* run here: ret == 0. */
2934
2935
2936 err_find:
2937         kfree(ctx);
2938 err_ctx:
2939         kfree(names);
2940 err_names:
2941         kfree(callbacks);
2942 err_callback:
2943         kfree(vqs);
2944 err_vq:
2945         return ret;
2946 }
2947
2948 static int virtnet_alloc_queues(struct virtnet_info *vi)
2949 {
2950         int i;
2951
2952         if (vi->has_cvq) {
2953                 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2954                 if (!vi->ctrl)
2955                         goto err_ctrl;
2956         } else {
2957                 vi->ctrl = NULL;
2958         }
2959         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2960         if (!vi->sq)
2961                 goto err_sq;
2962         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2963         if (!vi->rq)
2964                 goto err_rq;
2965
2966         INIT_DELAYED_WORK(&vi->refill, refill_work);
2967         for (i = 0; i < vi->max_queue_pairs; i++) {
2968                 vi->rq[i].pages = NULL;
2969                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2970                                napi_weight);
2971                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2972                                   napi_tx ? napi_weight : 0);
2973
2974                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2975                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2976                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2977
2978                 u64_stats_init(&vi->rq[i].stats.syncp);
2979                 u64_stats_init(&vi->sq[i].stats.syncp);
2980         }
2981
2982         return 0;
2983
2984 err_rq:
2985         kfree(vi->sq);
2986 err_sq:
2987         kfree(vi->ctrl);
2988 err_ctrl:
2989         return -ENOMEM;
2990 }
2991
2992 static int init_vqs(struct virtnet_info *vi)
2993 {
2994         int ret;
2995
2996         /* Allocate send & receive queues */
2997         ret = virtnet_alloc_queues(vi);
2998         if (ret)
2999                 goto err;
3000
3001         ret = virtnet_find_vqs(vi);
3002         if (ret)
3003                 goto err_free;
3004
3005         cpus_read_lock();
3006         virtnet_set_affinity(vi);
3007         cpus_read_unlock();
3008
3009         return 0;
3010
3011 err_free:
3012         virtnet_free_queues(vi);
3013 err:
3014         return ret;
3015 }
3016
3017 #ifdef CONFIG_SYSFS
3018 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
3019                 char *buf)
3020 {
3021         struct virtnet_info *vi = netdev_priv(queue->dev);
3022         unsigned int queue_index = get_netdev_rx_queue_index(queue);
3023         unsigned int headroom = virtnet_get_headroom(vi);
3024         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
3025         struct ewma_pkt_len *avg;
3026
3027         BUG_ON(queue_index >= vi->max_queue_pairs);
3028         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
3029         return sprintf(buf, "%u\n",
3030                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
3031                                        SKB_DATA_ALIGN(headroom + tailroom)));
3032 }
3033
3034 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
3035         __ATTR_RO(mergeable_rx_buffer_size);
3036
3037 static struct attribute *virtio_net_mrg_rx_attrs[] = {
3038         &mergeable_rx_buffer_size_attribute.attr,
3039         NULL
3040 };
3041
3042 static const struct attribute_group virtio_net_mrg_rx_group = {
3043         .name = "virtio_net",
3044         .attrs = virtio_net_mrg_rx_attrs
3045 };
3046 #endif
3047
3048 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3049                                     unsigned int fbit,
3050                                     const char *fname, const char *dname)
3051 {
3052         if (!virtio_has_feature(vdev, fbit))
3053                 return false;
3054
3055         dev_err(&vdev->dev, "device advertises feature %s but not %s",
3056                 fname, dname);
3057
3058         return true;
3059 }
3060
3061 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
3062         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3063
3064 static bool virtnet_validate_features(struct virtio_device *vdev)
3065 {
3066         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3067             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3068                              "VIRTIO_NET_F_CTRL_VQ") ||
3069              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3070                              "VIRTIO_NET_F_CTRL_VQ") ||
3071              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3072                              "VIRTIO_NET_F_CTRL_VQ") ||
3073              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3074              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3075                              "VIRTIO_NET_F_CTRL_VQ"))) {
3076                 return false;
3077         }
3078
3079         return true;
3080 }
3081
3082 #define MIN_MTU ETH_MIN_MTU
3083 #define MAX_MTU ETH_MAX_MTU
3084
3085 static int virtnet_validate(struct virtio_device *vdev)
3086 {
3087         if (!vdev->config->get) {
3088                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3089                         __func__);
3090                 return -EINVAL;
3091         }
3092
3093         if (!virtnet_validate_features(vdev))
3094                 return -EINVAL;
3095
3096         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3097                 int mtu = virtio_cread16(vdev,
3098                                          offsetof(struct virtio_net_config,
3099                                                   mtu));
3100                 if (mtu < MIN_MTU)
3101                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3102         }
3103
3104         return 0;
3105 }
3106
3107 static int virtnet_probe(struct virtio_device *vdev)
3108 {
3109         int i, err = -ENOMEM;
3110         struct net_device *dev;
3111         struct virtnet_info *vi;
3112         u16 max_queue_pairs;
3113         int mtu;
3114
3115         /* Find if host supports multiqueue virtio_net device */
3116         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3117                                    struct virtio_net_config,
3118                                    max_virtqueue_pairs, &max_queue_pairs);
3119
3120         /* We need at least 2 queue's */
3121         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3122             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3123             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3124                 max_queue_pairs = 1;
3125
3126         /* Allocate ourselves a network device with room for our info */
3127         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3128         if (!dev)
3129                 return -ENOMEM;
3130
3131         /* Set up network device as normal. */
3132         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3133                            IFF_TX_SKB_NO_LINEAR;
3134         dev->netdev_ops = &virtnet_netdev;
3135         dev->features = NETIF_F_HIGHDMA;
3136
3137         dev->ethtool_ops = &virtnet_ethtool_ops;
3138         SET_NETDEV_DEV(dev, &vdev->dev);
3139
3140         /* Do we support "hardware" checksums? */
3141         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3142                 /* This opens up the world of extra features. */
3143                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3144                 if (csum)
3145                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3146
3147                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3148                         dev->hw_features |= NETIF_F_TSO
3149                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3150                 }
3151                 /* Individual feature bits: what can host handle? */
3152                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3153                         dev->hw_features |= NETIF_F_TSO;
3154                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3155                         dev->hw_features |= NETIF_F_TSO6;
3156                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3157                         dev->hw_features |= NETIF_F_TSO_ECN;
3158
3159                 dev->features |= NETIF_F_GSO_ROBUST;
3160
3161                 if (gso)
3162                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3163                 /* (!csum && gso) case will be fixed by register_netdev() */
3164         }
3165         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3166                 dev->features |= NETIF_F_RXCSUM;
3167         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3168             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3169                 dev->features |= NETIF_F_GRO_HW;
3170         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3171                 dev->hw_features |= NETIF_F_GRO_HW;
3172
3173         dev->vlan_features = dev->features;
3174
3175         /* MTU range: 68 - 65535 */
3176         dev->min_mtu = MIN_MTU;
3177         dev->max_mtu = MAX_MTU;
3178
3179         /* Configuration may specify what MAC to use.  Otherwise random. */
3180         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
3181                 u8 addr[ETH_ALEN];
3182
3183                 virtio_cread_bytes(vdev,
3184                                    offsetof(struct virtio_net_config, mac),
3185                                    addr, ETH_ALEN);
3186                 eth_hw_addr_set(dev, addr);
3187         } else {
3188                 eth_hw_addr_random(dev);
3189         }
3190
3191         /* Set up our device-specific information */
3192         vi = netdev_priv(dev);
3193         vi->dev = dev;
3194         vi->vdev = vdev;
3195         vdev->priv = vi;
3196
3197         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3198
3199         /* If we can receive ANY GSO packets, we must allocate large ones. */
3200         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3201             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3202             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3203             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3204                 vi->big_packets = true;
3205
3206         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3207                 vi->mergeable_rx_bufs = true;
3208
3209         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3210             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3211                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3212         else
3213                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3214
3215         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3216             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3217                 vi->any_header_sg = true;
3218
3219         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3220                 vi->has_cvq = true;
3221
3222         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3223                 mtu = virtio_cread16(vdev,
3224                                      offsetof(struct virtio_net_config,
3225                                               mtu));
3226                 if (mtu < dev->min_mtu) {
3227                         /* Should never trigger: MTU was previously validated
3228                          * in virtnet_validate.
3229                          */
3230                         dev_err(&vdev->dev,
3231                                 "device MTU appears to have changed it is now %d < %d",
3232                                 mtu, dev->min_mtu);
3233                         err = -EINVAL;
3234                         goto free;
3235                 }
3236
3237                 dev->mtu = mtu;
3238                 dev->max_mtu = mtu;
3239
3240                 /* TODO: size buffers correctly in this case. */
3241                 if (dev->mtu > ETH_DATA_LEN)
3242                         vi->big_packets = true;
3243         }
3244
3245         if (vi->any_header_sg)
3246                 dev->needed_headroom = vi->hdr_len;
3247
3248         /* Enable multiqueue by default */
3249         if (num_online_cpus() >= max_queue_pairs)
3250                 vi->curr_queue_pairs = max_queue_pairs;
3251         else
3252                 vi->curr_queue_pairs = num_online_cpus();
3253         vi->max_queue_pairs = max_queue_pairs;
3254
3255         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3256         err = init_vqs(vi);
3257         if (err)
3258                 goto free;
3259
3260 #ifdef CONFIG_SYSFS
3261         if (vi->mergeable_rx_bufs)
3262                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3263 #endif
3264         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3265         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3266
3267         virtnet_init_settings(dev);
3268
3269         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3270                 vi->failover = net_failover_create(vi->dev);
3271                 if (IS_ERR(vi->failover)) {
3272                         err = PTR_ERR(vi->failover);
3273                         goto free_vqs;
3274                 }
3275         }
3276
3277         err = register_netdev(dev);
3278         if (err) {
3279                 pr_debug("virtio_net: registering device failed\n");
3280                 goto free_failover;
3281         }
3282
3283         virtio_device_ready(vdev);
3284
3285         err = virtnet_cpu_notif_add(vi);
3286         if (err) {
3287                 pr_debug("virtio_net: registering cpu notifier failed\n");
3288                 goto free_unregister_netdev;
3289         }
3290
3291         virtnet_set_queues(vi, vi->curr_queue_pairs);
3292
3293         /* Assume link up if device can't report link status,
3294            otherwise get link status from config. */
3295         netif_carrier_off(dev);
3296         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3297                 schedule_work(&vi->config_work);
3298         } else {
3299                 vi->status = VIRTIO_NET_S_LINK_UP;
3300                 virtnet_update_settings(vi);
3301                 netif_carrier_on(dev);
3302         }
3303
3304         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3305                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3306                         set_bit(guest_offloads[i], &vi->guest_offloads);
3307         vi->guest_offloads_capable = vi->guest_offloads;
3308
3309         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3310                  dev->name, max_queue_pairs);
3311
3312         return 0;
3313
3314 free_unregister_netdev:
3315         virtio_reset_device(vdev);
3316
3317         unregister_netdev(dev);
3318 free_failover:
3319         net_failover_destroy(vi->failover);
3320 free_vqs:
3321         cancel_delayed_work_sync(&vi->refill);
3322         free_receive_page_frags(vi);
3323         virtnet_del_vqs(vi);
3324 free:
3325         free_netdev(dev);
3326         return err;
3327 }
3328
3329 static void remove_vq_common(struct virtnet_info *vi)
3330 {
3331         virtio_reset_device(vi->vdev);
3332
3333         /* Free unused buffers in both send and recv, if any. */
3334         free_unused_bufs(vi);
3335
3336         free_receive_bufs(vi);
3337
3338         free_receive_page_frags(vi);
3339
3340         virtnet_del_vqs(vi);
3341 }
3342
3343 static void virtnet_remove(struct virtio_device *vdev)
3344 {
3345         struct virtnet_info *vi = vdev->priv;
3346
3347         virtnet_cpu_notif_remove(vi);
3348
3349         /* Make sure no work handler is accessing the device. */
3350         flush_work(&vi->config_work);
3351
3352         unregister_netdev(vi->dev);
3353
3354         net_failover_destroy(vi->failover);
3355
3356         remove_vq_common(vi);
3357
3358         free_netdev(vi->dev);
3359 }
3360
3361 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3362 {
3363         struct virtnet_info *vi = vdev->priv;
3364
3365         virtnet_cpu_notif_remove(vi);
3366         virtnet_freeze_down(vdev);
3367         remove_vq_common(vi);
3368
3369         return 0;
3370 }
3371
3372 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3373 {
3374         struct virtnet_info *vi = vdev->priv;
3375         int err;
3376
3377         err = virtnet_restore_up(vdev);
3378         if (err)
3379                 return err;
3380         virtnet_set_queues(vi, vi->curr_queue_pairs);
3381
3382         err = virtnet_cpu_notif_add(vi);
3383         if (err) {
3384                 virtnet_freeze_down(vdev);
3385                 remove_vq_common(vi);
3386                 return err;
3387         }
3388
3389         return 0;
3390 }
3391
3392 static struct virtio_device_id id_table[] = {
3393         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3394         { 0 },
3395 };
3396
3397 #define VIRTNET_FEATURES \
3398         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3399         VIRTIO_NET_F_MAC, \
3400         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3401         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3402         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3403         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3404         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3405         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3406         VIRTIO_NET_F_CTRL_MAC_ADDR, \
3407         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3408         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3409
3410 static unsigned int features[] = {
3411         VIRTNET_FEATURES,
3412 };
3413
3414 static unsigned int features_legacy[] = {
3415         VIRTNET_FEATURES,
3416         VIRTIO_NET_F_GSO,
3417         VIRTIO_F_ANY_LAYOUT,
3418 };
3419
3420 static struct virtio_driver virtio_net_driver = {
3421         .feature_table = features,
3422         .feature_table_size = ARRAY_SIZE(features),
3423         .feature_table_legacy = features_legacy,
3424         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3425         .driver.name =  KBUILD_MODNAME,
3426         .driver.owner = THIS_MODULE,
3427         .id_table =     id_table,
3428         .validate =     virtnet_validate,
3429         .probe =        virtnet_probe,
3430         .remove =       virtnet_remove,
3431         .config_changed = virtnet_config_changed,
3432 #ifdef CONFIG_PM_SLEEP
3433         .freeze =       virtnet_freeze,
3434         .restore =      virtnet_restore,
3435 #endif
3436 };
3437
3438 static __init int virtio_net_driver_init(void)
3439 {
3440         int ret;
3441
3442         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3443                                       virtnet_cpu_online,
3444                                       virtnet_cpu_down_prep);
3445         if (ret < 0)
3446                 goto out;
3447         virtionet_online = ret;
3448         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3449                                       NULL, virtnet_cpu_dead);
3450         if (ret)
3451                 goto err_dead;
3452         ret = register_virtio_driver(&virtio_net_driver);
3453         if (ret)
3454                 goto err_virtio;
3455         return 0;
3456 err_virtio:
3457         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3458 err_dead:
3459         cpuhp_remove_multi_state(virtionet_online);
3460 out:
3461         return ret;
3462 }
3463 module_init(virtio_net_driver_init);
3464
3465 static __exit void virtio_net_driver_exit(void)
3466 {
3467         unregister_virtio_driver(&virtio_net_driver);
3468         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3469         cpuhp_remove_multi_state(virtionet_online);
3470 }
3471 module_exit(virtio_net_driver_exit);
3472
3473 MODULE_DEVICE_TABLE(virtio, id_table);
3474 MODULE_DESCRIPTION("Virtio network driver");
3475 MODULE_LICENSE("GPL");