drivers/net/virtio_net.c

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