drivers/net/virtio_net.c

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