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_len;
 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(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                 put_page(page);
 830                 goto err;
 831         }
 832         skb_reserve(skb, headroom - delta);
 833         skb_put(skb, len);
 834         if (!xdp_prog) {
 835                 buf += header_offset;
 836                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 837         } /* keep zeroed vnet hdr since XDP is loaded */
 838
 839         if (metasize)
 840                 skb_metadata_set(skb, metasize);
 841
 842 err:
 843         return skb;
 844
 845 err_xdp:
 846         rcu_read_unlock();
 847         stats->xdp_drops++;
 848 err_len:
 849         stats->drops++;
 850         put_page(page);
 851 xdp_xmit:
 852         return NULL;
 853 }
 854
 855 static struct sk_buff *receive_big(struct net_device *dev,
 856                                    struct virtnet_info *vi,
 857                                    struct receive_queue *rq,
 858                                    void *buf,
 859                                    unsigned int len,
 860                                    struct virtnet_rq_stats *stats)
 861 {
 862         struct page *page = buf;
 863         struct sk_buff *skb =
 864                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
 865
 866         stats->bytes += len - vi->hdr_len;
 867         if (unlikely(!skb))
 868                 goto err;
 869
 870         return skb;
 871
 872 err:
 873         stats->drops++;
 874         give_pages(rq, page);
 875         return NULL;
 876 }
 877
 878 static struct sk_buff *receive_mergeable(struct net_device *dev,
 879                                          struct virtnet_info *vi,
 880                                          struct receive_queue *rq,
 881                                          void *buf,
 882                                          void *ctx,
 883                                          unsigned int len,
 884                                          unsigned int *xdp_xmit,
 885                                          struct virtnet_rq_stats *stats)
 886 {
 887         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 888         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 889         struct page *page = virt_to_head_page(buf);
 890         int offset = buf - page_address(page);
 891         struct sk_buff *head_skb, *curr_skb;
 892         struct bpf_prog *xdp_prog;
 893         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
 894         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 895         unsigned int metasize = 0;
 896         unsigned int frame_sz;
 897         int err;
 898
 899         head_skb = NULL;
 900         stats->bytes += len - vi->hdr_len;
 901
 902         if (unlikely(len > truesize)) {
 903                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 904                          dev->name, len, (unsigned long)ctx);
 905                 dev->stats.rx_length_errors++;
 906                 goto err_skb;
 907         }
 908
 909         if (likely(!vi->xdp_enabled)) {
 910                 xdp_prog = NULL;
 911                 goto skip_xdp;
 912         }
 913
 914         rcu_read_lock();
 915         xdp_prog = rcu_dereference(rq->xdp_prog);
 916         if (xdp_prog) {
 917                 struct xdp_frame *xdpf;
 918                 struct page *xdp_page;
 919                 struct xdp_buff xdp;
 920                 void *data;
 921                 u32 act;
 922
 923                 /* Transient failure which in theory could occur if
 924                  * in-flight packets from before XDP was enabled reach
 925                  * the receive path after XDP is loaded.
 926                  */
 927                 if (unlikely(hdr->hdr.gso_type))
 928                         goto err_xdp;
 929
 930                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 931                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 932                  */
 933                 frame_sz = headroom ? PAGE_SIZE : truesize;
 934
 935                 /* This happens when rx buffer size is underestimated
 936                  * or headroom is not enough because of the buffer
 937                  * was refilled before XDP is set. This should only
 938                  * happen for the first several packets, so we don't
 939                  * care much about its performance.
 940                  */
 941                 if (unlikely(num_buf > 1 ||
 942                              headroom < virtnet_get_headroom(vi))) {
 943                         /* linearize data for XDP */
 944                         xdp_page = xdp_linearize_page(rq, &num_buf,
 945                                                       page, offset,
 946                                                       VIRTIO_XDP_HEADROOM,
 947                                                       &len);
 948                         frame_sz = PAGE_SIZE;
 949
 950                         if (!xdp_page)
 951                                 goto err_xdp;
 952                         offset = VIRTIO_XDP_HEADROOM;
 953                 } else {
 954                         xdp_page = page;
 955                 }
 956
 957                 /* Allow consuming headroom but reserve enough space to push
 958                  * the descriptor on if we get an XDP_TX return code.
 959                  */
 960                 data = page_address(xdp_page) + offset;
 961                 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
 962                 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
 963                                  VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
 964
 965                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 966                 stats->xdp_packets++;
 967
 968                 switch (act) {
 969                 case XDP_PASS:
 970                         metasize = xdp.data - xdp.data_meta;
 971
 972                         /* recalculate offset to account for any header
 973                          * adjustments and minus the metasize to copy the
 974                          * metadata in page_to_skb(). Note other cases do not
 975                          * build an skb and avoid using offset
 976                          */
 977                         offset = xdp.data - page_address(xdp_page) -
 978                                  vi->hdr_len - metasize;
 979
 980                         /* recalculate len if xdp.data, xdp.data_end or
 981                          * xdp.data_meta were adjusted
 982                          */
 983                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 984                         /* We can only create skb based on xdp_page. */
 985                         if (unlikely(xdp_page != page)) {
 986                                 rcu_read_unlock();
 987                                 put_page(page);
 988                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
 989                                                        len, PAGE_SIZE, false,
 990                                                        metasize,
 991                                                        VIRTIO_XDP_HEADROOM);
 992                                 return head_skb;
 993                         }
 994                         break;
 995                 case XDP_TX:
 996                         stats->xdp_tx++;
 997                         xdpf = xdp_convert_buff_to_frame(&xdp);
 998                         if (unlikely(!xdpf))
 999                                 goto err_xdp;
1000                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
1001                         if (unlikely(!err)) {
1002                                 xdp_return_frame_rx_napi(xdpf);
1003                         } else if (unlikely(err < 0)) {
1004                                 trace_xdp_exception(vi->dev, xdp_prog, act);
1005                                 if (unlikely(xdp_page != page))
1006                                         put_page(xdp_page);
1007                                 goto err_xdp;
1008                         }
1009                         *xdp_xmit |= VIRTIO_XDP_TX;
1010                         if (unlikely(xdp_page != page))
1011                                 put_page(page);
1012                         rcu_read_unlock();
1013                         goto xdp_xmit;
1014                 case XDP_REDIRECT:
1015                         stats->xdp_redirects++;
1016                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
1017                         if (err) {
1018                                 if (unlikely(xdp_page != page))
1019                                         put_page(xdp_page);
1020                                 goto err_xdp;
1021                         }
1022                         *xdp_xmit |= VIRTIO_XDP_REDIR;
1023                         if (unlikely(xdp_page != page))
1024                                 put_page(page);
1025                         rcu_read_unlock();
1026                         goto xdp_xmit;
1027                 default:
1028                         bpf_warn_invalid_xdp_action(act);
1029                         fallthrough;
1030                 case XDP_ABORTED:
1031                         trace_xdp_exception(vi->dev, xdp_prog, act);
1032                         fallthrough;
1033                 case XDP_DROP:
1034                         if (unlikely(xdp_page != page))
1035                                 __free_pages(xdp_page, 0);
1036                         goto err_xdp;
1037                 }
1038         }
1039         rcu_read_unlock();
1040
1041 skip_xdp:
1042         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1043                                metasize, headroom);
1044         curr_skb = head_skb;
1045
1046         if (unlikely(!curr_skb))
1047                 goto err_skb;
1048         while (--num_buf) {
1049                 int num_skb_frags;
1050
1051                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1052                 if (unlikely(!buf)) {
1053                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1054                                  dev->name, num_buf,
1055                                  virtio16_to_cpu(vi->vdev,
1056                                                  hdr->num_buffers));
1057                         dev->stats.rx_length_errors++;
1058                         goto err_buf;
1059                 }
1060
1061                 stats->bytes += len;
1062                 page = virt_to_head_page(buf);
1063
1064                 truesize = mergeable_ctx_to_truesize(ctx);
1065                 if (unlikely(len > truesize)) {
1066                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1067                                  dev->name, len, (unsigned long)ctx);
1068                         dev->stats.rx_length_errors++;
1069                         goto err_skb;
1070                 }
1071
1072                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1073                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1074                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1075
1076                         if (unlikely(!nskb))
1077                                 goto err_skb;
1078                         if (curr_skb == head_skb)
1079                                 skb_shinfo(curr_skb)->frag_list = nskb;
1080                         else
1081                                 curr_skb->next = nskb;
1082                         curr_skb = nskb;
1083                         head_skb->truesize += nskb->truesize;
1084                         num_skb_frags = 0;
1085                 }
1086                 if (curr_skb != head_skb) {
1087                         head_skb->data_len += len;
1088                         head_skb->len += len;
1089                         head_skb->truesize += truesize;
1090                 }
1091                 offset = buf - page_address(page);
1092                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1093                         put_page(page);
1094                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1095                                              len, truesize);
1096                 } else {
1097                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1098                                         offset, len, truesize);
1099                 }
1100         }
1101
1102         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1103         return head_skb;
1104
1105 err_xdp:
1106         rcu_read_unlock();
1107         stats->xdp_drops++;
1108 err_skb:
1109         put_page(page);
1110         while (num_buf-- > 1) {
1111                 buf = virtqueue_get_buf(rq->vq, &len);
1112                 if (unlikely(!buf)) {
1113                         pr_debug("%s: rx error: %d buffers missing\n",
1114                                  dev->name, num_buf);
1115                         dev->stats.rx_length_errors++;
1116                         break;
1117                 }
1118                 stats->bytes += len;
1119                 page = virt_to_head_page(buf);
1120                 put_page(page);
1121         }
1122 err_buf:
1123         stats->drops++;
1124         dev_kfree_skb(head_skb);
1125 xdp_xmit:
1126         return NULL;
1127 }
1128
1129 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1130                         void *buf, unsigned int len, void **ctx,
1131                         unsigned int *xdp_xmit,
1132                         struct virtnet_rq_stats *stats)
1133 {
1134         struct net_device *dev = vi->dev;
1135         struct sk_buff *skb;
1136         struct virtio_net_hdr_mrg_rxbuf *hdr;
1137
1138         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1139                 pr_debug("%s: short packet %i\n", dev->name, len);
1140                 dev->stats.rx_length_errors++;
1141                 if (vi->mergeable_rx_bufs) {
1142                         put_page(virt_to_head_page(buf));
1143                 } else if (vi->big_packets) {
1144                         give_pages(rq, buf);
1145                 } else {
1146                         put_page(virt_to_head_page(buf));
1147                 }
1148                 return;
1149         }
1150
1151         if (vi->mergeable_rx_bufs)
1152                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1153                                         stats);
1154         else if (vi->big_packets)
1155                 skb = receive_big(dev, vi, rq, buf, len, stats);
1156         else
1157                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1158
1159         if (unlikely(!skb))
1160                 return;
1161
1162         hdr = skb_vnet_hdr(skb);
1163
1164         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1165                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1166
1167         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1168                                   virtio_is_little_endian(vi->vdev))) {
1169                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1170                                      dev->name, hdr->hdr.gso_type,
1171                                      hdr->hdr.gso_size);
1172                 goto frame_err;
1173         }
1174
1175         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1176         skb->protocol = eth_type_trans(skb, dev);
1177         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1178                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1179
1180         napi_gro_receive(&rq->napi, skb);
1181         return;
1182
1183 frame_err:
1184         dev->stats.rx_frame_errors++;
1185         dev_kfree_skb(skb);
1186 }
1187
1188 /* Unlike mergeable buffers, all buffers are allocated to the
1189  * same size, except for the headroom. For this reason we do
1190  * not need to use  mergeable_len_to_ctx here - it is enough
1191  * to store the headroom as the context ignoring the truesize.
1192  */
1193 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1194                              gfp_t gfp)
1195 {
1196         struct page_frag *alloc_frag = &rq->alloc_frag;
1197         char *buf;
1198         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1199         void *ctx = (void *)(unsigned long)xdp_headroom;
1200         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1201         int err;
1202
1203         len = SKB_DATA_ALIGN(len) +
1204               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1205         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1206                 return -ENOMEM;
1207
1208         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1209         get_page(alloc_frag->page);
1210         alloc_frag->offset += len;
1211         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1212                     vi->hdr_len + GOOD_PACKET_LEN);
1213         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1214         if (err < 0)
1215                 put_page(virt_to_head_page(buf));
1216         return err;
1217 }
1218
1219 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1220                            gfp_t gfp)
1221 {
1222         struct page *first, *list = NULL;
1223         char *p;
1224         int i, err, offset;
1225
1226         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1227
1228         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1229         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1230                 first = get_a_page(rq, gfp);
1231                 if (!first) {
1232                         if (list)
1233                                 give_pages(rq, list);
1234                         return -ENOMEM;
1235                 }
1236                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1237
1238                 /* chain new page in list head to match sg */
1239                 first->private = (unsigned long)list;
1240                 list = first;
1241         }
1242
1243         first = get_a_page(rq, gfp);
1244         if (!first) {
1245                 give_pages(rq, list);
1246                 return -ENOMEM;
1247         }
1248         p = page_address(first);
1249
1250         /* rq->sg[0], rq->sg[1] share the same page */
1251         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1252         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1253
1254         /* rq->sg[1] for data packet, from offset */
1255         offset = sizeof(struct padded_vnet_hdr);
1256         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1257
1258         /* chain first in list head */
1259         first->private = (unsigned long)list;
1260         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1261                                   first, gfp);
1262         if (err < 0)
1263                 give_pages(rq, first);
1264
1265         return err;
1266 }
1267
1268 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1269                                           struct ewma_pkt_len *avg_pkt_len,
1270                                           unsigned int room)
1271 {
1272         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1273         unsigned int len;
1274
1275         if (room)
1276                 return PAGE_SIZE - room;
1277
1278         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1279                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1280
1281         return ALIGN(len, L1_CACHE_BYTES);
1282 }
1283
1284 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1285                                  struct receive_queue *rq, gfp_t gfp)
1286 {
1287         struct page_frag *alloc_frag = &rq->alloc_frag;
1288         unsigned int headroom = virtnet_get_headroom(vi);
1289         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1290         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1291         char *buf;
1292         void *ctx;
1293         int err;
1294         unsigned int len, hole;
1295
1296         /* Extra tailroom is needed to satisfy XDP's assumption. This
1297          * means rx frags coalescing won't work, but consider we've
1298          * disabled GSO for XDP, it won't be a big issue.
1299          */
1300         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1301         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1302                 return -ENOMEM;
1303
1304         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1305         buf += headroom; /* advance address leaving hole at front of pkt */
1306         get_page(alloc_frag->page);
1307         alloc_frag->offset += len + room;
1308         hole = alloc_frag->size - alloc_frag->offset;
1309         if (hole < len + room) {
1310                 /* To avoid internal fragmentation, if there is very likely not
1311                  * enough space for another buffer, add the remaining space to
1312                  * the current buffer.
1313                  */
1314                 len += hole;
1315                 alloc_frag->offset += hole;
1316         }
1317
1318         sg_init_one(rq->sg, buf, len);
1319         ctx = mergeable_len_to_ctx(len, headroom);
1320         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1321         if (err < 0)
1322                 put_page(virt_to_head_page(buf));
1323
1324         return err;
1325 }
1326
1327 /*
1328  * Returns false if we couldn't fill entirely (OOM).
1329  *
1330  * Normally run in the receive path, but can also be run from ndo_open
1331  * before we're receiving packets, or from refill_work which is
1332  * careful to disable receiving (using napi_disable).
1333  */
1334 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1335                           gfp_t gfp)
1336 {
1337         int err;
1338         bool oom;
1339
1340         do {
1341                 if (vi->mergeable_rx_bufs)
1342                         err = add_recvbuf_mergeable(vi, rq, gfp);
1343                 else if (vi->big_packets)
1344                         err = add_recvbuf_big(vi, rq, gfp);
1345                 else
1346                         err = add_recvbuf_small(vi, rq, gfp);
1347
1348                 oom = err == -ENOMEM;
1349                 if (err)
1350                         break;
1351         } while (rq->vq->num_free);
1352         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1353                 unsigned long flags;
1354
1355                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1356                 rq->stats.kicks++;
1357                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1358         }
1359
1360         return !oom;
1361 }
1362
1363 static void skb_recv_done(struct virtqueue *rvq)
1364 {
1365         struct virtnet_info *vi = rvq->vdev->priv;
1366         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1367
1368         virtqueue_napi_schedule(&rq->napi, rvq);
1369 }
1370
1371 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1372 {
1373         napi_enable(napi);
1374
1375         /* If all buffers were filled by other side before we napi_enabled, we
1376          * won't get another interrupt, so process any outstanding packets now.
1377          * Call local_bh_enable after to trigger softIRQ processing.
1378          */
1379         local_bh_disable();
1380         virtqueue_napi_schedule(napi, vq);
1381         local_bh_enable();
1382 }
1383
1384 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1385                                    struct virtqueue *vq,
1386                                    struct napi_struct *napi)
1387 {
1388         if (!napi->weight)
1389                 return;
1390
1391         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1392          * enable the feature if this is likely affine with the transmit path.
1393          */
1394         if (!vi->affinity_hint_set) {
1395                 napi->weight = 0;
1396                 return;
1397         }
1398
1399         return virtnet_napi_enable(vq, napi);
1400 }
1401
1402 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1403 {
1404         if (napi->weight)
1405                 napi_disable(napi);
1406 }
1407
1408 static void refill_work(struct work_struct *work)
1409 {
1410         struct virtnet_info *vi =
1411                 container_of(work, struct virtnet_info, refill.work);
1412         bool still_empty;
1413         int i;
1414
1415         for (i = 0; i < vi->curr_queue_pairs; i++) {
1416                 struct receive_queue *rq = &vi->rq[i];
1417
1418                 napi_disable(&rq->napi);
1419                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1420                 virtnet_napi_enable(rq->vq, &rq->napi);
1421
1422                 /* In theory, this can happen: if we don't get any buffers in
1423                  * we will *never* try to fill again.
1424                  */
1425                 if (still_empty)
1426                         schedule_delayed_work(&vi->refill, HZ/2);
1427         }
1428 }
1429
1430 static int virtnet_receive(struct receive_queue *rq, int budget,
1431                            unsigned int *xdp_xmit)
1432 {
1433         struct virtnet_info *vi = rq->vq->vdev->priv;
1434         struct virtnet_rq_stats stats = {};
1435         unsigned int len;
1436         void *buf;
1437         int i;
1438
1439         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1440                 void *ctx;
1441
1442                 while (stats.packets < budget &&
1443                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1444                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1445                         stats.packets++;
1446                 }
1447         } else {
1448                 while (stats.packets < budget &&
1449                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1450                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1451                         stats.packets++;
1452                 }
1453         }
1454
1455         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1456                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1457                         schedule_delayed_work(&vi->refill, 0);
1458         }
1459
1460         u64_stats_update_begin(&rq->stats.syncp);
1461         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1462                 size_t offset = virtnet_rq_stats_desc[i].offset;
1463                 u64 *item;
1464
1465                 item = (u64 *)((u8 *)&rq->stats + offset);
1466                 *item += *(u64 *)((u8 *)&stats + offset);
1467         }
1468         u64_stats_update_end(&rq->stats.syncp);
1469
1470         return stats.packets;
1471 }
1472
1473 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1474 {
1475         unsigned int len;
1476         unsigned int packets = 0;
1477         unsigned int bytes = 0;
1478         void *ptr;
1479
1480         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1481                 if (likely(!is_xdp_frame(ptr))) {
1482                         struct sk_buff *skb = ptr;
1483
1484                         pr_debug("Sent skb %p\n", skb);
1485
1486                         bytes += skb->len;
1487                         napi_consume_skb(skb, in_napi);
1488                 } else {
1489                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1490
1491                         bytes += frame->len;
1492                         xdp_return_frame(frame);
1493                 }
1494                 packets++;
1495         }
1496
1497         /* Avoid overhead when no packets have been processed
1498          * happens when called speculatively from start_xmit.
1499          */
1500         if (!packets)
1501                 return;
1502
1503         u64_stats_update_begin(&sq->stats.syncp);
1504         sq->stats.bytes += bytes;
1505         sq->stats.packets += packets;
1506         u64_stats_update_end(&sq->stats.syncp);
1507 }
1508
1509 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1510 {
1511         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1512                 return false;
1513         else if (q < vi->curr_queue_pairs)
1514                 return true;
1515         else
1516                 return false;
1517 }
1518
1519 static void virtnet_poll_cleantx(struct receive_queue *rq)
1520 {
1521         struct virtnet_info *vi = rq->vq->vdev->priv;
1522         unsigned int index = vq2rxq(rq->vq);
1523         struct send_queue *sq = &vi->sq[index];
1524         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1525
1526         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1527                 return;
1528
1529         if (__netif_tx_trylock(txq)) {
1530                 do {
1531                         virtqueue_disable_cb(sq->vq);
1532                         free_old_xmit_skbs(sq, true);
1533                 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1534
1535                 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1536                         netif_tx_wake_queue(txq);
1537
1538                 __netif_tx_unlock(txq);
1539         }
1540 }
1541
1542 static int virtnet_poll(struct napi_struct *napi, int budget)
1543 {
1544         struct receive_queue *rq =
1545                 container_of(napi, struct receive_queue, napi);
1546         struct virtnet_info *vi = rq->vq->vdev->priv;
1547         struct send_queue *sq;
1548         unsigned int received;
1549         unsigned int xdp_xmit = 0;
1550
1551         virtnet_poll_cleantx(rq);
1552
1553         received = virtnet_receive(rq, budget, &xdp_xmit);
1554
1555         /* Out of packets? */
1556         if (received < budget)
1557                 virtqueue_napi_complete(napi, rq->vq, received);
1558
1559         if (xdp_xmit & VIRTIO_XDP_REDIR)
1560                 xdp_do_flush();
1561
1562         if (xdp_xmit & VIRTIO_XDP_TX) {
1563                 sq = virtnet_xdp_get_sq(vi);
1564                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1565                         u64_stats_update_begin(&sq->stats.syncp);
1566                         sq->stats.kicks++;
1567                         u64_stats_update_end(&sq->stats.syncp);
1568                 }
1569                 virtnet_xdp_put_sq(vi, sq);
1570         }
1571
1572         return received;
1573 }
1574
1575 static int virtnet_open(struct net_device *dev)
1576 {
1577         struct virtnet_info *vi = netdev_priv(dev);
1578         int i, err;
1579
1580         for (i = 0; i < vi->max_queue_pairs; i++) {
1581                 if (i < vi->curr_queue_pairs)
1582                         /* Make sure we have some buffers: if oom use wq. */
1583                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1584                                 schedule_delayed_work(&vi->refill, 0);
1585
1586                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1587                 if (err < 0)
1588                         return err;
1589
1590                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1591                                                  MEM_TYPE_PAGE_SHARED, NULL);
1592                 if (err < 0) {
1593                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1594                         return err;
1595                 }
1596
1597                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1598                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1599         }
1600
1601         return 0;
1602 }
1603
1604 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1605 {
1606         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1607         struct virtnet_info *vi = sq->vq->vdev->priv;
1608         unsigned int index = vq2txq(sq->vq);
1609         struct netdev_queue *txq;
1610         int opaque;
1611         bool done;
1612
1613         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1614                 /* We don't need to enable cb for XDP */
1615                 napi_complete_done(napi, 0);
1616                 return 0;
1617         }
1618
1619         txq = netdev_get_tx_queue(vi->dev, index);
1620         __netif_tx_lock(txq, raw_smp_processor_id());
1621         virtqueue_disable_cb(sq->vq);
1622         free_old_xmit_skbs(sq, true);
1623
1624         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1625                 netif_tx_wake_queue(txq);
1626
1627         opaque = virtqueue_enable_cb_prepare(sq->vq);
1628
1629         done = napi_complete_done(napi, 0);
1630
1631         if (!done)
1632                 virtqueue_disable_cb(sq->vq);
1633
1634         __netif_tx_unlock(txq);
1635
1636         if (done) {
1637                 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1638                         if (napi_schedule_prep(napi)) {
1639                                 __netif_tx_lock(txq, raw_smp_processor_id());
1640                                 virtqueue_disable_cb(sq->vq);
1641                                 __netif_tx_unlock(txq);
1642                                 __napi_schedule(napi);
1643                         }
1644                 }
1645         }
1646
1647         return 0;
1648 }
1649
1650 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1651 {
1652         struct virtio_net_hdr_mrg_rxbuf *hdr;
1653         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1654         struct virtnet_info *vi = sq->vq->vdev->priv;
1655         int num_sg;
1656         unsigned hdr_len = vi->hdr_len;
1657         bool can_push;
1658
1659         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1660
1661         can_push = vi->any_header_sg &&
1662                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1663                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1664         /* Even if we can, don't push here yet as this would skew
1665          * csum_start offset below. */
1666         if (can_push)
1667                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1668         else
1669                 hdr = skb_vnet_hdr(skb);
1670
1671         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1672                                     virtio_is_little_endian(vi->vdev), false,
1673                                     0))
1674                 return -EPROTO;
1675
1676         if (vi->mergeable_rx_bufs)
1677                 hdr->num_buffers = 0;
1678
1679         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1680         if (can_push) {
1681                 __skb_push(skb, hdr_len);
1682                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1683                 if (unlikely(num_sg < 0))
1684                         return num_sg;
1685                 /* Pull header back to avoid skew in tx bytes calculations. */
1686                 __skb_pull(skb, hdr_len);
1687         } else {
1688                 sg_set_buf(sq->sg, hdr, hdr_len);
1689                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1690                 if (unlikely(num_sg < 0))
1691                         return num_sg;
1692                 num_sg++;
1693         }
1694         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1695 }
1696
1697 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1698 {
1699         struct virtnet_info *vi = netdev_priv(dev);
1700         int qnum = skb_get_queue_mapping(skb);
1701         struct send_queue *sq = &vi->sq[qnum];
1702         int err;
1703         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1704         bool kick = !netdev_xmit_more();
1705         bool use_napi = sq->napi.weight;
1706
1707         /* Free up any pending old buffers before queueing new ones. */
1708         do {
1709                 if (use_napi)
1710                         virtqueue_disable_cb(sq->vq);
1711
1712                 free_old_xmit_skbs(sq, false);
1713
1714         } while (use_napi && kick &&
1715                unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1716
1717         /* timestamp packet in software */
1718         skb_tx_timestamp(skb);
1719
1720         /* Try to transmit */
1721         err = xmit_skb(sq, skb);
1722
1723         /* This should not happen! */
1724         if (unlikely(err)) {
1725                 dev->stats.tx_fifo_errors++;
1726                 if (net_ratelimit())
1727                         dev_warn(&dev->dev,
1728                                  "Unexpected TXQ (%d) queue failure: %d\n",
1729                                  qnum, err);
1730                 dev->stats.tx_dropped++;
1731                 dev_kfree_skb_any(skb);
1732                 return NETDEV_TX_OK;
1733         }
1734
1735         /* Don't wait up for transmitted skbs to be freed. */
1736         if (!use_napi) {
1737                 skb_orphan(skb);
1738                 nf_reset_ct(skb);
1739         }
1740
1741         /* If running out of space, stop queue to avoid getting packets that we
1742          * are then unable to transmit.
1743          * An alternative would be to force queuing layer to requeue the skb by
1744          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1745          * returned in a normal path of operation: it means that driver is not
1746          * maintaining the TX queue stop/start state properly, and causes
1747          * the stack to do a non-trivial amount of useless work.
1748          * Since most packets only take 1 or 2 ring slots, stopping the queue
1749          * early means 16 slots are typically wasted.
1750          */
1751         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1752                 netif_stop_subqueue(dev, qnum);
1753                 if (!use_napi &&
1754                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1755                         /* More just got used, free them then recheck. */
1756                         free_old_xmit_skbs(sq, false);
1757                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1758                                 netif_start_subqueue(dev, qnum);
1759                                 virtqueue_disable_cb(sq->vq);
1760                         }
1761                 }
1762         }
1763
1764         if (kick || netif_xmit_stopped(txq)) {
1765                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1766                         u64_stats_update_begin(&sq->stats.syncp);
1767                         sq->stats.kicks++;
1768                         u64_stats_update_end(&sq->stats.syncp);
1769                 }
1770         }
1771
1772         return NETDEV_TX_OK;
1773 }
1774
1775 /*
1776  * Send command via the control virtqueue and check status.  Commands
1777  * supported by the hypervisor, as indicated by feature bits, should
1778  * never fail unless improperly formatted.
1779  */
1780 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1781                                  struct scatterlist *out)
1782 {
1783         struct scatterlist *sgs[4], hdr, stat;
1784         unsigned out_num = 0, tmp;
1785         int ret;
1786
1787         /* Caller should know better */
1788         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1789
1790         vi->ctrl->status = ~0;
1791         vi->ctrl->hdr.class = class;
1792         vi->ctrl->hdr.cmd = cmd;
1793         /* Add header */
1794         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1795         sgs[out_num++] = &hdr;
1796
1797         if (out)
1798                 sgs[out_num++] = out;
1799
1800         /* Add return status. */
1801         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1802         sgs[out_num] = &stat;
1803
1804         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1805         ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1806         if (ret < 0) {
1807                 dev_warn(&vi->vdev->dev,
1808                          "Failed to add sgs for command vq: %d\n.", ret);
1809                 return false;
1810         }
1811
1812         if (unlikely(!virtqueue_kick(vi->cvq)))
1813                 return vi->ctrl->status == VIRTIO_NET_OK;
1814
1815         /* Spin for a response, the kick causes an ioport write, trapping
1816          * into the hypervisor, so the request should be handled immediately.
1817          */
1818         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1819                !virtqueue_is_broken(vi->cvq))
1820                 cpu_relax();
1821
1822         return vi->ctrl->status == VIRTIO_NET_OK;
1823 }
1824
1825 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1826 {
1827         struct virtnet_info *vi = netdev_priv(dev);
1828         struct virtio_device *vdev = vi->vdev;
1829         int ret;
1830         struct sockaddr *addr;
1831         struct scatterlist sg;
1832
1833         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1834                 return -EOPNOTSUPP;
1835
1836         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1837         if (!addr)
1838                 return -ENOMEM;
1839
1840         ret = eth_prepare_mac_addr_change(dev, addr);
1841         if (ret)
1842                 goto out;
1843
1844         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1845                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1846                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1847                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1848                         dev_warn(&vdev->dev,
1849                                  "Failed to set mac address by vq command.\n");
1850                         ret = -EINVAL;
1851                         goto out;
1852                 }
1853         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1854                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1855                 unsigned int i;
1856
1857                 /* Naturally, this has an atomicity problem. */
1858                 for (i = 0; i < dev->addr_len; i++)
1859                         virtio_cwrite8(vdev,
1860                                        offsetof(struct virtio_net_config, mac) +
1861                                        i, addr->sa_data[i]);
1862         }
1863
1864         eth_commit_mac_addr_change(dev, p);
1865         ret = 0;
1866
1867 out:
1868         kfree(addr);
1869         return ret;
1870 }
1871
1872 static void virtnet_stats(struct net_device *dev,
1873                           struct rtnl_link_stats64 *tot)
1874 {
1875         struct virtnet_info *vi = netdev_priv(dev);
1876         unsigned int start;
1877         int i;
1878
1879         for (i = 0; i < vi->max_queue_pairs; i++) {
1880                 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
1881                 struct receive_queue *rq = &vi->rq[i];
1882                 struct send_queue *sq = &vi->sq[i];
1883
1884                 do {
1885                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1886                         tpackets = sq->stats.packets;
1887                         tbytes   = sq->stats.bytes;
1888                         terrors  = sq->stats.tx_timeouts;
1889                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1890
1891                 do {
1892                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1893                         rpackets = rq->stats.packets;
1894                         rbytes   = rq->stats.bytes;
1895                         rdrops   = rq->stats.drops;
1896                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1897
1898                 tot->rx_packets += rpackets;
1899                 tot->tx_packets += tpackets;
1900                 tot->rx_bytes   += rbytes;
1901                 tot->tx_bytes   += tbytes;
1902                 tot->rx_dropped += rdrops;
1903                 tot->tx_errors  += terrors;
1904         }
1905
1906         tot->tx_dropped = dev->stats.tx_dropped;
1907         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1908         tot->rx_length_errors = dev->stats.rx_length_errors;
1909         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1910 }
1911
1912 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1913 {
1914         rtnl_lock();
1915         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1916                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1917                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1918         rtnl_unlock();
1919 }
1920
1921 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1922 {
1923         struct scatterlist sg;
1924         struct net_device *dev = vi->dev;
1925
1926         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1927                 return 0;
1928
1929         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1930         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1931
1932         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1933                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1934                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1935                          queue_pairs);
1936                 return -EINVAL;
1937         } else {
1938                 vi->curr_queue_pairs = queue_pairs;
1939                 /* virtnet_open() will refill when device is going to up. */
1940                 if (dev->flags & IFF_UP)
1941                         schedule_delayed_work(&vi->refill, 0);
1942         }
1943
1944         return 0;
1945 }
1946
1947 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1948 {
1949         int err;
1950
1951         rtnl_lock();
1952         err = _virtnet_set_queues(vi, queue_pairs);
1953         rtnl_unlock();
1954         return err;
1955 }
1956
1957 static int virtnet_close(struct net_device *dev)
1958 {
1959         struct virtnet_info *vi = netdev_priv(dev);
1960         int i;
1961
1962         /* Make sure refill_work doesn't re-enable napi! */
1963         cancel_delayed_work_sync(&vi->refill);
1964
1965         for (i = 0; i < vi->max_queue_pairs; i++) {
1966                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1967                 napi_disable(&vi->rq[i].napi);
1968                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1969         }
1970
1971         return 0;
1972 }
1973
1974 static void virtnet_set_rx_mode(struct net_device *dev)
1975 {
1976         struct virtnet_info *vi = netdev_priv(dev);
1977         struct scatterlist sg[2];
1978         struct virtio_net_ctrl_mac *mac_data;
1979         struct netdev_hw_addr *ha;
1980         int uc_count;
1981         int mc_count;
1982         void *buf;
1983         int i;
1984
1985         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1986         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1987                 return;
1988
1989         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1990         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1991
1992         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1993
1994         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1995                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1996                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1997                          vi->ctrl->promisc ? "en" : "dis");
1998
1999         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
2000
2001         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2002                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2003                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2004                          vi->ctrl->allmulti ? "en" : "dis");
2005
2006         uc_count = netdev_uc_count(dev);
2007         mc_count = netdev_mc_count(dev);
2008         /* MAC filter - use one buffer for both lists */
2009         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2010                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2011         mac_data = buf;
2012         if (!buf)
2013                 return;
2014
2015         sg_init_table(sg, 2);
2016
2017         /* Store the unicast list and count in the front of the buffer */
2018         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2019         i = 0;
2020         netdev_for_each_uc_addr(ha, dev)
2021                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2022
2023         sg_set_buf(&sg[0], mac_data,
2024                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2025
2026         /* multicast list and count fill the end */
2027         mac_data = (void *)&mac_data->macs[uc_count][0];
2028
2029         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2030         i = 0;
2031         netdev_for_each_mc_addr(ha, dev)
2032                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2033
2034         sg_set_buf(&sg[1], mac_data,
2035                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2036
2037         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2038                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2039                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2040
2041         kfree(buf);
2042 }
2043
2044 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2045                                    __be16 proto, u16 vid)
2046 {
2047         struct virtnet_info *vi = netdev_priv(dev);
2048         struct scatterlist sg;
2049
2050         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2051         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2052
2053         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2054                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2055                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2056         return 0;
2057 }
2058
2059 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2060                                     __be16 proto, u16 vid)
2061 {
2062         struct virtnet_info *vi = netdev_priv(dev);
2063         struct scatterlist sg;
2064
2065         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2066         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2067
2068         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2069                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2070                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2071         return 0;
2072 }
2073
2074 static void virtnet_clean_affinity(struct virtnet_info *vi)
2075 {
2076         int i;
2077
2078         if (vi->affinity_hint_set) {
2079                 for (i = 0; i < vi->max_queue_pairs; i++) {
2080                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
2081                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
2082                 }
2083
2084                 vi->affinity_hint_set = false;
2085         }
2086 }
2087
2088 static void virtnet_set_affinity(struct virtnet_info *vi)
2089 {
2090         cpumask_var_t mask;
2091         int stragglers;
2092         int group_size;
2093         int i, j, cpu;
2094         int num_cpu;
2095         int stride;
2096
2097         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2098                 virtnet_clean_affinity(vi);
2099                 return;
2100         }
2101
2102         num_cpu = num_online_cpus();
2103         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2104         stragglers = num_cpu >= vi->curr_queue_pairs ?
2105                         num_cpu % vi->curr_queue_pairs :
2106                         0;
2107         cpu = cpumask_next(-1, cpu_online_mask);
2108
2109         for (i = 0; i < vi->curr_queue_pairs; i++) {
2110                 group_size = stride + (i < stragglers ? 1 : 0);
2111
2112                 for (j = 0; j < group_size; j++) {
2113                         cpumask_set_cpu(cpu, mask);
2114                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2115                                                 nr_cpu_ids, false);
2116                 }
2117                 virtqueue_set_affinity(vi->rq[i].vq, mask);
2118                 virtqueue_set_affinity(vi->sq[i].vq, mask);
2119                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2120                 cpumask_clear(mask);
2121         }
2122
2123         vi->affinity_hint_set = true;
2124         free_cpumask_var(mask);
2125 }
2126
2127 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2128 {
2129         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2130                                                    node);
2131         virtnet_set_affinity(vi);
2132         return 0;
2133 }
2134
2135 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2136 {
2137         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2138                                                    node_dead);
2139         virtnet_set_affinity(vi);
2140         return 0;
2141 }
2142
2143 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2144 {
2145         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2146                                                    node);
2147
2148         virtnet_clean_affinity(vi);
2149         return 0;
2150 }
2151
2152 static enum cpuhp_state virtionet_online;
2153
2154 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2155 {
2156         int ret;
2157
2158         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2159         if (ret)
2160                 return ret;
2161         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2162                                                &vi->node_dead);
2163         if (!ret)
2164                 return ret;
2165         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2166         return ret;
2167 }
2168
2169 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2170 {
2171         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2172         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2173                                             &vi->node_dead);
2174 }
2175
2176 static void virtnet_get_ringparam(struct net_device *dev,
2177                                 struct ethtool_ringparam *ring)
2178 {
2179         struct virtnet_info *vi = netdev_priv(dev);
2180
2181         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2182         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2183         ring->rx_pending = ring->rx_max_pending;
2184         ring->tx_pending = ring->tx_max_pending;
2185 }
2186
2187
2188 static void virtnet_get_drvinfo(struct net_device *dev,
2189                                 struct ethtool_drvinfo *info)
2190 {
2191         struct virtnet_info *vi = netdev_priv(dev);
2192         struct virtio_device *vdev = vi->vdev;
2193
2194         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2195         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2196         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2197
2198 }
2199
2200 /* TODO: Eliminate OOO packets during switching */
2201 static int virtnet_set_channels(struct net_device *dev,
2202                                 struct ethtool_channels *channels)
2203 {
2204         struct virtnet_info *vi = netdev_priv(dev);
2205         u16 queue_pairs = channels->combined_count;
2206         int err;
2207
2208         /* We don't support separate rx/tx channels.
2209          * We don't allow setting 'other' channels.
2210          */
2211         if (channels->rx_count || channels->tx_count || channels->other_count)
2212                 return -EINVAL;
2213
2214         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2215                 return -EINVAL;
2216
2217         /* For now we don't support modifying channels while XDP is loaded
2218          * also when XDP is loaded all RX queues have XDP programs so we only
2219          * need to check a single RX queue.
2220          */
2221         if (vi->rq[0].xdp_prog)
2222                 return -EINVAL;
2223
2224         cpus_read_lock();
2225         err = _virtnet_set_queues(vi, queue_pairs);
2226         if (err) {
2227                 cpus_read_unlock();
2228                 goto err;
2229         }
2230         virtnet_set_affinity(vi);
2231         cpus_read_unlock();
2232
2233         netif_set_real_num_tx_queues(dev, queue_pairs);
2234         netif_set_real_num_rx_queues(dev, queue_pairs);
2235  err:
2236         return err;
2237 }
2238
2239 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2240 {
2241         struct virtnet_info *vi = netdev_priv(dev);
2242         unsigned int i, j;
2243         u8 *p = data;
2244
2245         switch (stringset) {
2246         case ETH_SS_STATS:
2247                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2248                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2249                                 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2250                                                 virtnet_rq_stats_desc[j].desc);
2251                 }
2252
2253                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2254                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2255                                 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2256                                                 virtnet_sq_stats_desc[j].desc);
2257                 }
2258                 break;
2259         }
2260 }
2261
2262 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2263 {
2264         struct virtnet_info *vi = netdev_priv(dev);
2265
2266         switch (sset) {
2267         case ETH_SS_STATS:
2268                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2269                                                VIRTNET_SQ_STATS_LEN);
2270         default:
2271                 return -EOPNOTSUPP;
2272         }
2273 }
2274
2275 static void virtnet_get_ethtool_stats(struct net_device *dev,
2276                                       struct ethtool_stats *stats, u64 *data)
2277 {
2278         struct virtnet_info *vi = netdev_priv(dev);
2279         unsigned int idx = 0, start, i, j;
2280         const u8 *stats_base;
2281         size_t offset;
2282
2283         for (i = 0; i < vi->curr_queue_pairs; i++) {
2284                 struct receive_queue *rq = &vi->rq[i];
2285
2286                 stats_base = (u8 *)&rq->stats;
2287                 do {
2288                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2289                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2290                                 offset = virtnet_rq_stats_desc[j].offset;
2291                                 data[idx + j] = *(u64 *)(stats_base + offset);
2292                         }
2293                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2294                 idx += VIRTNET_RQ_STATS_LEN;
2295         }
2296
2297         for (i = 0; i < vi->curr_queue_pairs; i++) {
2298                 struct send_queue *sq = &vi->sq[i];
2299
2300                 stats_base = (u8 *)&sq->stats;
2301                 do {
2302                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2303                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2304                                 offset = virtnet_sq_stats_desc[j].offset;
2305                                 data[idx + j] = *(u64 *)(stats_base + offset);
2306                         }
2307                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2308                 idx += VIRTNET_SQ_STATS_LEN;
2309         }
2310 }
2311
2312 static void virtnet_get_channels(struct net_device *dev,
2313                                  struct ethtool_channels *channels)
2314 {
2315         struct virtnet_info *vi = netdev_priv(dev);
2316
2317         channels->combined_count = vi->curr_queue_pairs;
2318         channels->max_combined = vi->max_queue_pairs;
2319         channels->max_other = 0;
2320         channels->rx_count = 0;
2321         channels->tx_count = 0;
2322         channels->other_count = 0;
2323 }
2324
2325 static int virtnet_set_link_ksettings(struct net_device *dev,
2326                                       const struct ethtool_link_ksettings *cmd)
2327 {
2328         struct virtnet_info *vi = netdev_priv(dev);
2329
2330         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2331                                                   &vi->speed, &vi->duplex);
2332 }
2333
2334 static int virtnet_get_link_ksettings(struct net_device *dev,
2335                                       struct ethtool_link_ksettings *cmd)
2336 {
2337         struct virtnet_info *vi = netdev_priv(dev);
2338
2339         cmd->base.speed = vi->speed;
2340         cmd->base.duplex = vi->duplex;
2341         cmd->base.port = PORT_OTHER;
2342
2343         return 0;
2344 }
2345
2346 static int virtnet_set_coalesce(struct net_device *dev,
2347                                 struct ethtool_coalesce *ec,
2348                                 struct kernel_ethtool_coalesce *kernel_coal,
2349                                 struct netlink_ext_ack *extack)
2350 {
2351         struct virtnet_info *vi = netdev_priv(dev);
2352         int i, napi_weight;
2353
2354         if (ec->tx_max_coalesced_frames > 1 ||
2355             ec->rx_max_coalesced_frames != 1)
2356                 return -EINVAL;
2357
2358         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2359         if (napi_weight ^ vi->sq[0].napi.weight) {
2360                 if (dev->flags & IFF_UP)
2361                         return -EBUSY;
2362                 for (i = 0; i < vi->max_queue_pairs; i++)
2363                         vi->sq[i].napi.weight = napi_weight;
2364         }
2365
2366         return 0;
2367 }
2368
2369 static int virtnet_get_coalesce(struct net_device *dev,
2370                                 struct ethtool_coalesce *ec,
2371                                 struct kernel_ethtool_coalesce *kernel_coal,
2372                                 struct netlink_ext_ack *extack)
2373 {
2374         struct ethtool_coalesce ec_default = {
2375                 .cmd = ETHTOOL_GCOALESCE,
2376                 .rx_max_coalesced_frames = 1,
2377         };
2378         struct virtnet_info *vi = netdev_priv(dev);
2379
2380         memcpy(ec, &ec_default, sizeof(ec_default));
2381
2382         if (vi->sq[0].napi.weight)
2383                 ec->tx_max_coalesced_frames = 1;
2384
2385         return 0;
2386 }
2387
2388 static void virtnet_init_settings(struct net_device *dev)
2389 {
2390         struct virtnet_info *vi = netdev_priv(dev);
2391
2392         vi->speed = SPEED_UNKNOWN;
2393         vi->duplex = DUPLEX_UNKNOWN;
2394 }
2395
2396 static void virtnet_update_settings(struct virtnet_info *vi)
2397 {
2398         u32 speed;
2399         u8 duplex;
2400
2401         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2402                 return;
2403
2404         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2405
2406         if (ethtool_validate_speed(speed))
2407                 vi->speed = speed;
2408
2409         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2410
2411         if (ethtool_validate_duplex(duplex))
2412                 vi->duplex = duplex;
2413 }
2414
2415 static const struct ethtool_ops virtnet_ethtool_ops = {
2416         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2417         .get_drvinfo = virtnet_get_drvinfo,
2418         .get_link = ethtool_op_get_link,
2419         .get_ringparam = virtnet_get_ringparam,
2420         .get_strings = virtnet_get_strings,
2421         .get_sset_count = virtnet_get_sset_count,
2422         .get_ethtool_stats = virtnet_get_ethtool_stats,
2423         .set_channels = virtnet_set_channels,
2424         .get_channels = virtnet_get_channels,
2425         .get_ts_info = ethtool_op_get_ts_info,
2426         .get_link_ksettings = virtnet_get_link_ksettings,
2427         .set_link_ksettings = virtnet_set_link_ksettings,
2428         .set_coalesce = virtnet_set_coalesce,
2429         .get_coalesce = virtnet_get_coalesce,
2430 };
2431
2432 static void virtnet_freeze_down(struct virtio_device *vdev)
2433 {
2434         struct virtnet_info *vi = vdev->priv;
2435         int i;
2436
2437         /* Make sure no work handler is accessing the device */
2438         flush_work(&vi->config_work);
2439
2440         netif_tx_lock_bh(vi->dev);
2441         netif_device_detach(vi->dev);
2442         netif_tx_unlock_bh(vi->dev);
2443         cancel_delayed_work_sync(&vi->refill);
2444
2445         if (netif_running(vi->dev)) {
2446                 for (i = 0; i < vi->max_queue_pairs; i++) {
2447                         napi_disable(&vi->rq[i].napi);
2448                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2449                 }
2450         }
2451 }
2452
2453 static int init_vqs(struct virtnet_info *vi);
2454
2455 static int virtnet_restore_up(struct virtio_device *vdev)
2456 {
2457         struct virtnet_info *vi = vdev->priv;
2458         int err, i;
2459
2460         err = init_vqs(vi);
2461         if (err)
2462                 return err;
2463
2464         virtio_device_ready(vdev);
2465
2466         if (netif_running(vi->dev)) {
2467                 for (i = 0; i < vi->curr_queue_pairs; i++)
2468                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2469                                 schedule_delayed_work(&vi->refill, 0);
2470
2471                 for (i = 0; i < vi->max_queue_pairs; i++) {
2472                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2473                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2474                                                &vi->sq[i].napi);
2475                 }
2476         }
2477
2478         netif_tx_lock_bh(vi->dev);
2479         netif_device_attach(vi->dev);
2480         netif_tx_unlock_bh(vi->dev);
2481         return err;
2482 }
2483
2484 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2485 {
2486         struct scatterlist sg;
2487         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2488
2489         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2490
2491         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2492                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2493                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2494                 return -EINVAL;
2495         }
2496
2497         return 0;
2498 }
2499
2500 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2501 {
2502         u64 offloads = 0;
2503
2504         if (!vi->guest_offloads)
2505                 return 0;
2506
2507         return virtnet_set_guest_offloads(vi, offloads);
2508 }
2509
2510 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2511 {
2512         u64 offloads = vi->guest_offloads;
2513
2514         if (!vi->guest_offloads)
2515                 return 0;
2516
2517         return virtnet_set_guest_offloads(vi, offloads);
2518 }
2519
2520 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2521                            struct netlink_ext_ack *extack)
2522 {
2523         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2524         struct virtnet_info *vi = netdev_priv(dev);
2525         struct bpf_prog *old_prog;
2526         u16 xdp_qp = 0, curr_qp;
2527         int i, err;
2528
2529         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2530             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2531                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2532                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2533                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2534                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2535                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
2536                 return -EOPNOTSUPP;
2537         }
2538
2539         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2540                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2541                 return -EINVAL;
2542         }
2543
2544         if (dev->mtu > max_sz) {
2545                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2546                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2547                 return -EINVAL;
2548         }
2549
2550         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2551         if (prog)
2552                 xdp_qp = nr_cpu_ids;
2553
2554         /* XDP requires extra queues for XDP_TX */
2555         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2556                 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",
2557                                  curr_qp + xdp_qp, vi->max_queue_pairs);
2558                 xdp_qp = 0;
2559         }
2560
2561         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2562         if (!prog && !old_prog)
2563                 return 0;
2564
2565         if (prog)
2566                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2567
2568         /* Make sure NAPI is not using any XDP TX queues for RX. */
2569         if (netif_running(dev)) {
2570                 for (i = 0; i < vi->max_queue_pairs; i++) {
2571                         napi_disable(&vi->rq[i].napi);
2572                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2573                 }
2574         }
2575
2576         if (!prog) {
2577                 for (i = 0; i < vi->max_queue_pairs; i++) {
2578                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2579                         if (i == 0)
2580                                 virtnet_restore_guest_offloads(vi);
2581                 }
2582                 synchronize_net();
2583         }
2584
2585         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2586         if (err)
2587                 goto err;
2588         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2589         vi->xdp_queue_pairs = xdp_qp;
2590
2591         if (prog) {
2592                 vi->xdp_enabled = true;
2593                 for (i = 0; i < vi->max_queue_pairs; i++) {
2594                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2595                         if (i == 0 && !old_prog)
2596                                 virtnet_clear_guest_offloads(vi);
2597                 }
2598         } else {
2599                 vi->xdp_enabled = false;
2600         }
2601
2602         for (i = 0; i < vi->max_queue_pairs; i++) {
2603                 if (old_prog)
2604                         bpf_prog_put(old_prog);
2605                 if (netif_running(dev)) {
2606                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2607                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2608                                                &vi->sq[i].napi);
2609                 }
2610         }
2611
2612         return 0;
2613
2614 err:
2615         if (!prog) {
2616                 virtnet_clear_guest_offloads(vi);
2617                 for (i = 0; i < vi->max_queue_pairs; i++)
2618                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2619         }
2620
2621         if (netif_running(dev)) {
2622                 for (i = 0; i < vi->max_queue_pairs; i++) {
2623                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2624                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2625                                                &vi->sq[i].napi);
2626                 }
2627         }
2628         if (prog)
2629                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2630         return err;
2631 }
2632
2633 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2634 {
2635         switch (xdp->command) {
2636         case XDP_SETUP_PROG:
2637                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2638         default:
2639                 return -EINVAL;
2640         }
2641 }
2642
2643 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2644                                       size_t len)
2645 {
2646         struct virtnet_info *vi = netdev_priv(dev);
2647         int ret;
2648
2649         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2650                 return -EOPNOTSUPP;
2651
2652         ret = snprintf(buf, len, "sby");
2653         if (ret >= len)
2654                 return -EOPNOTSUPP;
2655
2656         return 0;
2657 }
2658
2659 static int virtnet_set_features(struct net_device *dev,
2660                                 netdev_features_t features)
2661 {
2662         struct virtnet_info *vi = netdev_priv(dev);
2663         u64 offloads;
2664         int err;
2665
2666         if ((dev->features ^ features) & NETIF_F_GRO_HW) {
2667                 if (vi->xdp_enabled)
2668                         return -EBUSY;
2669
2670                 if (features & NETIF_F_GRO_HW)
2671                         offloads = vi->guest_offloads_capable;
2672                 else
2673                         offloads = vi->guest_offloads_capable &
2674                                    ~GUEST_OFFLOAD_GRO_HW_MASK;
2675
2676                 err = virtnet_set_guest_offloads(vi, offloads);
2677                 if (err)
2678                         return err;
2679                 vi->guest_offloads = offloads;
2680         }
2681
2682         return 0;
2683 }
2684
2685 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
2686 {
2687         struct virtnet_info *priv = netdev_priv(dev);
2688         struct send_queue *sq = &priv->sq[txqueue];
2689         struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
2690
2691         u64_stats_update_begin(&sq->stats.syncp);
2692         sq->stats.tx_timeouts++;
2693         u64_stats_update_end(&sq->stats.syncp);
2694
2695         netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
2696                    txqueue, sq->name, sq->vq->index, sq->vq->name,
2697                    jiffies_to_usecs(jiffies - txq->trans_start));
2698 }
2699
2700 static const struct net_device_ops virtnet_netdev = {
2701         .ndo_open            = virtnet_open,
2702         .ndo_stop            = virtnet_close,
2703         .ndo_start_xmit      = start_xmit,
2704         .ndo_validate_addr   = eth_validate_addr,
2705         .ndo_set_mac_address = virtnet_set_mac_address,
2706         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2707         .ndo_get_stats64     = virtnet_stats,
2708         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2709         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2710         .ndo_bpf                = virtnet_xdp,
2711         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2712         .ndo_features_check     = passthru_features_check,
2713         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2714         .ndo_set_features       = virtnet_set_features,
2715         .ndo_tx_timeout         = virtnet_tx_timeout,
2716 };
2717
2718 static void virtnet_config_changed_work(struct work_struct *work)
2719 {
2720         struct virtnet_info *vi =
2721                 container_of(work, struct virtnet_info, config_work);
2722         u16 v;
2723
2724         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2725                                  struct virtio_net_config, status, &v) < 0)
2726                 return;
2727
2728         if (v & VIRTIO_NET_S_ANNOUNCE) {
2729                 netdev_notify_peers(vi->dev);
2730                 virtnet_ack_link_announce(vi);
2731         }
2732
2733         /* Ignore unknown (future) status bits */
2734         v &= VIRTIO_NET_S_LINK_UP;
2735
2736         if (vi->status == v)
2737                 return;
2738
2739         vi->status = v;
2740
2741         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2742                 virtnet_update_settings(vi);
2743                 netif_carrier_on(vi->dev);
2744                 netif_tx_wake_all_queues(vi->dev);
2745         } else {
2746                 netif_carrier_off(vi->dev);
2747                 netif_tx_stop_all_queues(vi->dev);
2748         }
2749 }
2750
2751 static void virtnet_config_changed(struct virtio_device *vdev)
2752 {
2753         struct virtnet_info *vi = vdev->priv;
2754
2755         schedule_work(&vi->config_work);
2756 }
2757
2758 static void virtnet_free_queues(struct virtnet_info *vi)
2759 {
2760         int i;
2761
2762         for (i = 0; i < vi->max_queue_pairs; i++) {
2763                 __netif_napi_del(&vi->rq[i].napi);
2764                 __netif_napi_del(&vi->sq[i].napi);
2765         }
2766
2767         /* We called __netif_napi_del(),
2768          * we need to respect an RCU grace period before freeing vi->rq
2769          */
2770         synchronize_net();
2771
2772         kfree(vi->rq);
2773         kfree(vi->sq);
2774         kfree(vi->ctrl);
2775 }
2776
2777 static void _free_receive_bufs(struct virtnet_info *vi)
2778 {
2779         struct bpf_prog *old_prog;
2780         int i;
2781
2782         for (i = 0; i < vi->max_queue_pairs; i++) {
2783                 while (vi->rq[i].pages)
2784                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2785
2786                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2787                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2788                 if (old_prog)
2789                         bpf_prog_put(old_prog);
2790         }
2791 }
2792
2793 static void free_receive_bufs(struct virtnet_info *vi)
2794 {
2795         rtnl_lock();
2796         _free_receive_bufs(vi);
2797         rtnl_unlock();
2798 }
2799
2800 static void free_receive_page_frags(struct virtnet_info *vi)
2801 {
2802         int i;
2803         for (i = 0; i < vi->max_queue_pairs; i++)
2804                 if (vi->rq[i].alloc_frag.page)
2805                         put_page(vi->rq[i].alloc_frag.page);
2806 }
2807
2808 static void free_unused_bufs(struct virtnet_info *vi)
2809 {
2810         void *buf;
2811         int i;
2812
2813         for (i = 0; i < vi->max_queue_pairs; i++) {
2814                 struct virtqueue *vq = vi->sq[i].vq;
2815                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2816                         if (!is_xdp_frame(buf))
2817                                 dev_kfree_skb(buf);
2818                         else
2819                                 xdp_return_frame(ptr_to_xdp(buf));
2820                 }
2821         }
2822
2823         for (i = 0; i < vi->max_queue_pairs; i++) {
2824                 struct virtqueue *vq = vi->rq[i].vq;
2825
2826                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2827                         if (vi->mergeable_rx_bufs) {
2828                                 put_page(virt_to_head_page(buf));
2829                         } else if (vi->big_packets) {
2830                                 give_pages(&vi->rq[i], buf);
2831                         } else {
2832                                 put_page(virt_to_head_page(buf));
2833                         }
2834                 }
2835         }
2836 }
2837
2838 static void virtnet_del_vqs(struct virtnet_info *vi)
2839 {
2840         struct virtio_device *vdev = vi->vdev;
2841
2842         virtnet_clean_affinity(vi);
2843
2844         vdev->config->del_vqs(vdev);
2845
2846         virtnet_free_queues(vi);
2847 }
2848
2849 /* How large should a single buffer be so a queue full of these can fit at
2850  * least one full packet?
2851  * Logic below assumes the mergeable buffer header is used.
2852  */
2853 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2854 {
2855         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2856         unsigned int rq_size = virtqueue_get_vring_size(vq);
2857         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2858         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2859         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2860
2861         return max(max(min_buf_len, hdr_len) - hdr_len,
2862                    (unsigned int)GOOD_PACKET_LEN);
2863 }
2864
2865 static int virtnet_find_vqs(struct virtnet_info *vi)
2866 {
2867         vq_callback_t **callbacks;
2868         struct virtqueue **vqs;
2869         int ret = -ENOMEM;
2870         int i, total_vqs;
2871         const char **names;
2872         bool *ctx;
2873
2874         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2875          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2876          * possible control vq.
2877          */
2878         total_vqs = vi->max_queue_pairs * 2 +
2879                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2880
2881         /* Allocate space for find_vqs parameters */
2882         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2883         if (!vqs)
2884                 goto err_vq;
2885         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2886         if (!callbacks)
2887                 goto err_callback;
2888         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2889         if (!names)
2890                 goto err_names;
2891         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2892                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2893                 if (!ctx)
2894                         goto err_ctx;
2895         } else {
2896                 ctx = NULL;
2897         }
2898
2899         /* Parameters for control virtqueue, if any */
2900         if (vi->has_cvq) {
2901                 callbacks[total_vqs - 1] = NULL;
2902                 names[total_vqs - 1] = "control";
2903         }
2904
2905         /* Allocate/initialize parameters for send/receive virtqueues */
2906         for (i = 0; i < vi->max_queue_pairs; i++) {
2907                 callbacks[rxq2vq(i)] = skb_recv_done;
2908                 callbacks[txq2vq(i)] = skb_xmit_done;
2909                 sprintf(vi->rq[i].name, "input.%d", i);
2910                 sprintf(vi->sq[i].name, "output.%d", i);
2911                 names[rxq2vq(i)] = vi->rq[i].name;
2912                 names[txq2vq(i)] = vi->sq[i].name;
2913                 if (ctx)
2914                         ctx[rxq2vq(i)] = true;
2915         }
2916
2917         ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
2918                                   names, ctx, NULL);
2919         if (ret)
2920                 goto err_find;
2921
2922         if (vi->has_cvq) {
2923                 vi->cvq = vqs[total_vqs - 1];
2924                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2925                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2926         }
2927
2928         for (i = 0; i < vi->max_queue_pairs; i++) {
2929                 vi->rq[i].vq = vqs[rxq2vq(i)];
2930                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2931                 vi->sq[i].vq = vqs[txq2vq(i)];
2932         }
2933
2934         /* run here: ret == 0. */
2935
2936
2937 err_find:
2938         kfree(ctx);
2939 err_ctx:
2940         kfree(names);
2941 err_names:
2942         kfree(callbacks);
2943 err_callback:
2944         kfree(vqs);
2945 err_vq:
2946         return ret;
2947 }
2948
2949 static int virtnet_alloc_queues(struct virtnet_info *vi)
2950 {
2951         int i;
2952
2953         if (vi->has_cvq) {
2954                 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2955                 if (!vi->ctrl)
2956                         goto err_ctrl;
2957         } else {
2958                 vi->ctrl = NULL;
2959         }
2960         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2961         if (!vi->sq)
2962                 goto err_sq;
2963         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2964         if (!vi->rq)
2965                 goto err_rq;
2966
2967         INIT_DELAYED_WORK(&vi->refill, refill_work);
2968         for (i = 0; i < vi->max_queue_pairs; i++) {
2969                 vi->rq[i].pages = NULL;
2970                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2971                                napi_weight);
2972                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2973                                   napi_tx ? napi_weight : 0);
2974
2975                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2976                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2977                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2978
2979                 u64_stats_init(&vi->rq[i].stats.syncp);
2980                 u64_stats_init(&vi->sq[i].stats.syncp);
2981         }
2982
2983         return 0;
2984
2985 err_rq:
2986         kfree(vi->sq);
2987 err_sq:
2988         kfree(vi->ctrl);
2989 err_ctrl:
2990         return -ENOMEM;
2991 }
2992
2993 static int init_vqs(struct virtnet_info *vi)
2994 {
2995         int ret;
2996
2997         /* Allocate send & receive queues */
2998         ret = virtnet_alloc_queues(vi);
2999         if (ret)
3000                 goto err;
3001
3002         ret = virtnet_find_vqs(vi);
3003         if (ret)
3004                 goto err_free;
3005
3006         cpus_read_lock();
3007         virtnet_set_affinity(vi);
3008         cpus_read_unlock();
3009
3010         return 0;
3011
3012 err_free:
3013         virtnet_free_queues(vi);
3014 err:
3015         return ret;
3016 }
3017
3018 #ifdef CONFIG_SYSFS
3019 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
3020                 char *buf)
3021 {
3022         struct virtnet_info *vi = netdev_priv(queue->dev);
3023         unsigned int queue_index = get_netdev_rx_queue_index(queue);
3024         unsigned int headroom = virtnet_get_headroom(vi);
3025         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
3026         struct ewma_pkt_len *avg;
3027
3028         BUG_ON(queue_index >= vi->max_queue_pairs);
3029         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
3030         return sprintf(buf, "%u\n",
3031                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
3032                                        SKB_DATA_ALIGN(headroom + tailroom)));
3033 }
3034
3035 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
3036         __ATTR_RO(mergeable_rx_buffer_size);
3037
3038 static struct attribute *virtio_net_mrg_rx_attrs[] = {
3039         &mergeable_rx_buffer_size_attribute.attr,
3040         NULL
3041 };
3042
3043 static const struct attribute_group virtio_net_mrg_rx_group = {
3044         .name = "virtio_net",
3045         .attrs = virtio_net_mrg_rx_attrs
3046 };
3047 #endif
3048
3049 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3050                                     unsigned int fbit,
3051                                     const char *fname, const char *dname)
3052 {
3053         if (!virtio_has_feature(vdev, fbit))
3054                 return false;
3055
3056         dev_err(&vdev->dev, "device advertises feature %s but not %s",
3057                 fname, dname);
3058
3059         return true;
3060 }
3061
3062 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
3063         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3064
3065 static bool virtnet_validate_features(struct virtio_device *vdev)
3066 {
3067         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3068             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3069                              "VIRTIO_NET_F_CTRL_VQ") ||
3070              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3071                              "VIRTIO_NET_F_CTRL_VQ") ||
3072              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3073                              "VIRTIO_NET_F_CTRL_VQ") ||
3074              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3075              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3076                              "VIRTIO_NET_F_CTRL_VQ"))) {
3077                 return false;
3078         }
3079
3080         return true;
3081 }
3082
3083 #define MIN_MTU ETH_MIN_MTU
3084 #define MAX_MTU ETH_MAX_MTU
3085
3086 static int virtnet_validate(struct virtio_device *vdev)
3087 {
3088         if (!vdev->config->get) {
3089                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3090                         __func__);
3091                 return -EINVAL;
3092         }
3093
3094         if (!virtnet_validate_features(vdev))
3095                 return -EINVAL;
3096
3097         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3098                 int mtu = virtio_cread16(vdev,
3099                                          offsetof(struct virtio_net_config,
3100                                                   mtu));
3101                 if (mtu < MIN_MTU)
3102                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3103         }
3104
3105         return 0;
3106 }
3107
3108 static int virtnet_probe(struct virtio_device *vdev)
3109 {
3110         int i, err = -ENOMEM;
3111         struct net_device *dev;
3112         struct virtnet_info *vi;
3113         u16 max_queue_pairs;
3114         int mtu;
3115
3116         /* Find if host supports multiqueue virtio_net device */
3117         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3118                                    struct virtio_net_config,
3119                                    max_virtqueue_pairs, &max_queue_pairs);
3120
3121         /* We need at least 2 queue's */
3122         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3123             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3124             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3125                 max_queue_pairs = 1;
3126
3127         /* Allocate ourselves a network device with room for our info */
3128         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3129         if (!dev)
3130                 return -ENOMEM;
3131
3132         /* Set up network device as normal. */
3133         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3134                            IFF_TX_SKB_NO_LINEAR;
3135         dev->netdev_ops = &virtnet_netdev;
3136         dev->features = NETIF_F_HIGHDMA;
3137
3138         dev->ethtool_ops = &virtnet_ethtool_ops;
3139         SET_NETDEV_DEV(dev, &vdev->dev);
3140
3141         /* Do we support "hardware" checksums? */
3142         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3143                 /* This opens up the world of extra features. */
3144                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3145                 if (csum)
3146                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3147
3148                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3149                         dev->hw_features |= NETIF_F_TSO
3150                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3151                 }
3152                 /* Individual feature bits: what can host handle? */
3153                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3154                         dev->hw_features |= NETIF_F_TSO;
3155                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3156                         dev->hw_features |= NETIF_F_TSO6;
3157                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3158                         dev->hw_features |= NETIF_F_TSO_ECN;
3159
3160                 dev->features |= NETIF_F_GSO_ROBUST;
3161
3162                 if (gso)
3163                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3164                 /* (!csum && gso) case will be fixed by register_netdev() */
3165         }
3166         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3167                 dev->features |= NETIF_F_RXCSUM;
3168         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3169             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3170                 dev->features |= NETIF_F_GRO_HW;
3171         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3172                 dev->hw_features |= NETIF_F_GRO_HW;
3173
3174         dev->vlan_features = dev->features;
3175
3176         /* MTU range: 68 - 65535 */
3177         dev->min_mtu = MIN_MTU;
3178         dev->max_mtu = MAX_MTU;
3179
3180         /* Configuration may specify what MAC to use.  Otherwise random. */
3181         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
3182                 u8 addr[ETH_ALEN];
3183
3184                 virtio_cread_bytes(vdev,
3185                                    offsetof(struct virtio_net_config, mac),
3186                                    addr, ETH_ALEN);
3187                 eth_hw_addr_set(dev, addr);
3188         } else {
3189                 eth_hw_addr_random(dev);
3190         }
3191
3192         /* Set up our device-specific information */
3193         vi = netdev_priv(dev);
3194         vi->dev = dev;
3195         vi->vdev = vdev;
3196         vdev->priv = vi;
3197
3198         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3199
3200         /* If we can receive ANY GSO packets, we must allocate large ones. */
3201         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3202             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3203             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3204             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3205                 vi->big_packets = true;
3206
3207         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3208                 vi->mergeable_rx_bufs = true;
3209
3210         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3211             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3212                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3213         else
3214                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3215
3216         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3217             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3218                 vi->any_header_sg = true;
3219
3220         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3221                 vi->has_cvq = true;
3222
3223         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3224                 mtu = virtio_cread16(vdev,
3225                                      offsetof(struct virtio_net_config,
3226                                               mtu));
3227                 if (mtu < dev->min_mtu) {
3228                         /* Should never trigger: MTU was previously validated
3229                          * in virtnet_validate.
3230                          */
3231                         dev_err(&vdev->dev,
3232                                 "device MTU appears to have changed it is now %d < %d",
3233                                 mtu, dev->min_mtu);
3234                         err = -EINVAL;
3235                         goto free;
3236                 }
3237
3238                 dev->mtu = mtu;
3239                 dev->max_mtu = mtu;
3240
3241                 /* TODO: size buffers correctly in this case. */
3242                 if (dev->mtu > ETH_DATA_LEN)
3243                         vi->big_packets = true;
3244         }
3245
3246         if (vi->any_header_sg)
3247                 dev->needed_headroom = vi->hdr_len;
3248
3249         /* Enable multiqueue by default */
3250         if (num_online_cpus() >= max_queue_pairs)
3251                 vi->curr_queue_pairs = max_queue_pairs;
3252         else
3253                 vi->curr_queue_pairs = num_online_cpus();
3254         vi->max_queue_pairs = max_queue_pairs;
3255
3256         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3257         err = init_vqs(vi);
3258         if (err)
3259                 goto free;
3260
3261 #ifdef CONFIG_SYSFS
3262         if (vi->mergeable_rx_bufs)
3263                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3264 #endif
3265         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3266         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3267
3268         virtnet_init_settings(dev);
3269
3270         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3271                 vi->failover = net_failover_create(vi->dev);
3272                 if (IS_ERR(vi->failover)) {
3273                         err = PTR_ERR(vi->failover);
3274                         goto free_vqs;
3275                 }
3276         }
3277
3278         err = register_netdev(dev);
3279         if (err) {
3280                 pr_debug("virtio_net: registering device failed\n");
3281                 goto free_failover;
3282         }
3283
3284         virtio_device_ready(vdev);
3285
3286         err = virtnet_cpu_notif_add(vi);
3287         if (err) {
3288                 pr_debug("virtio_net: registering cpu notifier failed\n");
3289                 goto free_unregister_netdev;
3290         }
3291
3292         virtnet_set_queues(vi, vi->curr_queue_pairs);
3293
3294         /* Assume link up if device can't report link status,
3295            otherwise get link status from config. */
3296         netif_carrier_off(dev);
3297         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3298                 schedule_work(&vi->config_work);
3299         } else {
3300                 vi->status = VIRTIO_NET_S_LINK_UP;
3301                 virtnet_update_settings(vi);
3302                 netif_carrier_on(dev);
3303         }
3304
3305         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3306                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3307                         set_bit(guest_offloads[i], &vi->guest_offloads);
3308         vi->guest_offloads_capable = vi->guest_offloads;
3309
3310         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3311                  dev->name, max_queue_pairs);
3312
3313         return 0;
3314
3315 free_unregister_netdev:
3316         vi->vdev->config->reset(vdev);
3317
3318         unregister_netdev(dev);
3319 free_failover:
3320         net_failover_destroy(vi->failover);
3321 free_vqs:
3322         cancel_delayed_work_sync(&vi->refill);
3323         free_receive_page_frags(vi);
3324         virtnet_del_vqs(vi);
3325 free:
3326         free_netdev(dev);
3327         return err;
3328 }
3329
3330 static void remove_vq_common(struct virtnet_info *vi)
3331 {
3332         vi->vdev->config->reset(vi->vdev);
3333
3334         /* Free unused buffers in both send and recv, if any. */
3335         free_unused_bufs(vi);
3336
3337         free_receive_bufs(vi);
3338
3339         free_receive_page_frags(vi);
3340
3341         virtnet_del_vqs(vi);
3342 }
3343
3344 static void virtnet_remove(struct virtio_device *vdev)
3345 {
3346         struct virtnet_info *vi = vdev->priv;
3347
3348         virtnet_cpu_notif_remove(vi);
3349
3350         /* Make sure no work handler is accessing the device. */
3351         flush_work(&vi->config_work);
3352
3353         unregister_netdev(vi->dev);
3354
3355         net_failover_destroy(vi->failover);
3356
3357         remove_vq_common(vi);
3358
3359         free_netdev(vi->dev);
3360 }
3361
3362 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3363 {
3364         struct virtnet_info *vi = vdev->priv;
3365
3366         virtnet_cpu_notif_remove(vi);
3367         virtnet_freeze_down(vdev);
3368         remove_vq_common(vi);
3369
3370         return 0;
3371 }
3372
3373 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3374 {
3375         struct virtnet_info *vi = vdev->priv;
3376         int err;
3377
3378         err = virtnet_restore_up(vdev);
3379         if (err)
3380                 return err;
3381         virtnet_set_queues(vi, vi->curr_queue_pairs);
3382
3383         err = virtnet_cpu_notif_add(vi);
3384         if (err) {
3385                 virtnet_freeze_down(vdev);
3386                 remove_vq_common(vi);
3387                 return err;
3388         }
3389
3390         return 0;
3391 }
3392
3393 static struct virtio_device_id id_table[] = {
3394         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3395         { 0 },
3396 };
3397
3398 #define VIRTNET_FEATURES \
3399         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3400         VIRTIO_NET_F_MAC, \
3401         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3402         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3403         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3404         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3405         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3406         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3407         VIRTIO_NET_F_CTRL_MAC_ADDR, \
3408         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3409         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3410
3411 static unsigned int features[] = {
3412         VIRTNET_FEATURES,
3413 };
3414
3415 static unsigned int features_legacy[] = {
3416         VIRTNET_FEATURES,
3417         VIRTIO_NET_F_GSO,
3418         VIRTIO_F_ANY_LAYOUT,
3419 };
3420
3421 static struct virtio_driver virtio_net_driver = {
3422         .feature_table = features,
3423         .feature_table_size = ARRAY_SIZE(features),
3424         .feature_table_legacy = features_legacy,
3425         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3426         .driver.name =  KBUILD_MODNAME,
3427         .driver.owner = THIS_MODULE,
3428         .id_table =     id_table,
3429         .validate =     virtnet_validate,
3430         .probe =        virtnet_probe,
3431         .remove =       virtnet_remove,
3432         .config_changed = virtnet_config_changed,
3433 #ifdef CONFIG_PM_SLEEP
3434         .freeze =       virtnet_freeze,
3435         .restore =      virtnet_restore,
3436 #endif
3437 };
3438
3439 static __init int virtio_net_driver_init(void)
3440 {
3441         int ret;
3442
3443         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3444                                       virtnet_cpu_online,
3445                                       virtnet_cpu_down_prep);
3446         if (ret < 0)
3447                 goto out;
3448         virtionet_online = ret;
3449         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3450                                       NULL, virtnet_cpu_dead);
3451         if (ret)
3452                 goto err_dead;
3453
3454         ret = register_virtio_driver(&virtio_net_driver);
3455         if (ret)
3456                 goto err_virtio;
3457         return 0;
3458 err_virtio:
3459         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3460 err_dead:
3461         cpuhp_remove_multi_state(virtionet_online);
3462 out:
3463         return ret;
3464 }
3465 module_init(virtio_net_driver_init);
3466
3467 static __exit void virtio_net_driver_exit(void)
3468 {
3469         unregister_virtio_driver(&virtio_net_driver);
3470         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3471         cpuhp_remove_multi_state(virtionet_online);
3472 }
3473 module_exit(virtio_net_driver_exit);
3474
3475 MODULE_DEVICE_TABLE(virtio, id_table);
3476 MODULE_DESCRIPTION("Virtio network driver");
3477 MODULE_LICENSE("GPL");