drivers/net/virtio_net.c

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