drivers/net/virtio_net.c

   1 /* A network driver using virtio.
   2  *
   3  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation; either version 2 of the License, or
   8  * (at your option) any later version.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, see <http://www.gnu.org/licenses/>.
  17  */
  18 //#define DEBUG
  19 #include <linux/netdevice.h>
  20 #include <linux/etherdevice.h>
  21 #include <linux/ethtool.h>
  22 #include <linux/module.h>
  23 #include <linux/virtio.h>
  24 #include <linux/virtio_net.h>
  25 #include <linux/bpf.h>
  26 #include <linux/bpf_trace.h>
  27 #include <linux/scatterlist.h>
  28 #include <linux/if_vlan.h>
  29 #include <linux/slab.h>
  30 #include <linux/cpu.h>
  31 #include <linux/average.h>
  32 #include <linux/filter.h>
  33 #include <net/route.h>
  34 #include <net/xdp.h>
  35
  36 static int napi_weight = NAPI_POLL_WEIGHT;
  37 module_param(napi_weight, int, 0444);
  38
  39 static bool csum = true, gso = true, napi_tx;
  40 module_param(csum, bool, 0444);
  41 module_param(gso, bool, 0444);
  42 module_param(napi_tx, bool, 0644);
  43
  44 /* FIXME: MTU in config. */
  45 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  46 #define GOOD_COPY_LEN   128
  47
  48 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  49
  50 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  51 #define VIRTIO_XDP_HEADROOM 256
  52
  53 /* RX packet size EWMA. The average packet size is used to determine the packet
  54  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  55  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  56  * term, transient changes in packet size.
  57  */
  58 DECLARE_EWMA(pkt_len, 0, 64)
  59
  60 #define VIRTNET_DRIVER_VERSION "1.0.0"
  61
  62 static const unsigned long guest_offloads[] = {
  63         VIRTIO_NET_F_GUEST_TSO4,
  64         VIRTIO_NET_F_GUEST_TSO6,
  65         VIRTIO_NET_F_GUEST_ECN,
  66         VIRTIO_NET_F_GUEST_UFO
  67 };
  68
  69 struct virtnet_stat_desc {
  70         char desc[ETH_GSTRING_LEN];
  71         size_t offset;
  72 };
  73
  74 struct virtnet_sq_stats {
  75         struct u64_stats_sync syncp;
  76         u64 packets;
  77         u64 bytes;
  78 };
  79
  80 struct virtnet_rq_stats {
  81         struct u64_stats_sync syncp;
  82         u64 packets;
  83         u64 bytes;
  84 };
  85
  86 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
  87 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
  88
  89 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
  90         { "packets",    VIRTNET_SQ_STAT(packets) },
  91         { "bytes",      VIRTNET_SQ_STAT(bytes) },
  92 };
  93
  94 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
  95         { "packets",    VIRTNET_RQ_STAT(packets) },
  96         { "bytes",      VIRTNET_RQ_STAT(bytes) },
  97 };
  98
  99 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 100 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 101
 102 /* Internal representation of a send virtqueue */
 103 struct send_queue {
 104         /* Virtqueue associated with this send _queue */
 105         struct virtqueue *vq;
 106
 107         /* TX: fragments + linear part + virtio header */
 108         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 109
 110         /* Name of the send queue: output.$index */
 111         char name[40];
 112
 113         struct virtnet_sq_stats stats;
 114
 115         struct napi_struct napi;
 116 };
 117
 118 /* Internal representation of a receive virtqueue */
 119 struct receive_queue {
 120         /* Virtqueue associated with this receive_queue */
 121         struct virtqueue *vq;
 122
 123         struct napi_struct napi;
 124
 125         struct bpf_prog __rcu *xdp_prog;
 126
 127         struct virtnet_rq_stats stats;
 128
 129         /* Chain pages by the private ptr. */
 130         struct page *pages;
 131
 132         /* Average packet length for mergeable receive buffers. */
 133         struct ewma_pkt_len mrg_avg_pkt_len;
 134
 135         /* Page frag for packet buffer allocation. */
 136         struct page_frag alloc_frag;
 137
 138         /* RX: fragments + linear part + virtio header */
 139         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 140
 141         /* Min single buffer size for mergeable buffers case. */
 142         unsigned int min_buf_len;
 143
 144         /* Name of this receive queue: input.$index */
 145         char name[40];
 146
 147         struct xdp_rxq_info xdp_rxq;
 148 };
 149
 150 struct virtnet_info {
 151         struct virtio_device *vdev;
 152         struct virtqueue *cvq;
 153         struct net_device *dev;
 154         struct send_queue *sq;
 155         struct receive_queue *rq;
 156         unsigned int status;
 157
 158         /* Max # of queue pairs supported by the device */
 159         u16 max_queue_pairs;
 160
 161         /* # of queue pairs currently used by the driver */
 162         u16 curr_queue_pairs;
 163
 164         /* # of XDP queue pairs currently used by the driver */
 165         u16 xdp_queue_pairs;
 166
 167         /* I like... big packets and I cannot lie! */
 168         bool big_packets;
 169
 170         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 171         bool mergeable_rx_bufs;
 172
 173         /* Has control virtqueue */
 174         bool has_cvq;
 175
 176         /* Host can handle any s/g split between our header and packet data */
 177         bool any_header_sg;
 178
 179         /* Packet virtio header size */
 180         u8 hdr_len;
 181
 182         /* Work struct for refilling if we run low on memory. */
 183         struct delayed_work refill;
 184
 185         /* Work struct for config space updates */
 186         struct work_struct config_work;
 187
 188         /* Does the affinity hint is set for virtqueues? */
 189         bool affinity_hint_set;
 190
 191         /* CPU hotplug instances for online & dead */
 192         struct hlist_node node;
 193         struct hlist_node node_dead;
 194
 195         /* Control VQ buffers: protected by the rtnl lock */
 196         struct virtio_net_ctrl_hdr ctrl_hdr;
 197         virtio_net_ctrl_ack ctrl_status;
 198         struct virtio_net_ctrl_mq ctrl_mq;
 199         u8 ctrl_promisc;
 200         u8 ctrl_allmulti;
 201         u16 ctrl_vid;
 202         u64 ctrl_offloads;
 203
 204         /* Ethtool settings */
 205         u8 duplex;
 206         u32 speed;
 207
 208         unsigned long guest_offloads;
 209 };
 210
 211 struct padded_vnet_hdr {
 212         struct virtio_net_hdr_mrg_rxbuf hdr;
 213         /*
 214          * hdr is in a separate sg buffer, and data sg buffer shares same page
 215          * with this header sg. This padding makes next sg 16 byte aligned
 216          * after the header.
 217          */
 218         char padding[4];
 219 };
 220
 221 /* Converting between virtqueue no. and kernel tx/rx queue no.
 222  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 223  */
 224 static int vq2txq(struct virtqueue *vq)
 225 {
 226         return (vq->index - 1) / 2;
 227 }
 228
 229 static int txq2vq(int txq)
 230 {
 231         return txq * 2 + 1;
 232 }
 233
 234 static int vq2rxq(struct virtqueue *vq)
 235 {
 236         return vq->index / 2;
 237 }
 238
 239 static int rxq2vq(int rxq)
 240 {
 241         return rxq * 2;
 242 }
 243
 244 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 245 {
 246         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 247 }
 248
 249 /*
 250  * private is used to chain pages for big packets, put the whole
 251  * most recent used list in the beginning for reuse
 252  */
 253 static void give_pages(struct receive_queue *rq, struct page *page)
 254 {
 255         struct page *end;
 256
 257         /* Find end of list, sew whole thing into vi->rq.pages. */
 258         for (end = page; end->private; end = (struct page *)end->private);
 259         end->private = (unsigned long)rq->pages;
 260         rq->pages = page;
 261 }
 262
 263 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 264 {
 265         struct page *p = rq->pages;
 266
 267         if (p) {
 268                 rq->pages = (struct page *)p->private;
 269                 /* clear private here, it is used to chain pages */
 270                 p->private = 0;
 271         } else
 272                 p = alloc_page(gfp_mask);
 273         return p;
 274 }
 275
 276 static void virtqueue_napi_schedule(struct napi_struct *napi,
 277                                     struct virtqueue *vq)
 278 {
 279         if (napi_schedule_prep(napi)) {
 280                 virtqueue_disable_cb(vq);
 281                 __napi_schedule(napi);
 282         }
 283 }
 284
 285 static void virtqueue_napi_complete(struct napi_struct *napi,
 286                                     struct virtqueue *vq, int processed)
 287 {
 288         int opaque;
 289
 290         opaque = virtqueue_enable_cb_prepare(vq);
 291         if (napi_complete_done(napi, processed)) {
 292                 if (unlikely(virtqueue_poll(vq, opaque)))
 293                         virtqueue_napi_schedule(napi, vq);
 294         } else {
 295                 virtqueue_disable_cb(vq);
 296         }
 297 }
 298
 299 static void skb_xmit_done(struct virtqueue *vq)
 300 {
 301         struct virtnet_info *vi = vq->vdev->priv;
 302         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 303
 304         /* Suppress further interrupts. */
 305         virtqueue_disable_cb(vq);
 306
 307         if (napi->weight)
 308                 virtqueue_napi_schedule(napi, vq);
 309         else
 310                 /* We were probably waiting for more output buffers. */
 311                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 312 }
 313
 314 #define MRG_CTX_HEADER_SHIFT 22
 315 static void *mergeable_len_to_ctx(unsigned int truesize,
 316                                   unsigned int headroom)
 317 {
 318         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 319 }
 320
 321 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 322 {
 323         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 324 }
 325
 326 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 327 {
 328         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 329 }
 330
 331 /* Called from bottom half context */
 332 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 333                                    struct receive_queue *rq,
 334                                    struct page *page, unsigned int offset,
 335                                    unsigned int len, unsigned int truesize)
 336 {
 337         struct sk_buff *skb;
 338         struct virtio_net_hdr_mrg_rxbuf *hdr;
 339         unsigned int copy, hdr_len, hdr_padded_len;
 340         char *p;
 341
 342         p = page_address(page) + offset;
 343
 344         /* copy small packet so we can reuse these pages for small data */
 345         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 346         if (unlikely(!skb))
 347                 return NULL;
 348
 349         hdr = skb_vnet_hdr(skb);
 350
 351         hdr_len = vi->hdr_len;
 352         if (vi->mergeable_rx_bufs)
 353                 hdr_padded_len = sizeof(*hdr);
 354         else
 355                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 356
 357         memcpy(hdr, p, hdr_len);
 358
 359         len -= hdr_len;
 360         offset += hdr_padded_len;
 361         p += hdr_padded_len;
 362
 363         copy = len;
 364         if (copy > skb_tailroom(skb))
 365                 copy = skb_tailroom(skb);
 366         skb_put_data(skb, p, copy);
 367
 368         len -= copy;
 369         offset += copy;
 370
 371         if (vi->mergeable_rx_bufs) {
 372                 if (len)
 373                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 374                 else
 375                         put_page(page);
 376                 return skb;
 377         }
 378
 379         /*
 380          * Verify that we can indeed put this data into a skb.
 381          * This is here to handle cases when the device erroneously
 382          * tries to receive more than is possible. This is usually
 383          * the case of a broken device.
 384          */
 385         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 386                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 387                 dev_kfree_skb(skb);
 388                 return NULL;
 389         }
 390         BUG_ON(offset >= PAGE_SIZE);
 391         while (len) {
 392                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 393                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 394                                 frag_size, truesize);
 395                 len -= frag_size;
 396                 page = (struct page *)page->private;
 397                 offset = 0;
 398         }
 399
 400         if (page)
 401                 give_pages(rq, page);
 402
 403         return skb;
 404 }
 405
 406 static void virtnet_xdp_flush(struct net_device *dev)
 407 {
 408         struct virtnet_info *vi = netdev_priv(dev);
 409         struct send_queue *sq;
 410         unsigned int qp;
 411
 412         qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
 413         sq = &vi->sq[qp];
 414
 415         virtqueue_kick(sq->vq);
 416 }
 417
 418 static bool __virtnet_xdp_xmit(struct virtnet_info *vi,
 419                                struct xdp_buff *xdp)
 420 {
 421         struct virtio_net_hdr_mrg_rxbuf *hdr;
 422         unsigned int len;
 423         struct send_queue *sq;
 424         unsigned int qp;
 425         void *xdp_sent;
 426         int err;
 427
 428         qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
 429         sq = &vi->sq[qp];
 430
 431         /* Free up any pending old buffers before queueing new ones. */
 432         while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 433                 struct page *sent_page = virt_to_head_page(xdp_sent);
 434
 435                 put_page(sent_page);
 436         }
 437
 438         xdp->data -= vi->hdr_len;
 439         /* Zero header and leave csum up to XDP layers */
 440         hdr = xdp->data;
 441         memset(hdr, 0, vi->hdr_len);
 442
 443         sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
 444
 445         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp->data, GFP_ATOMIC);
 446         if (unlikely(err)) {
 447                 struct page *page = virt_to_head_page(xdp->data);
 448
 449                 put_page(page);
 450                 return false;
 451         }
 452
 453         return true;
 454 }
 455
 456 static int virtnet_xdp_xmit(struct net_device *dev, struct xdp_buff *xdp)
 457 {
 458         struct virtnet_info *vi = netdev_priv(dev);
 459         bool sent = __virtnet_xdp_xmit(vi, xdp);
 460
 461         if (!sent)
 462                 return -ENOSPC;
 463         return 0;
 464 }
 465
 466 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 467 {
 468         return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
 469 }
 470
 471 /* We copy the packet for XDP in the following cases:
 472  *
 473  * 1) Packet is scattered across multiple rx buffers.
 474  * 2) Headroom space is insufficient.
 475  *
 476  * This is inefficient but it's a temporary condition that
 477  * we hit right after XDP is enabled and until queue is refilled
 478  * with large buffers with sufficient headroom - so it should affect
 479  * at most queue size packets.
 480  * Afterwards, the conditions to enable
 481  * XDP should preclude the underlying device from sending packets
 482  * across multiple buffers (num_buf > 1), and we make sure buffers
 483  * have enough headroom.
 484  */
 485 static struct page *xdp_linearize_page(struct receive_queue *rq,
 486                                        u16 *num_buf,
 487                                        struct page *p,
 488                                        int offset,
 489                                        int page_off,
 490                                        unsigned int *len)
 491 {
 492         struct page *page = alloc_page(GFP_ATOMIC);
 493
 494         if (!page)
 495                 return NULL;
 496
 497         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 498         page_off += *len;
 499
 500         while (--*num_buf) {
 501                 unsigned int buflen;
 502                 void *buf;
 503                 int off;
 504
 505                 buf = virtqueue_get_buf(rq->vq, &buflen);
 506                 if (unlikely(!buf))
 507                         goto err_buf;
 508
 509                 p = virt_to_head_page(buf);
 510                 off = buf - page_address(p);
 511
 512                 /* guard against a misconfigured or uncooperative backend that
 513                  * is sending packet larger than the MTU.
 514                  */
 515                 if ((page_off + buflen) > PAGE_SIZE) {
 516                         put_page(p);
 517                         goto err_buf;
 518                 }
 519
 520                 memcpy(page_address(page) + page_off,
 521                        page_address(p) + off, buflen);
 522                 page_off += buflen;
 523                 put_page(p);
 524         }
 525
 526         /* Headroom does not contribute to packet length */
 527         *len = page_off - VIRTIO_XDP_HEADROOM;
 528         return page;
 529 err_buf:
 530         __free_pages(page, 0);
 531         return NULL;
 532 }
 533
 534 static struct sk_buff *receive_small(struct net_device *dev,
 535                                      struct virtnet_info *vi,
 536                                      struct receive_queue *rq,
 537                                      void *buf, void *ctx,
 538                                      unsigned int len,
 539                                      bool *xdp_xmit)
 540 {
 541         struct sk_buff *skb;
 542         struct bpf_prog *xdp_prog;
 543         unsigned int xdp_headroom = (unsigned long)ctx;
 544         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 545         unsigned int headroom = vi->hdr_len + header_offset;
 546         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 547                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 548         struct page *page = virt_to_head_page(buf);
 549         unsigned int delta = 0, err;
 550         struct page *xdp_page;
 551         len -= vi->hdr_len;
 552
 553         rcu_read_lock();
 554         xdp_prog = rcu_dereference(rq->xdp_prog);
 555         if (xdp_prog) {
 556                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 557                 struct xdp_buff xdp;
 558                 void *orig_data;
 559                 u32 act;
 560
 561                 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
 562                         goto err_xdp;
 563
 564                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 565                         int offset = buf - page_address(page) + header_offset;
 566                         unsigned int tlen = len + vi->hdr_len;
 567                         u16 num_buf = 1;
 568
 569                         xdp_headroom = virtnet_get_headroom(vi);
 570                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 571                         headroom = vi->hdr_len + header_offset;
 572                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 573                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 574                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 575                                                       offset, header_offset,
 576                                                       &tlen);
 577                         if (!xdp_page)
 578                                 goto err_xdp;
 579
 580                         buf = page_address(xdp_page);
 581                         put_page(page);
 582                         page = xdp_page;
 583                 }
 584
 585                 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
 586                 xdp.data = xdp.data_hard_start + xdp_headroom;
 587                 xdp_set_data_meta_invalid(&xdp);
 588                 xdp.data_end = xdp.data + len;
 589                 xdp.rxq = &rq->xdp_rxq;
 590                 orig_data = xdp.data;
 591                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 592
 593                 switch (act) {
 594                 case XDP_PASS:
 595                         /* Recalculate length in case bpf program changed it */
 596                         delta = orig_data - xdp.data;
 597                         break;
 598                 case XDP_TX:
 599                         if (unlikely(!__virtnet_xdp_xmit(vi, &xdp)))
 600                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 601                         else
 602                                 *xdp_xmit = true;
 603                         rcu_read_unlock();
 604                         goto xdp_xmit;
 605                 case XDP_REDIRECT:
 606                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 607                         if (!err)
 608                                 *xdp_xmit = true;
 609                         rcu_read_unlock();
 610                         goto xdp_xmit;
 611                 default:
 612                         bpf_warn_invalid_xdp_action(act);
 613                 case XDP_ABORTED:
 614                         trace_xdp_exception(vi->dev, xdp_prog, act);
 615                 case XDP_DROP:
 616                         goto err_xdp;
 617                 }
 618         }
 619         rcu_read_unlock();
 620
 621         skb = build_skb(buf, buflen);
 622         if (!skb) {
 623                 put_page(page);
 624                 goto err;
 625         }
 626         skb_reserve(skb, headroom - delta);
 627         skb_put(skb, len + delta);
 628         if (!delta) {
 629                 buf += header_offset;
 630                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 631         } /* keep zeroed vnet hdr since packet was changed by bpf */
 632
 633 err:
 634         return skb;
 635
 636 err_xdp:
 637         rcu_read_unlock();
 638         dev->stats.rx_dropped++;
 639         put_page(page);
 640 xdp_xmit:
 641         return NULL;
 642 }
 643
 644 static struct sk_buff *receive_big(struct net_device *dev,
 645                                    struct virtnet_info *vi,
 646                                    struct receive_queue *rq,
 647                                    void *buf,
 648                                    unsigned int len)
 649 {
 650         struct page *page = buf;
 651         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 652
 653         if (unlikely(!skb))
 654                 goto err;
 655
 656         return skb;
 657
 658 err:
 659         dev->stats.rx_dropped++;
 660         give_pages(rq, page);
 661         return NULL;
 662 }
 663
 664 static struct sk_buff *receive_mergeable(struct net_device *dev,
 665                                          struct virtnet_info *vi,
 666                                          struct receive_queue *rq,
 667                                          void *buf,
 668                                          void *ctx,
 669                                          unsigned int len,
 670                                          bool *xdp_xmit)
 671 {
 672         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 673         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 674         struct page *page = virt_to_head_page(buf);
 675         int offset = buf - page_address(page);
 676         struct sk_buff *head_skb, *curr_skb;
 677         struct bpf_prog *xdp_prog;
 678         unsigned int truesize;
 679         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 680         int err;
 681
 682         head_skb = NULL;
 683
 684         rcu_read_lock();
 685         xdp_prog = rcu_dereference(rq->xdp_prog);
 686         if (xdp_prog) {
 687                 struct page *xdp_page;
 688                 struct xdp_buff xdp;
 689                 void *data;
 690                 u32 act;
 691
 692                 /* This happens when rx buffer size is underestimated */
 693                 if (unlikely(num_buf > 1 ||
 694                              headroom < virtnet_get_headroom(vi))) {
 695                         /* linearize data for XDP */
 696                         xdp_page = xdp_linearize_page(rq, &num_buf,
 697                                                       page, offset,
 698                                                       VIRTIO_XDP_HEADROOM,
 699                                                       &len);
 700                         if (!xdp_page)
 701                                 goto err_xdp;
 702                         offset = VIRTIO_XDP_HEADROOM;
 703                 } else {
 704                         xdp_page = page;
 705                 }
 706
 707                 /* Transient failure which in theory could occur if
 708                  * in-flight packets from before XDP was enabled reach
 709                  * the receive path after XDP is loaded. In practice I
 710                  * was not able to create this condition.
 711                  */
 712                 if (unlikely(hdr->hdr.gso_type))
 713                         goto err_xdp;
 714
 715                 /* Allow consuming headroom but reserve enough space to push
 716                  * the descriptor on if we get an XDP_TX return code.
 717                  */
 718                 data = page_address(xdp_page) + offset;
 719                 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
 720                 xdp.data = data + vi->hdr_len;
 721                 xdp_set_data_meta_invalid(&xdp);
 722                 xdp.data_end = xdp.data + (len - vi->hdr_len);
 723                 xdp.rxq = &rq->xdp_rxq;
 724
 725                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 726
 727                 if (act != XDP_PASS)
 728                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 729
 730                 switch (act) {
 731                 case XDP_PASS:
 732                         /* recalculate offset to account for any header
 733                          * adjustments. Note other cases do not build an
 734                          * skb and avoid using offset
 735                          */
 736                         offset = xdp.data -
 737                                         page_address(xdp_page) - vi->hdr_len;
 738
 739                         /* We can only create skb based on xdp_page. */
 740                         if (unlikely(xdp_page != page)) {
 741                                 rcu_read_unlock();
 742                                 put_page(page);
 743                                 head_skb = page_to_skb(vi, rq, xdp_page,
 744                                                        offset, len, PAGE_SIZE);
 745                                 return head_skb;
 746                         }
 747                         break;
 748                 case XDP_TX:
 749                         if (unlikely(!__virtnet_xdp_xmit(vi, &xdp)))
 750                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 751                         else
 752                                 *xdp_xmit = true;
 753                         if (unlikely(xdp_page != page))
 754                                 goto err_xdp;
 755                         rcu_read_unlock();
 756                         goto xdp_xmit;
 757                 case XDP_REDIRECT:
 758                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 759                         if (!err)
 760                                 *xdp_xmit = true;
 761                         rcu_read_unlock();
 762                         goto xdp_xmit;
 763                 default:
 764                         bpf_warn_invalid_xdp_action(act);
 765                 case XDP_ABORTED:
 766                         trace_xdp_exception(vi->dev, xdp_prog, act);
 767                 case XDP_DROP:
 768                         if (unlikely(xdp_page != page))
 769                                 __free_pages(xdp_page, 0);
 770                         goto err_xdp;
 771                 }
 772         }
 773         rcu_read_unlock();
 774
 775         truesize = mergeable_ctx_to_truesize(ctx);
 776         if (unlikely(len > truesize)) {
 777                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 778                          dev->name, len, (unsigned long)ctx);
 779                 dev->stats.rx_length_errors++;
 780                 goto err_skb;
 781         }
 782
 783         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
 784         curr_skb = head_skb;
 785
 786         if (unlikely(!curr_skb))
 787                 goto err_skb;
 788         while (--num_buf) {
 789                 int num_skb_frags;
 790
 791                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
 792                 if (unlikely(!buf)) {
 793                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 794                                  dev->name, num_buf,
 795                                  virtio16_to_cpu(vi->vdev,
 796                                                  hdr->num_buffers));
 797                         dev->stats.rx_length_errors++;
 798                         goto err_buf;
 799                 }
 800
 801                 page = virt_to_head_page(buf);
 802
 803                 truesize = mergeable_ctx_to_truesize(ctx);
 804                 if (unlikely(len > truesize)) {
 805                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 806                                  dev->name, len, (unsigned long)ctx);
 807                         dev->stats.rx_length_errors++;
 808                         goto err_skb;
 809                 }
 810
 811                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 812                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 813                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 814
 815                         if (unlikely(!nskb))
 816                                 goto err_skb;
 817                         if (curr_skb == head_skb)
 818                                 skb_shinfo(curr_skb)->frag_list = nskb;
 819                         else
 820                                 curr_skb->next = nskb;
 821                         curr_skb = nskb;
 822                         head_skb->truesize += nskb->truesize;
 823                         num_skb_frags = 0;
 824                 }
 825                 if (curr_skb != head_skb) {
 826                         head_skb->data_len += len;
 827                         head_skb->len += len;
 828                         head_skb->truesize += truesize;
 829                 }
 830                 offset = buf - page_address(page);
 831                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 832                         put_page(page);
 833                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 834                                              len, truesize);
 835                 } else {
 836                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 837                                         offset, len, truesize);
 838                 }
 839         }
 840
 841         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 842         return head_skb;
 843
 844 err_xdp:
 845         rcu_read_unlock();
 846 err_skb:
 847         put_page(page);
 848         while (--num_buf) {
 849                 buf = virtqueue_get_buf(rq->vq, &len);
 850                 if (unlikely(!buf)) {
 851                         pr_debug("%s: rx error: %d buffers missing\n",
 852                                  dev->name, num_buf);
 853                         dev->stats.rx_length_errors++;
 854                         break;
 855                 }
 856                 page = virt_to_head_page(buf);
 857                 put_page(page);
 858         }
 859 err_buf:
 860         dev->stats.rx_dropped++;
 861         dev_kfree_skb(head_skb);
 862 xdp_xmit:
 863         return NULL;
 864 }
 865
 866 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 867                        void *buf, unsigned int len, void **ctx, bool *xdp_xmit)
 868 {
 869         struct net_device *dev = vi->dev;
 870         struct sk_buff *skb;
 871         struct virtio_net_hdr_mrg_rxbuf *hdr;
 872         int ret;
 873
 874         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 875                 pr_debug("%s: short packet %i\n", dev->name, len);
 876                 dev->stats.rx_length_errors++;
 877                 if (vi->mergeable_rx_bufs) {
 878                         put_page(virt_to_head_page(buf));
 879                 } else if (vi->big_packets) {
 880                         give_pages(rq, buf);
 881                 } else {
 882                         put_page(virt_to_head_page(buf));
 883                 }
 884                 return 0;
 885         }
 886
 887         if (vi->mergeable_rx_bufs)
 888                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit);
 889         else if (vi->big_packets)
 890                 skb = receive_big(dev, vi, rq, buf, len);
 891         else
 892                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit);
 893
 894         if (unlikely(!skb))
 895                 return 0;
 896
 897         hdr = skb_vnet_hdr(skb);
 898
 899         ret = skb->len;
 900
 901         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 902                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 903
 904         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 905                                   virtio_is_little_endian(vi->vdev))) {
 906                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 907                                      dev->name, hdr->hdr.gso_type,
 908                                      hdr->hdr.gso_size);
 909                 goto frame_err;
 910         }
 911
 912         skb->protocol = eth_type_trans(skb, dev);
 913         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 914                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 915
 916         napi_gro_receive(&rq->napi, skb);
 917         return ret;
 918
 919 frame_err:
 920         dev->stats.rx_frame_errors++;
 921         dev_kfree_skb(skb);
 922         return 0;
 923 }
 924
 925 /* Unlike mergeable buffers, all buffers are allocated to the
 926  * same size, except for the headroom. For this reason we do
 927  * not need to use  mergeable_len_to_ctx here - it is enough
 928  * to store the headroom as the context ignoring the truesize.
 929  */
 930 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 931                              gfp_t gfp)
 932 {
 933         struct page_frag *alloc_frag = &rq->alloc_frag;
 934         char *buf;
 935         unsigned int xdp_headroom = virtnet_get_headroom(vi);
 936         void *ctx = (void *)(unsigned long)xdp_headroom;
 937         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
 938         int err;
 939
 940         len = SKB_DATA_ALIGN(len) +
 941               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 942         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 943                 return -ENOMEM;
 944
 945         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 946         get_page(alloc_frag->page);
 947         alloc_frag->offset += len;
 948         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
 949                     vi->hdr_len + GOOD_PACKET_LEN);
 950         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
 951         if (err < 0)
 952                 put_page(virt_to_head_page(buf));
 953         return err;
 954 }
 955
 956 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 957                            gfp_t gfp)
 958 {
 959         struct page *first, *list = NULL;
 960         char *p;
 961         int i, err, offset;
 962
 963         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 964
 965         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 966         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 967                 first = get_a_page(rq, gfp);
 968                 if (!first) {
 969                         if (list)
 970                                 give_pages(rq, list);
 971                         return -ENOMEM;
 972                 }
 973                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 974
 975                 /* chain new page in list head to match sg */
 976                 first->private = (unsigned long)list;
 977                 list = first;
 978         }
 979
 980         first = get_a_page(rq, gfp);
 981         if (!first) {
 982                 give_pages(rq, list);
 983                 return -ENOMEM;
 984         }
 985         p = page_address(first);
 986
 987         /* rq->sg[0], rq->sg[1] share the same page */
 988         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 989         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 990
 991         /* rq->sg[1] for data packet, from offset */
 992         offset = sizeof(struct padded_vnet_hdr);
 993         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 994
 995         /* chain first in list head */
 996         first->private = (unsigned long)list;
 997         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 998                                   first, gfp);
 999         if (err < 0)
1000                 give_pages(rq, first);
1001
1002         return err;
1003 }
1004
1005 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1006                                           struct ewma_pkt_len *avg_pkt_len)
1007 {
1008         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1009         unsigned int len;
1010
1011         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1012                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1013         return ALIGN(len, L1_CACHE_BYTES);
1014 }
1015
1016 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1017                                  struct receive_queue *rq, gfp_t gfp)
1018 {
1019         struct page_frag *alloc_frag = &rq->alloc_frag;
1020         unsigned int headroom = virtnet_get_headroom(vi);
1021         char *buf;
1022         void *ctx;
1023         int err;
1024         unsigned int len, hole;
1025
1026         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len);
1027         if (unlikely(!skb_page_frag_refill(len + headroom, alloc_frag, gfp)))
1028                 return -ENOMEM;
1029
1030         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1031         buf += headroom; /* advance address leaving hole at front of pkt */
1032         get_page(alloc_frag->page);
1033         alloc_frag->offset += len + headroom;
1034         hole = alloc_frag->size - alloc_frag->offset;
1035         if (hole < len + headroom) {
1036                 /* To avoid internal fragmentation, if there is very likely not
1037                  * enough space for another buffer, add the remaining space to
1038                  * the current buffer.
1039                  */
1040                 len += hole;
1041                 alloc_frag->offset += hole;
1042         }
1043
1044         sg_init_one(rq->sg, buf, len);
1045         ctx = mergeable_len_to_ctx(len, headroom);
1046         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1047         if (err < 0)
1048                 put_page(virt_to_head_page(buf));
1049
1050         return err;
1051 }
1052
1053 /*
1054  * Returns false if we couldn't fill entirely (OOM).
1055  *
1056  * Normally run in the receive path, but can also be run from ndo_open
1057  * before we're receiving packets, or from refill_work which is
1058  * careful to disable receiving (using napi_disable).
1059  */
1060 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1061                           gfp_t gfp)
1062 {
1063         int err;
1064         bool oom;
1065
1066         do {
1067                 if (vi->mergeable_rx_bufs)
1068                         err = add_recvbuf_mergeable(vi, rq, gfp);
1069                 else if (vi->big_packets)
1070                         err = add_recvbuf_big(vi, rq, gfp);
1071                 else
1072                         err = add_recvbuf_small(vi, rq, gfp);
1073
1074                 oom = err == -ENOMEM;
1075                 if (err)
1076                         break;
1077         } while (rq->vq->num_free);
1078         virtqueue_kick(rq->vq);
1079         return !oom;
1080 }
1081
1082 static void skb_recv_done(struct virtqueue *rvq)
1083 {
1084         struct virtnet_info *vi = rvq->vdev->priv;
1085         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1086
1087         virtqueue_napi_schedule(&rq->napi, rvq);
1088 }
1089
1090 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1091 {
1092         napi_enable(napi);
1093
1094         /* If all buffers were filled by other side before we napi_enabled, we
1095          * won't get another interrupt, so process any outstanding packets now.
1096          * Call local_bh_enable after to trigger softIRQ processing.
1097          */
1098         local_bh_disable();
1099         virtqueue_napi_schedule(napi, vq);
1100         local_bh_enable();
1101 }
1102
1103 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1104                                    struct virtqueue *vq,
1105                                    struct napi_struct *napi)
1106 {
1107         if (!napi->weight)
1108                 return;
1109
1110         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1111          * enable the feature if this is likely affine with the transmit path.
1112          */
1113         if (!vi->affinity_hint_set) {
1114                 napi->weight = 0;
1115                 return;
1116         }
1117
1118         return virtnet_napi_enable(vq, napi);
1119 }
1120
1121 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1122 {
1123         if (napi->weight)
1124                 napi_disable(napi);
1125 }
1126
1127 static void refill_work(struct work_struct *work)
1128 {
1129         struct virtnet_info *vi =
1130                 container_of(work, struct virtnet_info, refill.work);
1131         bool still_empty;
1132         int i;
1133
1134         for (i = 0; i < vi->curr_queue_pairs; i++) {
1135                 struct receive_queue *rq = &vi->rq[i];
1136
1137                 napi_disable(&rq->napi);
1138                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1139                 virtnet_napi_enable(rq->vq, &rq->napi);
1140
1141                 /* In theory, this can happen: if we don't get any buffers in
1142                  * we will *never* try to fill again.
1143                  */
1144                 if (still_empty)
1145                         schedule_delayed_work(&vi->refill, HZ/2);
1146         }
1147 }
1148
1149 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit)
1150 {
1151         struct virtnet_info *vi = rq->vq->vdev->priv;
1152         unsigned int len, received = 0, bytes = 0;
1153         void *buf;
1154
1155         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1156                 void *ctx;
1157
1158                 while (received < budget &&
1159                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1160                         bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit);
1161                         received++;
1162                 }
1163         } else {
1164                 while (received < budget &&
1165                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1166                         bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit);
1167                         received++;
1168                 }
1169         }
1170
1171         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1172                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1173                         schedule_delayed_work(&vi->refill, 0);
1174         }
1175
1176         u64_stats_update_begin(&rq->stats.syncp);
1177         rq->stats.bytes += bytes;
1178         rq->stats.packets += received;
1179         u64_stats_update_end(&rq->stats.syncp);
1180
1181         return received;
1182 }
1183
1184 static void free_old_xmit_skbs(struct send_queue *sq)
1185 {
1186         struct sk_buff *skb;
1187         unsigned int len;
1188         unsigned int packets = 0;
1189         unsigned int bytes = 0;
1190
1191         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1192                 pr_debug("Sent skb %p\n", skb);
1193
1194                 bytes += skb->len;
1195                 packets++;
1196
1197                 dev_consume_skb_any(skb);
1198         }
1199
1200         /* Avoid overhead when no packets have been processed
1201          * happens when called speculatively from start_xmit.
1202          */
1203         if (!packets)
1204                 return;
1205
1206         u64_stats_update_begin(&sq->stats.syncp);
1207         sq->stats.bytes += bytes;
1208         sq->stats.packets += packets;
1209         u64_stats_update_end(&sq->stats.syncp);
1210 }
1211
1212 static void virtnet_poll_cleantx(struct receive_queue *rq)
1213 {
1214         struct virtnet_info *vi = rq->vq->vdev->priv;
1215         unsigned int index = vq2rxq(rq->vq);
1216         struct send_queue *sq = &vi->sq[index];
1217         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1218
1219         if (!sq->napi.weight)
1220                 return;
1221
1222         if (__netif_tx_trylock(txq)) {
1223                 free_old_xmit_skbs(sq);
1224                 __netif_tx_unlock(txq);
1225         }
1226
1227         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1228                 netif_tx_wake_queue(txq);
1229 }
1230
1231 static int virtnet_poll(struct napi_struct *napi, int budget)
1232 {
1233         struct receive_queue *rq =
1234                 container_of(napi, struct receive_queue, napi);
1235         unsigned int received;
1236         bool xdp_xmit = false;
1237
1238         virtnet_poll_cleantx(rq);
1239
1240         received = virtnet_receive(rq, budget, &xdp_xmit);
1241
1242         /* Out of packets? */
1243         if (received < budget)
1244                 virtqueue_napi_complete(napi, rq->vq, received);
1245
1246         if (xdp_xmit)
1247                 xdp_do_flush_map();
1248
1249         return received;
1250 }
1251
1252 static int virtnet_open(struct net_device *dev)
1253 {
1254         struct virtnet_info *vi = netdev_priv(dev);
1255         int i, err;
1256
1257         for (i = 0; i < vi->max_queue_pairs; i++) {
1258                 if (i < vi->curr_queue_pairs)
1259                         /* Make sure we have some buffers: if oom use wq. */
1260                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1261                                 schedule_delayed_work(&vi->refill, 0);
1262
1263                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1264                 if (err < 0)
1265                         return err;
1266
1267                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1268                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1269         }
1270
1271         return 0;
1272 }
1273
1274 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1275 {
1276         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1277         struct virtnet_info *vi = sq->vq->vdev->priv;
1278         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1279
1280         __netif_tx_lock(txq, raw_smp_processor_id());
1281         free_old_xmit_skbs(sq);
1282         __netif_tx_unlock(txq);
1283
1284         virtqueue_napi_complete(napi, sq->vq, 0);
1285
1286         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1287                 netif_tx_wake_queue(txq);
1288
1289         return 0;
1290 }
1291
1292 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1293 {
1294         struct virtio_net_hdr_mrg_rxbuf *hdr;
1295         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1296         struct virtnet_info *vi = sq->vq->vdev->priv;
1297         int num_sg;
1298         unsigned hdr_len = vi->hdr_len;
1299         bool can_push;
1300
1301         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1302
1303         can_push = vi->any_header_sg &&
1304                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1305                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1306         /* Even if we can, don't push here yet as this would skew
1307          * csum_start offset below. */
1308         if (can_push)
1309                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1310         else
1311                 hdr = skb_vnet_hdr(skb);
1312
1313         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1314                                     virtio_is_little_endian(vi->vdev), false))
1315                 BUG();
1316
1317         if (vi->mergeable_rx_bufs)
1318                 hdr->num_buffers = 0;
1319
1320         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1321         if (can_push) {
1322                 __skb_push(skb, hdr_len);
1323                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1324                 if (unlikely(num_sg < 0))
1325                         return num_sg;
1326                 /* Pull header back to avoid skew in tx bytes calculations. */
1327                 __skb_pull(skb, hdr_len);
1328         } else {
1329                 sg_set_buf(sq->sg, hdr, hdr_len);
1330                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1331                 if (unlikely(num_sg < 0))
1332                         return num_sg;
1333                 num_sg++;
1334         }
1335         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1336 }
1337
1338 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1339 {
1340         struct virtnet_info *vi = netdev_priv(dev);
1341         int qnum = skb_get_queue_mapping(skb);
1342         struct send_queue *sq = &vi->sq[qnum];
1343         int err;
1344         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1345         bool kick = !skb->xmit_more;
1346         bool use_napi = sq->napi.weight;
1347
1348         /* Free up any pending old buffers before queueing new ones. */
1349         free_old_xmit_skbs(sq);
1350
1351         if (use_napi && kick)
1352                 virtqueue_enable_cb_delayed(sq->vq);
1353
1354         /* timestamp packet in software */
1355         skb_tx_timestamp(skb);
1356
1357         /* Try to transmit */
1358         err = xmit_skb(sq, skb);
1359
1360         /* This should not happen! */
1361         if (unlikely(err)) {
1362                 dev->stats.tx_fifo_errors++;
1363                 if (net_ratelimit())
1364                         dev_warn(&dev->dev,
1365                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1366                 dev->stats.tx_dropped++;
1367                 dev_kfree_skb_any(skb);
1368                 return NETDEV_TX_OK;
1369         }
1370
1371         /* Don't wait up for transmitted skbs to be freed. */
1372         if (!use_napi) {
1373                 skb_orphan(skb);
1374                 nf_reset(skb);
1375         }
1376
1377         /* If running out of space, stop queue to avoid getting packets that we
1378          * are then unable to transmit.
1379          * An alternative would be to force queuing layer to requeue the skb by
1380          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1381          * returned in a normal path of operation: it means that driver is not
1382          * maintaining the TX queue stop/start state properly, and causes
1383          * the stack to do a non-trivial amount of useless work.
1384          * Since most packets only take 1 or 2 ring slots, stopping the queue
1385          * early means 16 slots are typically wasted.
1386          */
1387         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1388                 netif_stop_subqueue(dev, qnum);
1389                 if (!use_napi &&
1390                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1391                         /* More just got used, free them then recheck. */
1392                         free_old_xmit_skbs(sq);
1393                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1394                                 netif_start_subqueue(dev, qnum);
1395                                 virtqueue_disable_cb(sq->vq);
1396                         }
1397                 }
1398         }
1399
1400         if (kick || netif_xmit_stopped(txq))
1401                 virtqueue_kick(sq->vq);
1402
1403         return NETDEV_TX_OK;
1404 }
1405
1406 /*
1407  * Send command via the control virtqueue and check status.  Commands
1408  * supported by the hypervisor, as indicated by feature bits, should
1409  * never fail unless improperly formatted.
1410  */
1411 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1412                                  struct scatterlist *out)
1413 {
1414         struct scatterlist *sgs[4], hdr, stat;
1415         unsigned out_num = 0, tmp;
1416
1417         /* Caller should know better */
1418         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1419
1420         vi->ctrl_status = ~0;
1421         vi->ctrl_hdr.class = class;
1422         vi->ctrl_hdr.cmd = cmd;
1423         /* Add header */
1424         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1425         sgs[out_num++] = &hdr;
1426
1427         if (out)
1428                 sgs[out_num++] = out;
1429
1430         /* Add return status. */
1431         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1432         sgs[out_num] = &stat;
1433
1434         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1435         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1436
1437         if (unlikely(!virtqueue_kick(vi->cvq)))
1438                 return vi->ctrl_status == VIRTIO_NET_OK;
1439
1440         /* Spin for a response, the kick causes an ioport write, trapping
1441          * into the hypervisor, so the request should be handled immediately.
1442          */
1443         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1444                !virtqueue_is_broken(vi->cvq))
1445                 cpu_relax();
1446
1447         return vi->ctrl_status == VIRTIO_NET_OK;
1448 }
1449
1450 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1451 {
1452         struct virtnet_info *vi = netdev_priv(dev);
1453         struct virtio_device *vdev = vi->vdev;
1454         int ret;
1455         struct sockaddr *addr;
1456         struct scatterlist sg;
1457
1458         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1459         if (!addr)
1460                 return -ENOMEM;
1461
1462         ret = eth_prepare_mac_addr_change(dev, addr);
1463         if (ret)
1464                 goto out;
1465
1466         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1467                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1468                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1469                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1470                         dev_warn(&vdev->dev,
1471                                  "Failed to set mac address by vq command.\n");
1472                         ret = -EINVAL;
1473                         goto out;
1474                 }
1475         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1476                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1477                 unsigned int i;
1478
1479                 /* Naturally, this has an atomicity problem. */
1480                 for (i = 0; i < dev->addr_len; i++)
1481                         virtio_cwrite8(vdev,
1482                                        offsetof(struct virtio_net_config, mac) +
1483                                        i, addr->sa_data[i]);
1484         }
1485
1486         eth_commit_mac_addr_change(dev, p);
1487         ret = 0;
1488
1489 out:
1490         kfree(addr);
1491         return ret;
1492 }
1493
1494 static void virtnet_stats(struct net_device *dev,
1495                           struct rtnl_link_stats64 *tot)
1496 {
1497         struct virtnet_info *vi = netdev_priv(dev);
1498         unsigned int start;
1499         int i;
1500
1501         for (i = 0; i < vi->max_queue_pairs; i++) {
1502                 u64 tpackets, tbytes, rpackets, rbytes;
1503                 struct receive_queue *rq = &vi->rq[i];
1504                 struct send_queue *sq = &vi->sq[i];
1505
1506                 do {
1507                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1508                         tpackets = sq->stats.packets;
1509                         tbytes   = sq->stats.bytes;
1510                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1511
1512                 do {
1513                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1514                         rpackets = rq->stats.packets;
1515                         rbytes   = rq->stats.bytes;
1516                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1517
1518                 tot->rx_packets += rpackets;
1519                 tot->tx_packets += tpackets;
1520                 tot->rx_bytes   += rbytes;
1521                 tot->tx_bytes   += tbytes;
1522         }
1523
1524         tot->tx_dropped = dev->stats.tx_dropped;
1525         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1526         tot->rx_dropped = dev->stats.rx_dropped;
1527         tot->rx_length_errors = dev->stats.rx_length_errors;
1528         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1529 }
1530
1531 #ifdef CONFIG_NET_POLL_CONTROLLER
1532 static void virtnet_netpoll(struct net_device *dev)
1533 {
1534         struct virtnet_info *vi = netdev_priv(dev);
1535         int i;
1536
1537         for (i = 0; i < vi->curr_queue_pairs; i++)
1538                 napi_schedule(&vi->rq[i].napi);
1539 }
1540 #endif
1541
1542 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1543 {
1544         rtnl_lock();
1545         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1546                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1547                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1548         rtnl_unlock();
1549 }
1550
1551 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1552 {
1553         struct scatterlist sg;
1554         struct net_device *dev = vi->dev;
1555
1556         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1557                 return 0;
1558
1559         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1560         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1561
1562         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1563                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1564                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1565                          queue_pairs);
1566                 return -EINVAL;
1567         } else {
1568                 vi->curr_queue_pairs = queue_pairs;
1569                 /* virtnet_open() will refill when device is going to up. */
1570                 if (dev->flags & IFF_UP)
1571                         schedule_delayed_work(&vi->refill, 0);
1572         }
1573
1574         return 0;
1575 }
1576
1577 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1578 {
1579         int err;
1580
1581         rtnl_lock();
1582         err = _virtnet_set_queues(vi, queue_pairs);
1583         rtnl_unlock();
1584         return err;
1585 }
1586
1587 static int virtnet_close(struct net_device *dev)
1588 {
1589         struct virtnet_info *vi = netdev_priv(dev);
1590         int i;
1591
1592         /* Make sure refill_work doesn't re-enable napi! */
1593         cancel_delayed_work_sync(&vi->refill);
1594
1595         for (i = 0; i < vi->max_queue_pairs; i++) {
1596                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1597                 napi_disable(&vi->rq[i].napi);
1598                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1599         }
1600
1601         return 0;
1602 }
1603
1604 static void virtnet_set_rx_mode(struct net_device *dev)
1605 {
1606         struct virtnet_info *vi = netdev_priv(dev);
1607         struct scatterlist sg[2];
1608         struct virtio_net_ctrl_mac *mac_data;
1609         struct netdev_hw_addr *ha;
1610         int uc_count;
1611         int mc_count;
1612         void *buf;
1613         int i;
1614
1615         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1616         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1617                 return;
1618
1619         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1620         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1621
1622         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1623
1624         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1625                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1626                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1627                          vi->ctrl_promisc ? "en" : "dis");
1628
1629         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1630
1631         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1632                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1633                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1634                          vi->ctrl_allmulti ? "en" : "dis");
1635
1636         uc_count = netdev_uc_count(dev);
1637         mc_count = netdev_mc_count(dev);
1638         /* MAC filter - use one buffer for both lists */
1639         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1640                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1641         mac_data = buf;
1642         if (!buf)
1643                 return;
1644
1645         sg_init_table(sg, 2);
1646
1647         /* Store the unicast list and count in the front of the buffer */
1648         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1649         i = 0;
1650         netdev_for_each_uc_addr(ha, dev)
1651                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1652
1653         sg_set_buf(&sg[0], mac_data,
1654                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1655
1656         /* multicast list and count fill the end */
1657         mac_data = (void *)&mac_data->macs[uc_count][0];
1658
1659         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1660         i = 0;
1661         netdev_for_each_mc_addr(ha, dev)
1662                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1663
1664         sg_set_buf(&sg[1], mac_data,
1665                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1666
1667         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1668                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1669                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1670
1671         kfree(buf);
1672 }
1673
1674 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1675                                    __be16 proto, u16 vid)
1676 {
1677         struct virtnet_info *vi = netdev_priv(dev);
1678         struct scatterlist sg;
1679
1680         vi->ctrl_vid = vid;
1681         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1682
1683         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1684                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1685                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1686         return 0;
1687 }
1688
1689 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1690                                     __be16 proto, u16 vid)
1691 {
1692         struct virtnet_info *vi = netdev_priv(dev);
1693         struct scatterlist sg;
1694
1695         vi->ctrl_vid = vid;
1696         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1697
1698         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1699                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1700                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1701         return 0;
1702 }
1703
1704 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1705 {
1706         int i;
1707
1708         if (vi->affinity_hint_set) {
1709                 for (i = 0; i < vi->max_queue_pairs; i++) {
1710                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1711                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1712                 }
1713
1714                 vi->affinity_hint_set = false;
1715         }
1716 }
1717
1718 static void virtnet_set_affinity(struct virtnet_info *vi)
1719 {
1720         int i;
1721         int cpu;
1722
1723         /* In multiqueue mode, when the number of cpu is equal to the number of
1724          * queue pairs, we let the queue pairs to be private to one cpu by
1725          * setting the affinity hint to eliminate the contention.
1726          */
1727         if (vi->curr_queue_pairs == 1 ||
1728             vi->max_queue_pairs != num_online_cpus()) {
1729                 virtnet_clean_affinity(vi, -1);
1730                 return;
1731         }
1732
1733         i = 0;
1734         for_each_online_cpu(cpu) {
1735                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1736                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1737                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1738                 i++;
1739         }
1740
1741         vi->affinity_hint_set = true;
1742 }
1743
1744 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1745 {
1746         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1747                                                    node);
1748         virtnet_set_affinity(vi);
1749         return 0;
1750 }
1751
1752 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1753 {
1754         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1755                                                    node_dead);
1756         virtnet_set_affinity(vi);
1757         return 0;
1758 }
1759
1760 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1761 {
1762         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1763                                                    node);
1764
1765         virtnet_clean_affinity(vi, cpu);
1766         return 0;
1767 }
1768
1769 static enum cpuhp_state virtionet_online;
1770
1771 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1772 {
1773         int ret;
1774
1775         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1776         if (ret)
1777                 return ret;
1778         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1779                                                &vi->node_dead);
1780         if (!ret)
1781                 return ret;
1782         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1783         return ret;
1784 }
1785
1786 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1787 {
1788         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1789         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1790                                             &vi->node_dead);
1791 }
1792
1793 static void virtnet_get_ringparam(struct net_device *dev,
1794                                 struct ethtool_ringparam *ring)
1795 {
1796         struct virtnet_info *vi = netdev_priv(dev);
1797
1798         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1799         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1800         ring->rx_pending = ring->rx_max_pending;
1801         ring->tx_pending = ring->tx_max_pending;
1802 }
1803
1804
1805 static void virtnet_get_drvinfo(struct net_device *dev,
1806                                 struct ethtool_drvinfo *info)
1807 {
1808         struct virtnet_info *vi = netdev_priv(dev);
1809         struct virtio_device *vdev = vi->vdev;
1810
1811         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1812         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1813         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1814
1815 }
1816
1817 /* TODO: Eliminate OOO packets during switching */
1818 static int virtnet_set_channels(struct net_device *dev,
1819                                 struct ethtool_channels *channels)
1820 {
1821         struct virtnet_info *vi = netdev_priv(dev);
1822         u16 queue_pairs = channels->combined_count;
1823         int err;
1824
1825         /* We don't support separate rx/tx channels.
1826          * We don't allow setting 'other' channels.
1827          */
1828         if (channels->rx_count || channels->tx_count || channels->other_count)
1829                 return -EINVAL;
1830
1831         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1832                 return -EINVAL;
1833
1834         /* For now we don't support modifying channels while XDP is loaded
1835          * also when XDP is loaded all RX queues have XDP programs so we only
1836          * need to check a single RX queue.
1837          */
1838         if (vi->rq[0].xdp_prog)
1839                 return -EINVAL;
1840
1841         get_online_cpus();
1842         err = _virtnet_set_queues(vi, queue_pairs);
1843         if (!err) {
1844                 netif_set_real_num_tx_queues(dev, queue_pairs);
1845                 netif_set_real_num_rx_queues(dev, queue_pairs);
1846
1847                 virtnet_set_affinity(vi);
1848         }
1849         put_online_cpus();
1850
1851         return err;
1852 }
1853
1854 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
1855 {
1856         struct virtnet_info *vi = netdev_priv(dev);
1857         char *p = (char *)data;
1858         unsigned int i, j;
1859
1860         switch (stringset) {
1861         case ETH_SS_STATS:
1862                 for (i = 0; i < vi->curr_queue_pairs; i++) {
1863                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1864                                 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
1865                                          i, virtnet_rq_stats_desc[j].desc);
1866                                 p += ETH_GSTRING_LEN;
1867                         }
1868                 }
1869
1870                 for (i = 0; i < vi->curr_queue_pairs; i++) {
1871                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1872                                 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
1873                                          i, virtnet_sq_stats_desc[j].desc);
1874                                 p += ETH_GSTRING_LEN;
1875                         }
1876                 }
1877                 break;
1878         }
1879 }
1880
1881 static int virtnet_get_sset_count(struct net_device *dev, int sset)
1882 {
1883         struct virtnet_info *vi = netdev_priv(dev);
1884
1885         switch (sset) {
1886         case ETH_SS_STATS:
1887                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
1888                                                VIRTNET_SQ_STATS_LEN);
1889         default:
1890                 return -EOPNOTSUPP;
1891         }
1892 }
1893
1894 static void virtnet_get_ethtool_stats(struct net_device *dev,
1895                                       struct ethtool_stats *stats, u64 *data)
1896 {
1897         struct virtnet_info *vi = netdev_priv(dev);
1898         unsigned int idx = 0, start, i, j;
1899         const u8 *stats_base;
1900         size_t offset;
1901
1902         for (i = 0; i < vi->curr_queue_pairs; i++) {
1903                 struct receive_queue *rq = &vi->rq[i];
1904
1905                 stats_base = (u8 *)&rq->stats;
1906                 do {
1907                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1908                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1909                                 offset = virtnet_rq_stats_desc[j].offset;
1910                                 data[idx + j] = *(u64 *)(stats_base + offset);
1911                         }
1912                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1913                 idx += VIRTNET_RQ_STATS_LEN;
1914         }
1915
1916         for (i = 0; i < vi->curr_queue_pairs; i++) {
1917                 struct send_queue *sq = &vi->sq[i];
1918
1919                 stats_base = (u8 *)&sq->stats;
1920                 do {
1921                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1922                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1923                                 offset = virtnet_sq_stats_desc[j].offset;
1924                                 data[idx + j] = *(u64 *)(stats_base + offset);
1925                         }
1926                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1927                 idx += VIRTNET_SQ_STATS_LEN;
1928         }
1929 }
1930
1931 static void virtnet_get_channels(struct net_device *dev,
1932                                  struct ethtool_channels *channels)
1933 {
1934         struct virtnet_info *vi = netdev_priv(dev);
1935
1936         channels->combined_count = vi->curr_queue_pairs;
1937         channels->max_combined = vi->max_queue_pairs;
1938         channels->max_other = 0;
1939         channels->rx_count = 0;
1940         channels->tx_count = 0;
1941         channels->other_count = 0;
1942 }
1943
1944 /* Check if the user is trying to change anything besides speed/duplex */
1945 static bool
1946 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1947 {
1948         struct ethtool_link_ksettings diff1 = *cmd;
1949         struct ethtool_link_ksettings diff2 = {};
1950
1951         /* cmd is always set so we need to clear it, validate the port type
1952          * and also without autonegotiation we can ignore advertising
1953          */
1954         diff1.base.speed = 0;
1955         diff2.base.port = PORT_OTHER;
1956         ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
1957         diff1.base.duplex = 0;
1958         diff1.base.cmd = 0;
1959         diff1.base.link_mode_masks_nwords = 0;
1960
1961         return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
1962                 bitmap_empty(diff1.link_modes.supported,
1963                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1964                 bitmap_empty(diff1.link_modes.advertising,
1965                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1966                 bitmap_empty(diff1.link_modes.lp_advertising,
1967                              __ETHTOOL_LINK_MODE_MASK_NBITS);
1968 }
1969
1970 static int virtnet_set_link_ksettings(struct net_device *dev,
1971                                       const struct ethtool_link_ksettings *cmd)
1972 {
1973         struct virtnet_info *vi = netdev_priv(dev);
1974         u32 speed;
1975
1976         speed = cmd->base.speed;
1977         /* don't allow custom speed and duplex */
1978         if (!ethtool_validate_speed(speed) ||
1979             !ethtool_validate_duplex(cmd->base.duplex) ||
1980             !virtnet_validate_ethtool_cmd(cmd))
1981                 return -EINVAL;
1982         vi->speed = speed;
1983         vi->duplex = cmd->base.duplex;
1984
1985         return 0;
1986 }
1987
1988 static int virtnet_get_link_ksettings(struct net_device *dev,
1989                                       struct ethtool_link_ksettings *cmd)
1990 {
1991         struct virtnet_info *vi = netdev_priv(dev);
1992
1993         cmd->base.speed = vi->speed;
1994         cmd->base.duplex = vi->duplex;
1995         cmd->base.port = PORT_OTHER;
1996
1997         return 0;
1998 }
1999
2000 static void virtnet_init_settings(struct net_device *dev)
2001 {
2002         struct virtnet_info *vi = netdev_priv(dev);
2003
2004         vi->speed = SPEED_UNKNOWN;
2005         vi->duplex = DUPLEX_UNKNOWN;
2006 }
2007
2008 static void virtnet_update_settings(struct virtnet_info *vi)
2009 {
2010         u32 speed;
2011         u8 duplex;
2012
2013         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2014                 return;
2015
2016         speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2017                                                   speed));
2018         if (ethtool_validate_speed(speed))
2019                 vi->speed = speed;
2020         duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2021                                                   duplex));
2022         if (ethtool_validate_duplex(duplex))
2023                 vi->duplex = duplex;
2024 }
2025
2026 static const struct ethtool_ops virtnet_ethtool_ops = {
2027         .get_drvinfo = virtnet_get_drvinfo,
2028         .get_link = ethtool_op_get_link,
2029         .get_ringparam = virtnet_get_ringparam,
2030         .get_strings = virtnet_get_strings,
2031         .get_sset_count = virtnet_get_sset_count,
2032         .get_ethtool_stats = virtnet_get_ethtool_stats,
2033         .set_channels = virtnet_set_channels,
2034         .get_channels = virtnet_get_channels,
2035         .get_ts_info = ethtool_op_get_ts_info,
2036         .get_link_ksettings = virtnet_get_link_ksettings,
2037         .set_link_ksettings = virtnet_set_link_ksettings,
2038 };
2039
2040 static void virtnet_freeze_down(struct virtio_device *vdev)
2041 {
2042         struct virtnet_info *vi = vdev->priv;
2043         int i;
2044
2045         /* Make sure no work handler is accessing the device */
2046         flush_work(&vi->config_work);
2047
2048         netif_device_detach(vi->dev);
2049         netif_tx_disable(vi->dev);
2050         cancel_delayed_work_sync(&vi->refill);
2051
2052         if (netif_running(vi->dev)) {
2053                 for (i = 0; i < vi->max_queue_pairs; i++) {
2054                         napi_disable(&vi->rq[i].napi);
2055                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2056                 }
2057         }
2058 }
2059
2060 static int init_vqs(struct virtnet_info *vi);
2061
2062 static int virtnet_restore_up(struct virtio_device *vdev)
2063 {
2064         struct virtnet_info *vi = vdev->priv;
2065         int err, i;
2066
2067         err = init_vqs(vi);
2068         if (err)
2069                 return err;
2070
2071         virtio_device_ready(vdev);
2072
2073         if (netif_running(vi->dev)) {
2074                 for (i = 0; i < vi->curr_queue_pairs; i++)
2075                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2076                                 schedule_delayed_work(&vi->refill, 0);
2077
2078                 for (i = 0; i < vi->max_queue_pairs; i++) {
2079                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2080                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2081                                                &vi->sq[i].napi);
2082                 }
2083         }
2084
2085         netif_device_attach(vi->dev);
2086         return err;
2087 }
2088
2089 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2090 {
2091         struct scatterlist sg;
2092         vi->ctrl_offloads = cpu_to_virtio64(vi->vdev, offloads);
2093
2094         sg_init_one(&sg, &vi->ctrl_offloads, sizeof(vi->ctrl_offloads));
2095
2096         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2097                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2098                 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2099                 return -EINVAL;
2100         }
2101
2102         return 0;
2103 }
2104
2105 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2106 {
2107         u64 offloads = 0;
2108
2109         if (!vi->guest_offloads)
2110                 return 0;
2111
2112         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2113                 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2114
2115         return virtnet_set_guest_offloads(vi, offloads);
2116 }
2117
2118 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2119 {
2120         u64 offloads = vi->guest_offloads;
2121
2122         if (!vi->guest_offloads)
2123                 return 0;
2124         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2125                 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2126
2127         return virtnet_set_guest_offloads(vi, offloads);
2128 }
2129
2130 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2131                            struct netlink_ext_ack *extack)
2132 {
2133         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2134         struct virtnet_info *vi = netdev_priv(dev);
2135         struct bpf_prog *old_prog;
2136         u16 xdp_qp = 0, curr_qp;
2137         int i, err;
2138
2139         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2140             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2141                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2142                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2143                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
2144                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
2145                 return -EOPNOTSUPP;
2146         }
2147
2148         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2149                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2150                 return -EINVAL;
2151         }
2152
2153         if (dev->mtu > max_sz) {
2154                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2155                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2156                 return -EINVAL;
2157         }
2158
2159         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2160         if (prog)
2161                 xdp_qp = nr_cpu_ids;
2162
2163         /* XDP requires extra queues for XDP_TX */
2164         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2165                 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2166                 netdev_warn(dev, "request %i queues but max is %i\n",
2167                             curr_qp + xdp_qp, vi->max_queue_pairs);
2168                 return -ENOMEM;
2169         }
2170
2171         if (prog) {
2172                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2173                 if (IS_ERR(prog))
2174                         return PTR_ERR(prog);
2175         }
2176
2177         /* Make sure NAPI is not using any XDP TX queues for RX. */
2178         for (i = 0; i < vi->max_queue_pairs; i++)
2179                 napi_disable(&vi->rq[i].napi);
2180
2181         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2182         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2183         if (err)
2184                 goto err;
2185         vi->xdp_queue_pairs = xdp_qp;
2186
2187         for (i = 0; i < vi->max_queue_pairs; i++) {
2188                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2189                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2190                 if (i == 0) {
2191                         if (!old_prog)
2192                                 virtnet_clear_guest_offloads(vi);
2193                         if (!prog)
2194                                 virtnet_restore_guest_offloads(vi);
2195                 }
2196                 if (old_prog)
2197                         bpf_prog_put(old_prog);
2198                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2199         }
2200
2201         return 0;
2202
2203 err:
2204         for (i = 0; i < vi->max_queue_pairs; i++)
2205                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2206         if (prog)
2207                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2208         return err;
2209 }
2210
2211 static u32 virtnet_xdp_query(struct net_device *dev)
2212 {
2213         struct virtnet_info *vi = netdev_priv(dev);
2214         const struct bpf_prog *xdp_prog;
2215         int i;
2216
2217         for (i = 0; i < vi->max_queue_pairs; i++) {
2218                 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2219                 if (xdp_prog)
2220                         return xdp_prog->aux->id;
2221         }
2222         return 0;
2223 }
2224
2225 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2226 {
2227         switch (xdp->command) {
2228         case XDP_SETUP_PROG:
2229                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2230         case XDP_QUERY_PROG:
2231                 xdp->prog_id = virtnet_xdp_query(dev);
2232                 xdp->prog_attached = !!xdp->prog_id;
2233                 return 0;
2234         default:
2235                 return -EINVAL;
2236         }
2237 }
2238
2239 static const struct net_device_ops virtnet_netdev = {
2240         .ndo_open            = virtnet_open,
2241         .ndo_stop            = virtnet_close,
2242         .ndo_start_xmit      = start_xmit,
2243         .ndo_validate_addr   = eth_validate_addr,
2244         .ndo_set_mac_address = virtnet_set_mac_address,
2245         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2246         .ndo_get_stats64     = virtnet_stats,
2247         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2248         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2249 #ifdef CONFIG_NET_POLL_CONTROLLER
2250         .ndo_poll_controller = virtnet_netpoll,
2251 #endif
2252         .ndo_bpf                = virtnet_xdp,
2253         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2254         .ndo_xdp_flush          = virtnet_xdp_flush,
2255         .ndo_features_check     = passthru_features_check,
2256 };
2257
2258 static void virtnet_config_changed_work(struct work_struct *work)
2259 {
2260         struct virtnet_info *vi =
2261                 container_of(work, struct virtnet_info, config_work);
2262         u16 v;
2263
2264         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2265                                  struct virtio_net_config, status, &v) < 0)
2266                 return;
2267
2268         if (v & VIRTIO_NET_S_ANNOUNCE) {
2269                 netdev_notify_peers(vi->dev);
2270                 virtnet_ack_link_announce(vi);
2271         }
2272
2273         /* Ignore unknown (future) status bits */
2274         v &= VIRTIO_NET_S_LINK_UP;
2275
2276         if (vi->status == v)
2277                 return;
2278
2279         vi->status = v;
2280
2281         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2282                 virtnet_update_settings(vi);
2283                 netif_carrier_on(vi->dev);
2284                 netif_tx_wake_all_queues(vi->dev);
2285         } else {
2286                 netif_carrier_off(vi->dev);
2287                 netif_tx_stop_all_queues(vi->dev);
2288         }
2289 }
2290
2291 static void virtnet_config_changed(struct virtio_device *vdev)
2292 {
2293         struct virtnet_info *vi = vdev->priv;
2294
2295         schedule_work(&vi->config_work);
2296 }
2297
2298 static void virtnet_free_queues(struct virtnet_info *vi)
2299 {
2300         int i;
2301
2302         for (i = 0; i < vi->max_queue_pairs; i++) {
2303                 napi_hash_del(&vi->rq[i].napi);
2304                 netif_napi_del(&vi->rq[i].napi);
2305                 netif_napi_del(&vi->sq[i].napi);
2306         }
2307
2308         /* We called napi_hash_del() before netif_napi_del(),
2309          * we need to respect an RCU grace period before freeing vi->rq
2310          */
2311         synchronize_net();
2312
2313         kfree(vi->rq);
2314         kfree(vi->sq);
2315 }
2316
2317 static void _free_receive_bufs(struct virtnet_info *vi)
2318 {
2319         struct bpf_prog *old_prog;
2320         int i;
2321
2322         for (i = 0; i < vi->max_queue_pairs; i++) {
2323                 while (vi->rq[i].pages)
2324                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2325
2326                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2327                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2328                 if (old_prog)
2329                         bpf_prog_put(old_prog);
2330         }
2331 }
2332
2333 static void free_receive_bufs(struct virtnet_info *vi)
2334 {
2335         rtnl_lock();
2336         _free_receive_bufs(vi);
2337         rtnl_unlock();
2338 }
2339
2340 static void free_receive_page_frags(struct virtnet_info *vi)
2341 {
2342         int i;
2343         for (i = 0; i < vi->max_queue_pairs; i++)
2344                 if (vi->rq[i].alloc_frag.page)
2345                         put_page(vi->rq[i].alloc_frag.page);
2346 }
2347
2348 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2349 {
2350         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2351                 return false;
2352         else if (q < vi->curr_queue_pairs)
2353                 return true;
2354         else
2355                 return false;
2356 }
2357
2358 static void free_unused_bufs(struct virtnet_info *vi)
2359 {
2360         void *buf;
2361         int i;
2362
2363         for (i = 0; i < vi->max_queue_pairs; i++) {
2364                 struct virtqueue *vq = vi->sq[i].vq;
2365                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2366                         if (!is_xdp_raw_buffer_queue(vi, i))
2367                                 dev_kfree_skb(buf);
2368                         else
2369                                 put_page(virt_to_head_page(buf));
2370                 }
2371         }
2372
2373         for (i = 0; i < vi->max_queue_pairs; i++) {
2374                 struct virtqueue *vq = vi->rq[i].vq;
2375
2376                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2377                         if (vi->mergeable_rx_bufs) {
2378                                 put_page(virt_to_head_page(buf));
2379                         } else if (vi->big_packets) {
2380                                 give_pages(&vi->rq[i], buf);
2381                         } else {
2382                                 put_page(virt_to_head_page(buf));
2383                         }
2384                 }
2385         }
2386 }
2387
2388 static void virtnet_del_vqs(struct virtnet_info *vi)
2389 {
2390         struct virtio_device *vdev = vi->vdev;
2391
2392         virtnet_clean_affinity(vi, -1);
2393
2394         vdev->config->del_vqs(vdev);
2395
2396         virtnet_free_queues(vi);
2397 }
2398
2399 /* How large should a single buffer be so a queue full of these can fit at
2400  * least one full packet?
2401  * Logic below assumes the mergeable buffer header is used.
2402  */
2403 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2404 {
2405         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2406         unsigned int rq_size = virtqueue_get_vring_size(vq);
2407         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2408         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2409         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2410
2411         return max(max(min_buf_len, hdr_len) - hdr_len,
2412                    (unsigned int)GOOD_PACKET_LEN);
2413 }
2414
2415 static int virtnet_find_vqs(struct virtnet_info *vi)
2416 {
2417         vq_callback_t **callbacks;
2418         struct virtqueue **vqs;
2419         int ret = -ENOMEM;
2420         int i, total_vqs;
2421         const char **names;
2422         bool *ctx;
2423
2424         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2425          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2426          * possible control vq.
2427          */
2428         total_vqs = vi->max_queue_pairs * 2 +
2429                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2430
2431         /* Allocate space for find_vqs parameters */
2432         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2433         if (!vqs)
2434                 goto err_vq;
2435         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2436         if (!callbacks)
2437                 goto err_callback;
2438         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2439         if (!names)
2440                 goto err_names;
2441         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2442                 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2443                 if (!ctx)
2444                         goto err_ctx;
2445         } else {
2446                 ctx = NULL;
2447         }
2448
2449         /* Parameters for control virtqueue, if any */
2450         if (vi->has_cvq) {
2451                 callbacks[total_vqs - 1] = NULL;
2452                 names[total_vqs - 1] = "control";
2453         }
2454
2455         /* Allocate/initialize parameters for send/receive virtqueues */
2456         for (i = 0; i < vi->max_queue_pairs; i++) {
2457                 callbacks[rxq2vq(i)] = skb_recv_done;
2458                 callbacks[txq2vq(i)] = skb_xmit_done;
2459                 sprintf(vi->rq[i].name, "input.%d", i);
2460                 sprintf(vi->sq[i].name, "output.%d", i);
2461                 names[rxq2vq(i)] = vi->rq[i].name;
2462                 names[txq2vq(i)] = vi->sq[i].name;
2463                 if (ctx)
2464                         ctx[rxq2vq(i)] = true;
2465         }
2466
2467         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2468                                          names, ctx, NULL);
2469         if (ret)
2470                 goto err_find;
2471
2472         if (vi->has_cvq) {
2473                 vi->cvq = vqs[total_vqs - 1];
2474                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2475                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2476         }
2477
2478         for (i = 0; i < vi->max_queue_pairs; i++) {
2479                 vi->rq[i].vq = vqs[rxq2vq(i)];
2480                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2481                 vi->sq[i].vq = vqs[txq2vq(i)];
2482         }
2483
2484         kfree(names);
2485         kfree(callbacks);
2486         kfree(vqs);
2487         kfree(ctx);
2488
2489         return 0;
2490
2491 err_find:
2492         kfree(ctx);
2493 err_ctx:
2494         kfree(names);
2495 err_names:
2496         kfree(callbacks);
2497 err_callback:
2498         kfree(vqs);
2499 err_vq:
2500         return ret;
2501 }
2502
2503 static int virtnet_alloc_queues(struct virtnet_info *vi)
2504 {
2505         int i;
2506
2507         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2508         if (!vi->sq)
2509                 goto err_sq;
2510         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2511         if (!vi->rq)
2512                 goto err_rq;
2513
2514         INIT_DELAYED_WORK(&vi->refill, refill_work);
2515         for (i = 0; i < vi->max_queue_pairs; i++) {
2516                 vi->rq[i].pages = NULL;
2517                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2518                                napi_weight);
2519                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2520                                   napi_tx ? napi_weight : 0);
2521
2522                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2523                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2524                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2525
2526                 u64_stats_init(&vi->rq[i].stats.syncp);
2527                 u64_stats_init(&vi->sq[i].stats.syncp);
2528         }
2529
2530         return 0;
2531
2532 err_rq:
2533         kfree(vi->sq);
2534 err_sq:
2535         return -ENOMEM;
2536 }
2537
2538 static int init_vqs(struct virtnet_info *vi)
2539 {
2540         int ret;
2541
2542         /* Allocate send & receive queues */
2543         ret = virtnet_alloc_queues(vi);
2544         if (ret)
2545                 goto err;
2546
2547         ret = virtnet_find_vqs(vi);
2548         if (ret)
2549                 goto err_free;
2550
2551         get_online_cpus();
2552         virtnet_set_affinity(vi);
2553         put_online_cpus();
2554
2555         return 0;
2556
2557 err_free:
2558         virtnet_free_queues(vi);
2559 err:
2560         return ret;
2561 }
2562
2563 #ifdef CONFIG_SYSFS
2564 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2565                 char *buf)
2566 {
2567         struct virtnet_info *vi = netdev_priv(queue->dev);
2568         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2569         struct ewma_pkt_len *avg;
2570
2571         BUG_ON(queue_index >= vi->max_queue_pairs);
2572         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2573         return sprintf(buf, "%u\n",
2574                        get_mergeable_buf_len(&vi->rq[queue_index], avg));
2575 }
2576
2577 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2578         __ATTR_RO(mergeable_rx_buffer_size);
2579
2580 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2581         &mergeable_rx_buffer_size_attribute.attr,
2582         NULL
2583 };
2584
2585 static const struct attribute_group virtio_net_mrg_rx_group = {
2586         .name = "virtio_net",
2587         .attrs = virtio_net_mrg_rx_attrs
2588 };
2589 #endif
2590
2591 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2592                                     unsigned int fbit,
2593                                     const char *fname, const char *dname)
2594 {
2595         if (!virtio_has_feature(vdev, fbit))
2596                 return false;
2597
2598         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2599                 fname, dname);
2600
2601         return true;
2602 }
2603
2604 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2605         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2606
2607 static bool virtnet_validate_features(struct virtio_device *vdev)
2608 {
2609         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2610             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2611                              "VIRTIO_NET_F_CTRL_VQ") ||
2612              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2613                              "VIRTIO_NET_F_CTRL_VQ") ||
2614              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2615                              "VIRTIO_NET_F_CTRL_VQ") ||
2616              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2617              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2618                              "VIRTIO_NET_F_CTRL_VQ"))) {
2619                 return false;
2620         }
2621
2622         return true;
2623 }
2624
2625 #define MIN_MTU ETH_MIN_MTU
2626 #define MAX_MTU ETH_MAX_MTU
2627
2628 static int virtnet_validate(struct virtio_device *vdev)
2629 {
2630         if (!vdev->config->get) {
2631                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2632                         __func__);
2633                 return -EINVAL;
2634         }
2635
2636         if (!virtnet_validate_features(vdev))
2637                 return -EINVAL;
2638
2639         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2640                 int mtu = virtio_cread16(vdev,
2641                                          offsetof(struct virtio_net_config,
2642                                                   mtu));
2643                 if (mtu < MIN_MTU)
2644                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2645         }
2646
2647         return 0;
2648 }
2649
2650 static int virtnet_probe(struct virtio_device *vdev)
2651 {
2652         int i, err = -ENOMEM;
2653         struct net_device *dev;
2654         struct virtnet_info *vi;
2655         u16 max_queue_pairs;
2656         int mtu;
2657
2658         /* Find if host supports multiqueue virtio_net device */
2659         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2660                                    struct virtio_net_config,
2661                                    max_virtqueue_pairs, &max_queue_pairs);
2662
2663         /* We need at least 2 queue's */
2664         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2665             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2666             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2667                 max_queue_pairs = 1;
2668
2669         /* Allocate ourselves a network device with room for our info */
2670         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2671         if (!dev)
2672                 return -ENOMEM;
2673
2674         /* Set up network device as normal. */
2675         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2676         dev->netdev_ops = &virtnet_netdev;
2677         dev->features = NETIF_F_HIGHDMA;
2678
2679         dev->ethtool_ops = &virtnet_ethtool_ops;
2680         SET_NETDEV_DEV(dev, &vdev->dev);
2681
2682         /* Do we support "hardware" checksums? */
2683         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2684                 /* This opens up the world of extra features. */
2685                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2686                 if (csum)
2687                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2688
2689                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2690                         dev->hw_features |= NETIF_F_TSO
2691                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2692                 }
2693                 /* Individual feature bits: what can host handle? */
2694                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2695                         dev->hw_features |= NETIF_F_TSO;
2696                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2697                         dev->hw_features |= NETIF_F_TSO6;
2698                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2699                         dev->hw_features |= NETIF_F_TSO_ECN;
2700
2701                 dev->features |= NETIF_F_GSO_ROBUST;
2702
2703                 if (gso)
2704                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2705                 /* (!csum && gso) case will be fixed by register_netdev() */
2706         }
2707         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2708                 dev->features |= NETIF_F_RXCSUM;
2709
2710         dev->vlan_features = dev->features;
2711
2712         /* MTU range: 68 - 65535 */
2713         dev->min_mtu = MIN_MTU;
2714         dev->max_mtu = MAX_MTU;
2715
2716         /* Configuration may specify what MAC to use.  Otherwise random. */
2717         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2718                 virtio_cread_bytes(vdev,
2719                                    offsetof(struct virtio_net_config, mac),
2720                                    dev->dev_addr, dev->addr_len);
2721         else
2722                 eth_hw_addr_random(dev);
2723
2724         /* Set up our device-specific information */
2725         vi = netdev_priv(dev);
2726         vi->dev = dev;
2727         vi->vdev = vdev;
2728         vdev->priv = vi;
2729
2730         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2731
2732         /* If we can receive ANY GSO packets, we must allocate large ones. */
2733         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2734             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2735             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2736             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2737                 vi->big_packets = true;
2738
2739         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2740                 vi->mergeable_rx_bufs = true;
2741
2742         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2743             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2744                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2745         else
2746                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2747
2748         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2749             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2750                 vi->any_header_sg = true;
2751
2752         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2753                 vi->has_cvq = true;
2754
2755         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2756                 mtu = virtio_cread16(vdev,
2757                                      offsetof(struct virtio_net_config,
2758                                               mtu));
2759                 if (mtu < dev->min_mtu) {
2760                         /* Should never trigger: MTU was previously validated
2761                          * in virtnet_validate.
2762                          */
2763                         dev_err(&vdev->dev, "device MTU appears to have changed "
2764                                 "it is now %d < %d", mtu, dev->min_mtu);
2765                         goto free;
2766                 }
2767
2768                 dev->mtu = mtu;
2769                 dev->max_mtu = mtu;
2770
2771                 /* TODO: size buffers correctly in this case. */
2772                 if (dev->mtu > ETH_DATA_LEN)
2773                         vi->big_packets = true;
2774         }
2775
2776         if (vi->any_header_sg)
2777                 dev->needed_headroom = vi->hdr_len;
2778
2779         /* Enable multiqueue by default */
2780         if (num_online_cpus() >= max_queue_pairs)
2781                 vi->curr_queue_pairs = max_queue_pairs;
2782         else
2783                 vi->curr_queue_pairs = num_online_cpus();
2784         vi->max_queue_pairs = max_queue_pairs;
2785
2786         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2787         err = init_vqs(vi);
2788         if (err)
2789                 goto free;
2790
2791 #ifdef CONFIG_SYSFS
2792         if (vi->mergeable_rx_bufs)
2793                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2794 #endif
2795         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2796         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2797
2798         virtnet_init_settings(dev);
2799
2800         err = register_netdev(dev);
2801         if (err) {
2802                 pr_debug("virtio_net: registering device failed\n");
2803                 goto free_vqs;
2804         }
2805
2806         virtio_device_ready(vdev);
2807
2808         err = virtnet_cpu_notif_add(vi);
2809         if (err) {
2810                 pr_debug("virtio_net: registering cpu notifier failed\n");
2811                 goto free_unregister_netdev;
2812         }
2813
2814         virtnet_set_queues(vi, vi->curr_queue_pairs);
2815
2816         /* Assume link up if device can't report link status,
2817            otherwise get link status from config. */
2818         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2819                 netif_carrier_off(dev);
2820                 schedule_work(&vi->config_work);
2821         } else {
2822                 vi->status = VIRTIO_NET_S_LINK_UP;
2823                 virtnet_update_settings(vi);
2824                 netif_carrier_on(dev);
2825         }
2826
2827         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2828                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2829                         set_bit(guest_offloads[i], &vi->guest_offloads);
2830
2831         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2832                  dev->name, max_queue_pairs);
2833
2834         return 0;
2835
2836 free_unregister_netdev:
2837         vi->vdev->config->reset(vdev);
2838
2839         unregister_netdev(dev);
2840 free_vqs:
2841         cancel_delayed_work_sync(&vi->refill);
2842         free_receive_page_frags(vi);
2843         virtnet_del_vqs(vi);
2844 free:
2845         free_netdev(dev);
2846         return err;
2847 }
2848
2849 static void remove_vq_common(struct virtnet_info *vi)
2850 {
2851         vi->vdev->config->reset(vi->vdev);
2852
2853         /* Free unused buffers in both send and recv, if any. */
2854         free_unused_bufs(vi);
2855
2856         free_receive_bufs(vi);
2857
2858         free_receive_page_frags(vi);
2859
2860         virtnet_del_vqs(vi);
2861 }
2862
2863 static void virtnet_remove(struct virtio_device *vdev)
2864 {
2865         struct virtnet_info *vi = vdev->priv;
2866
2867         virtnet_cpu_notif_remove(vi);
2868
2869         /* Make sure no work handler is accessing the device. */
2870         flush_work(&vi->config_work);
2871
2872         unregister_netdev(vi->dev);
2873
2874         remove_vq_common(vi);
2875
2876         free_netdev(vi->dev);
2877 }
2878
2879 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
2880 {
2881         struct virtnet_info *vi = vdev->priv;
2882
2883         virtnet_cpu_notif_remove(vi);
2884         virtnet_freeze_down(vdev);
2885         remove_vq_common(vi);
2886
2887         return 0;
2888 }
2889
2890 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
2891 {
2892         struct virtnet_info *vi = vdev->priv;
2893         int err;
2894
2895         err = virtnet_restore_up(vdev);
2896         if (err)
2897                 return err;
2898         virtnet_set_queues(vi, vi->curr_queue_pairs);
2899
2900         err = virtnet_cpu_notif_add(vi);
2901         if (err)
2902                 return err;
2903
2904         return 0;
2905 }
2906
2907 static struct virtio_device_id id_table[] = {
2908         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2909         { 0 },
2910 };
2911
2912 #define VIRTNET_FEATURES \
2913         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2914         VIRTIO_NET_F_MAC, \
2915         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2916         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2917         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2918         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2919         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2920         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2921         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2922         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
2923         VIRTIO_NET_F_SPEED_DUPLEX
2924
2925 static unsigned int features[] = {
2926         VIRTNET_FEATURES,
2927 };
2928
2929 static unsigned int features_legacy[] = {
2930         VIRTNET_FEATURES,
2931         VIRTIO_NET_F_GSO,
2932         VIRTIO_F_ANY_LAYOUT,
2933 };
2934
2935 static struct virtio_driver virtio_net_driver = {
2936         .feature_table = features,
2937         .feature_table_size = ARRAY_SIZE(features),
2938         .feature_table_legacy = features_legacy,
2939         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2940         .driver.name =  KBUILD_MODNAME,
2941         .driver.owner = THIS_MODULE,
2942         .id_table =     id_table,
2943         .validate =     virtnet_validate,
2944         .probe =        virtnet_probe,
2945         .remove =       virtnet_remove,
2946         .config_changed = virtnet_config_changed,
2947 #ifdef CONFIG_PM_SLEEP
2948         .freeze =       virtnet_freeze,
2949         .restore =      virtnet_restore,
2950 #endif
2951 };
2952
2953 static __init int virtio_net_driver_init(void)
2954 {
2955         int ret;
2956
2957         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2958                                       virtnet_cpu_online,
2959                                       virtnet_cpu_down_prep);
2960         if (ret < 0)
2961                 goto out;
2962         virtionet_online = ret;
2963         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2964                                       NULL, virtnet_cpu_dead);
2965         if (ret)
2966                 goto err_dead;
2967
2968         ret = register_virtio_driver(&virtio_net_driver);
2969         if (ret)
2970                 goto err_virtio;
2971         return 0;
2972 err_virtio:
2973         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2974 err_dead:
2975         cpuhp_remove_multi_state(virtionet_online);
2976 out:
2977         return ret;
2978 }
2979 module_init(virtio_net_driver_init);
2980
2981 static __exit void virtio_net_driver_exit(void)
2982 {
2983         unregister_virtio_driver(&virtio_net_driver);
2984         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2985         cpuhp_remove_multi_state(virtionet_online);
2986 }
2987 module_exit(virtio_net_driver_exit);
2988
2989 MODULE_DEVICE_TABLE(virtio, id_table);
2990 MODULE_DESCRIPTION("Virtio network driver");
2991 MODULE_LICENSE("GPL");