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 /* Control VQ buffers: protected by the rtnl lock */
 151 struct control_buf {
 152         struct virtio_net_ctrl_hdr hdr;
 153         virtio_net_ctrl_ack status;
 154         struct virtio_net_ctrl_mq mq;
 155         u8 promisc;
 156         u8 allmulti;
 157         __virtio16 vid;
 158         __virtio64 offloads;
 159 };
 160
 161 struct virtnet_info {
 162         struct virtio_device *vdev;
 163         struct virtqueue *cvq;
 164         struct net_device *dev;
 165         struct send_queue *sq;
 166         struct receive_queue *rq;
 167         unsigned int status;
 168
 169         /* Max # of queue pairs supported by the device */
 170         u16 max_queue_pairs;
 171
 172         /* # of queue pairs currently used by the driver */
 173         u16 curr_queue_pairs;
 174
 175         /* # of XDP queue pairs currently used by the driver */
 176         u16 xdp_queue_pairs;
 177
 178         /* I like... big packets and I cannot lie! */
 179         bool big_packets;
 180
 181         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 182         bool mergeable_rx_bufs;
 183
 184         /* Has control virtqueue */
 185         bool has_cvq;
 186
 187         /* Host can handle any s/g split between our header and packet data */
 188         bool any_header_sg;
 189
 190         /* Packet virtio header size */
 191         u8 hdr_len;
 192
 193         /* Work struct for refilling if we run low on memory. */
 194         struct delayed_work refill;
 195
 196         /* Work struct for config space updates */
 197         struct work_struct config_work;
 198
 199         /* Does the affinity hint is set for virtqueues? */
 200         bool affinity_hint_set;
 201
 202         /* CPU hotplug instances for online & dead */
 203         struct hlist_node node;
 204         struct hlist_node node_dead;
 205
 206         struct control_buf *ctrl;
 207
 208         /* Ethtool settings */
 209         u8 duplex;
 210         u32 speed;
 211
 212         unsigned long guest_offloads;
 213 };
 214
 215 struct padded_vnet_hdr {
 216         struct virtio_net_hdr_mrg_rxbuf hdr;
 217         /*
 218          * hdr is in a separate sg buffer, and data sg buffer shares same page
 219          * with this header sg. This padding makes next sg 16 byte aligned
 220          * after the header.
 221          */
 222         char padding[4];
 223 };
 224
 225 /* Converting between virtqueue no. and kernel tx/rx queue no.
 226  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 227  */
 228 static int vq2txq(struct virtqueue *vq)
 229 {
 230         return (vq->index - 1) / 2;
 231 }
 232
 233 static int txq2vq(int txq)
 234 {
 235         return txq * 2 + 1;
 236 }
 237
 238 static int vq2rxq(struct virtqueue *vq)
 239 {
 240         return vq->index / 2;
 241 }
 242
 243 static int rxq2vq(int rxq)
 244 {
 245         return rxq * 2;
 246 }
 247
 248 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 249 {
 250         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 251 }
 252
 253 /*
 254  * private is used to chain pages for big packets, put the whole
 255  * most recent used list in the beginning for reuse
 256  */
 257 static void give_pages(struct receive_queue *rq, struct page *page)
 258 {
 259         struct page *end;
 260
 261         /* Find end of list, sew whole thing into vi->rq.pages. */
 262         for (end = page; end->private; end = (struct page *)end->private);
 263         end->private = (unsigned long)rq->pages;
 264         rq->pages = page;
 265 }
 266
 267 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 268 {
 269         struct page *p = rq->pages;
 270
 271         if (p) {
 272                 rq->pages = (struct page *)p->private;
 273                 /* clear private here, it is used to chain pages */
 274                 p->private = 0;
 275         } else
 276                 p = alloc_page(gfp_mask);
 277         return p;
 278 }
 279
 280 static void virtqueue_napi_schedule(struct napi_struct *napi,
 281                                     struct virtqueue *vq)
 282 {
 283         if (napi_schedule_prep(napi)) {
 284                 virtqueue_disable_cb(vq);
 285                 __napi_schedule(napi);
 286         }
 287 }
 288
 289 static void virtqueue_napi_complete(struct napi_struct *napi,
 290                                     struct virtqueue *vq, int processed)
 291 {
 292         int opaque;
 293
 294         opaque = virtqueue_enable_cb_prepare(vq);
 295         if (napi_complete_done(napi, processed)) {
 296                 if (unlikely(virtqueue_poll(vq, opaque)))
 297                         virtqueue_napi_schedule(napi, vq);
 298         } else {
 299                 virtqueue_disable_cb(vq);
 300         }
 301 }
 302
 303 static void skb_xmit_done(struct virtqueue *vq)
 304 {
 305         struct virtnet_info *vi = vq->vdev->priv;
 306         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 307
 308         /* Suppress further interrupts. */
 309         virtqueue_disable_cb(vq);
 310
 311         if (napi->weight)
 312                 virtqueue_napi_schedule(napi, vq);
 313         else
 314                 /* We were probably waiting for more output buffers. */
 315                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 316 }
 317
 318 #define MRG_CTX_HEADER_SHIFT 22
 319 static void *mergeable_len_to_ctx(unsigned int truesize,
 320                                   unsigned int headroom)
 321 {
 322         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 323 }
 324
 325 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 326 {
 327         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 328 }
 329
 330 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 331 {
 332         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 333 }
 334
 335 /* Called from bottom half context */
 336 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 337                                    struct receive_queue *rq,
 338                                    struct page *page, unsigned int offset,
 339                                    unsigned int len, unsigned int truesize)
 340 {
 341         struct sk_buff *skb;
 342         struct virtio_net_hdr_mrg_rxbuf *hdr;
 343         unsigned int copy, hdr_len, hdr_padded_len;
 344         char *p;
 345
 346         p = page_address(page) + offset;
 347
 348         /* copy small packet so we can reuse these pages for small data */
 349         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 350         if (unlikely(!skb))
 351                 return NULL;
 352
 353         hdr = skb_vnet_hdr(skb);
 354
 355         hdr_len = vi->hdr_len;
 356         if (vi->mergeable_rx_bufs)
 357                 hdr_padded_len = sizeof(*hdr);
 358         else
 359                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 360
 361         memcpy(hdr, p, hdr_len);
 362
 363         len -= hdr_len;
 364         offset += hdr_padded_len;
 365         p += hdr_padded_len;
 366
 367         copy = len;
 368         if (copy > skb_tailroom(skb))
 369                 copy = skb_tailroom(skb);
 370         skb_put_data(skb, p, copy);
 371
 372         len -= copy;
 373         offset += copy;
 374
 375         if (vi->mergeable_rx_bufs) {
 376                 if (len)
 377                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 378                 else
 379                         put_page(page);
 380                 return skb;
 381         }
 382
 383         /*
 384          * Verify that we can indeed put this data into a skb.
 385          * This is here to handle cases when the device erroneously
 386          * tries to receive more than is possible. This is usually
 387          * the case of a broken device.
 388          */
 389         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 390                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 391                 dev_kfree_skb(skb);
 392                 return NULL;
 393         }
 394         BUG_ON(offset >= PAGE_SIZE);
 395         while (len) {
 396                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 397                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 398                                 frag_size, truesize);
 399                 len -= frag_size;
 400                 page = (struct page *)page->private;
 401                 offset = 0;
 402         }
 403
 404         if (page)
 405                 give_pages(rq, page);
 406
 407         return skb;
 408 }
 409
 410 static void virtnet_xdp_flush(struct net_device *dev)
 411 {
 412         struct virtnet_info *vi = netdev_priv(dev);
 413         struct send_queue *sq;
 414         unsigned int qp;
 415
 416         qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
 417         sq = &vi->sq[qp];
 418
 419         virtqueue_kick(sq->vq);
 420 }
 421
 422 static bool __virtnet_xdp_xmit(struct virtnet_info *vi,
 423                                struct xdp_buff *xdp)
 424 {
 425         struct virtio_net_hdr_mrg_rxbuf *hdr;
 426         unsigned int len;
 427         struct send_queue *sq;
 428         unsigned int qp;
 429         void *xdp_sent;
 430         int err;
 431
 432         qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
 433         sq = &vi->sq[qp];
 434
 435         /* Free up any pending old buffers before queueing new ones. */
 436         while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 437                 struct page *sent_page = virt_to_head_page(xdp_sent);
 438
 439                 put_page(sent_page);
 440         }
 441
 442         xdp->data -= vi->hdr_len;
 443         /* Zero header and leave csum up to XDP layers */
 444         hdr = xdp->data;
 445         memset(hdr, 0, vi->hdr_len);
 446
 447         sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
 448
 449         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp->data, GFP_ATOMIC);
 450         if (unlikely(err))
 451                 return false; /* Caller handle free/refcnt */
 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         struct receive_queue *rq = vi->rq;
 460         struct bpf_prog *xdp_prog;
 461         bool sent;
 462
 463         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 464          * indicate XDP resources have been successfully allocated.
 465          */
 466         xdp_prog = rcu_dereference(rq->xdp_prog);
 467         if (!xdp_prog)
 468                 return -ENXIO;
 469
 470         sent = __virtnet_xdp_xmit(vi, xdp);
 471         if (!sent)
 472                 return -ENOSPC;
 473         return 0;
 474 }
 475
 476 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 477 {
 478         return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
 479 }
 480
 481 /* We copy the packet for XDP in the following cases:
 482  *
 483  * 1) Packet is scattered across multiple rx buffers.
 484  * 2) Headroom space is insufficient.
 485  *
 486  * This is inefficient but it's a temporary condition that
 487  * we hit right after XDP is enabled and until queue is refilled
 488  * with large buffers with sufficient headroom - so it should affect
 489  * at most queue size packets.
 490  * Afterwards, the conditions to enable
 491  * XDP should preclude the underlying device from sending packets
 492  * across multiple buffers (num_buf > 1), and we make sure buffers
 493  * have enough headroom.
 494  */
 495 static struct page *xdp_linearize_page(struct receive_queue *rq,
 496                                        u16 *num_buf,
 497                                        struct page *p,
 498                                        int offset,
 499                                        int page_off,
 500                                        unsigned int *len)
 501 {
 502         struct page *page = alloc_page(GFP_ATOMIC);
 503
 504         if (!page)
 505                 return NULL;
 506
 507         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 508         page_off += *len;
 509
 510         while (--*num_buf) {
 511                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 512                 unsigned int buflen;
 513                 void *buf;
 514                 int off;
 515
 516                 buf = virtqueue_get_buf(rq->vq, &buflen);
 517                 if (unlikely(!buf))
 518                         goto err_buf;
 519
 520                 p = virt_to_head_page(buf);
 521                 off = buf - page_address(p);
 522
 523                 /* guard against a misconfigured or uncooperative backend that
 524                  * is sending packet larger than the MTU.
 525                  */
 526                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 527                         put_page(p);
 528                         goto err_buf;
 529                 }
 530
 531                 memcpy(page_address(page) + page_off,
 532                        page_address(p) + off, buflen);
 533                 page_off += buflen;
 534                 put_page(p);
 535         }
 536
 537         /* Headroom does not contribute to packet length */
 538         *len = page_off - VIRTIO_XDP_HEADROOM;
 539         return page;
 540 err_buf:
 541         __free_pages(page, 0);
 542         return NULL;
 543 }
 544
 545 static struct sk_buff *receive_small(struct net_device *dev,
 546                                      struct virtnet_info *vi,
 547                                      struct receive_queue *rq,
 548                                      void *buf, void *ctx,
 549                                      unsigned int len,
 550                                      bool *xdp_xmit)
 551 {
 552         struct sk_buff *skb;
 553         struct bpf_prog *xdp_prog;
 554         unsigned int xdp_headroom = (unsigned long)ctx;
 555         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 556         unsigned int headroom = vi->hdr_len + header_offset;
 557         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 558                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 559         struct page *page = virt_to_head_page(buf);
 560         unsigned int delta = 0;
 561         struct page *xdp_page;
 562         bool sent;
 563         int err;
 564
 565         len -= vi->hdr_len;
 566
 567         rcu_read_lock();
 568         xdp_prog = rcu_dereference(rq->xdp_prog);
 569         if (xdp_prog) {
 570                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 571                 struct xdp_buff xdp;
 572                 void *orig_data;
 573                 u32 act;
 574
 575                 if (unlikely(hdr->hdr.gso_type))
 576                         goto err_xdp;
 577
 578                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 579                         int offset = buf - page_address(page) + header_offset;
 580                         unsigned int tlen = len + vi->hdr_len;
 581                         u16 num_buf = 1;
 582
 583                         xdp_headroom = virtnet_get_headroom(vi);
 584                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 585                         headroom = vi->hdr_len + header_offset;
 586                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 587                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 588                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 589                                                       offset, header_offset,
 590                                                       &tlen);
 591                         if (!xdp_page)
 592                                 goto err_xdp;
 593
 594                         buf = page_address(xdp_page);
 595                         put_page(page);
 596                         page = xdp_page;
 597                 }
 598
 599                 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
 600                 xdp.data = xdp.data_hard_start + xdp_headroom;
 601                 xdp_set_data_meta_invalid(&xdp);
 602                 xdp.data_end = xdp.data + len;
 603                 xdp.rxq = &rq->xdp_rxq;
 604                 orig_data = xdp.data;
 605                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 606
 607                 switch (act) {
 608                 case XDP_PASS:
 609                         /* Recalculate length in case bpf program changed it */
 610                         delta = orig_data - xdp.data;
 611                         break;
 612                 case XDP_TX:
 613                         sent = __virtnet_xdp_xmit(vi, &xdp);
 614                         if (unlikely(!sent)) {
 615                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 616                                 goto err_xdp;
 617                         }
 618                         *xdp_xmit = true;
 619                         rcu_read_unlock();
 620                         goto xdp_xmit;
 621                 case XDP_REDIRECT:
 622                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 623                         if (err)
 624                                 goto err_xdp;
 625                         *xdp_xmit = true;
 626                         rcu_read_unlock();
 627                         goto xdp_xmit;
 628                 default:
 629                         bpf_warn_invalid_xdp_action(act);
 630                 case XDP_ABORTED:
 631                         trace_xdp_exception(vi->dev, xdp_prog, act);
 632                 case XDP_DROP:
 633                         goto err_xdp;
 634                 }
 635         }
 636         rcu_read_unlock();
 637
 638         skb = build_skb(buf, buflen);
 639         if (!skb) {
 640                 put_page(page);
 641                 goto err;
 642         }
 643         skb_reserve(skb, headroom - delta);
 644         skb_put(skb, len + delta);
 645         if (!delta) {
 646                 buf += header_offset;
 647                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 648         } /* keep zeroed vnet hdr since packet was changed by bpf */
 649
 650 err:
 651         return skb;
 652
 653 err_xdp:
 654         rcu_read_unlock();
 655         dev->stats.rx_dropped++;
 656         put_page(page);
 657 xdp_xmit:
 658         return NULL;
 659 }
 660
 661 static struct sk_buff *receive_big(struct net_device *dev,
 662                                    struct virtnet_info *vi,
 663                                    struct receive_queue *rq,
 664                                    void *buf,
 665                                    unsigned int len)
 666 {
 667         struct page *page = buf;
 668         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 669
 670         if (unlikely(!skb))
 671                 goto err;
 672
 673         return skb;
 674
 675 err:
 676         dev->stats.rx_dropped++;
 677         give_pages(rq, page);
 678         return NULL;
 679 }
 680
 681 static struct sk_buff *receive_mergeable(struct net_device *dev,
 682                                          struct virtnet_info *vi,
 683                                          struct receive_queue *rq,
 684                                          void *buf,
 685                                          void *ctx,
 686                                          unsigned int len,
 687                                          bool *xdp_xmit)
 688 {
 689         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 690         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 691         struct page *page = virt_to_head_page(buf);
 692         int offset = buf - page_address(page);
 693         struct sk_buff *head_skb, *curr_skb;
 694         struct bpf_prog *xdp_prog;
 695         unsigned int truesize;
 696         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 697         bool sent;
 698         int err;
 699
 700         head_skb = NULL;
 701
 702         rcu_read_lock();
 703         xdp_prog = rcu_dereference(rq->xdp_prog);
 704         if (xdp_prog) {
 705                 struct page *xdp_page;
 706                 struct xdp_buff xdp;
 707                 void *data;
 708                 u32 act;
 709
 710                 /* Transient failure which in theory could occur if
 711                  * in-flight packets from before XDP was enabled reach
 712                  * the receive path after XDP is loaded.
 713                  */
 714                 if (unlikely(hdr->hdr.gso_type))
 715                         goto err_xdp;
 716
 717                 /* This happens when rx buffer size is underestimated
 718                  * or headroom is not enough because of the buffer
 719                  * was refilled before XDP is set. This should only
 720                  * happen for the first several packets, so we don't
 721                  * care much about its performance.
 722                  */
 723                 if (unlikely(num_buf > 1 ||
 724                              headroom < virtnet_get_headroom(vi))) {
 725                         /* linearize data for XDP */
 726                         xdp_page = xdp_linearize_page(rq, &num_buf,
 727                                                       page, offset,
 728                                                       VIRTIO_XDP_HEADROOM,
 729                                                       &len);
 730                         if (!xdp_page)
 731                                 goto err_xdp;
 732                         offset = VIRTIO_XDP_HEADROOM;
 733                 } else {
 734                         xdp_page = page;
 735                 }
 736
 737                 /* Allow consuming headroom but reserve enough space to push
 738                  * the descriptor on if we get an XDP_TX return code.
 739                  */
 740                 data = page_address(xdp_page) + offset;
 741                 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
 742                 xdp.data = data + vi->hdr_len;
 743                 xdp_set_data_meta_invalid(&xdp);
 744                 xdp.data_end = xdp.data + (len - vi->hdr_len);
 745                 xdp.rxq = &rq->xdp_rxq;
 746
 747                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 748
 749                 switch (act) {
 750                 case XDP_PASS:
 751                         /* recalculate offset to account for any header
 752                          * adjustments. Note other cases do not build an
 753                          * skb and avoid using offset
 754                          */
 755                         offset = xdp.data -
 756                                         page_address(xdp_page) - vi->hdr_len;
 757
 758                         /* We can only create skb based on xdp_page. */
 759                         if (unlikely(xdp_page != page)) {
 760                                 rcu_read_unlock();
 761                                 put_page(page);
 762                                 head_skb = page_to_skb(vi, rq, xdp_page,
 763                                                        offset, len, PAGE_SIZE);
 764                                 return head_skb;
 765                         }
 766                         break;
 767                 case XDP_TX:
 768                         sent = __virtnet_xdp_xmit(vi, &xdp);
 769                         if (unlikely(!sent)) {
 770                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 771                                 if (unlikely(xdp_page != page))
 772                                         put_page(xdp_page);
 773                                 goto err_xdp;
 774                         }
 775                         *xdp_xmit = true;
 776                         if (unlikely(xdp_page != page))
 777                                 put_page(page);
 778                         rcu_read_unlock();
 779                         goto xdp_xmit;
 780                 case XDP_REDIRECT:
 781                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 782                         if (err) {
 783                                 if (unlikely(xdp_page != page))
 784                                         put_page(xdp_page);
 785                                 goto err_xdp;
 786                         }
 787                         *xdp_xmit = true;
 788                         if (unlikely(xdp_page != page))
 789                                 put_page(page);
 790                         rcu_read_unlock();
 791                         goto xdp_xmit;
 792                 default:
 793                         bpf_warn_invalid_xdp_action(act);
 794                 case XDP_ABORTED:
 795                         trace_xdp_exception(vi->dev, xdp_prog, act);
 796                 case XDP_DROP:
 797                         if (unlikely(xdp_page != page))
 798                                 __free_pages(xdp_page, 0);
 799                         goto err_xdp;
 800                 }
 801         }
 802         rcu_read_unlock();
 803
 804         truesize = mergeable_ctx_to_truesize(ctx);
 805         if (unlikely(len > truesize)) {
 806                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 807                          dev->name, len, (unsigned long)ctx);
 808                 dev->stats.rx_length_errors++;
 809                 goto err_skb;
 810         }
 811
 812         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
 813         curr_skb = head_skb;
 814
 815         if (unlikely(!curr_skb))
 816                 goto err_skb;
 817         while (--num_buf) {
 818                 int num_skb_frags;
 819
 820                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
 821                 if (unlikely(!buf)) {
 822                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 823                                  dev->name, num_buf,
 824                                  virtio16_to_cpu(vi->vdev,
 825                                                  hdr->num_buffers));
 826                         dev->stats.rx_length_errors++;
 827                         goto err_buf;
 828                 }
 829
 830                 page = virt_to_head_page(buf);
 831
 832                 truesize = mergeable_ctx_to_truesize(ctx);
 833                 if (unlikely(len > truesize)) {
 834                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 835                                  dev->name, len, (unsigned long)ctx);
 836                         dev->stats.rx_length_errors++;
 837                         goto err_skb;
 838                 }
 839
 840                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 841                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 842                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 843
 844                         if (unlikely(!nskb))
 845                                 goto err_skb;
 846                         if (curr_skb == head_skb)
 847                                 skb_shinfo(curr_skb)->frag_list = nskb;
 848                         else
 849                                 curr_skb->next = nskb;
 850                         curr_skb = nskb;
 851                         head_skb->truesize += nskb->truesize;
 852                         num_skb_frags = 0;
 853                 }
 854                 if (curr_skb != head_skb) {
 855                         head_skb->data_len += len;
 856                         head_skb->len += len;
 857                         head_skb->truesize += truesize;
 858                 }
 859                 offset = buf - page_address(page);
 860                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 861                         put_page(page);
 862                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 863                                              len, truesize);
 864                 } else {
 865                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 866                                         offset, len, truesize);
 867                 }
 868         }
 869
 870         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 871         return head_skb;
 872
 873 err_xdp:
 874         rcu_read_unlock();
 875 err_skb:
 876         put_page(page);
 877         while (num_buf-- > 1) {
 878                 buf = virtqueue_get_buf(rq->vq, &len);
 879                 if (unlikely(!buf)) {
 880                         pr_debug("%s: rx error: %d buffers missing\n",
 881                                  dev->name, num_buf);
 882                         dev->stats.rx_length_errors++;
 883                         break;
 884                 }
 885                 page = virt_to_head_page(buf);
 886                 put_page(page);
 887         }
 888 err_buf:
 889         dev->stats.rx_dropped++;
 890         dev_kfree_skb(head_skb);
 891 xdp_xmit:
 892         return NULL;
 893 }
 894
 895 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 896                        void *buf, unsigned int len, void **ctx, bool *xdp_xmit)
 897 {
 898         struct net_device *dev = vi->dev;
 899         struct sk_buff *skb;
 900         struct virtio_net_hdr_mrg_rxbuf *hdr;
 901         int ret;
 902
 903         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 904                 pr_debug("%s: short packet %i\n", dev->name, len);
 905                 dev->stats.rx_length_errors++;
 906                 if (vi->mergeable_rx_bufs) {
 907                         put_page(virt_to_head_page(buf));
 908                 } else if (vi->big_packets) {
 909                         give_pages(rq, buf);
 910                 } else {
 911                         put_page(virt_to_head_page(buf));
 912                 }
 913                 return 0;
 914         }
 915
 916         if (vi->mergeable_rx_bufs)
 917                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit);
 918         else if (vi->big_packets)
 919                 skb = receive_big(dev, vi, rq, buf, len);
 920         else
 921                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit);
 922
 923         if (unlikely(!skb))
 924                 return 0;
 925
 926         hdr = skb_vnet_hdr(skb);
 927
 928         ret = skb->len;
 929
 930         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 931                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 932
 933         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 934                                   virtio_is_little_endian(vi->vdev))) {
 935                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 936                                      dev->name, hdr->hdr.gso_type,
 937                                      hdr->hdr.gso_size);
 938                 goto frame_err;
 939         }
 940
 941         skb->protocol = eth_type_trans(skb, dev);
 942         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 943                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 944
 945         napi_gro_receive(&rq->napi, skb);
 946         return ret;
 947
 948 frame_err:
 949         dev->stats.rx_frame_errors++;
 950         dev_kfree_skb(skb);
 951         return 0;
 952 }
 953
 954 /* Unlike mergeable buffers, all buffers are allocated to the
 955  * same size, except for the headroom. For this reason we do
 956  * not need to use  mergeable_len_to_ctx here - it is enough
 957  * to store the headroom as the context ignoring the truesize.
 958  */
 959 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 960                              gfp_t gfp)
 961 {
 962         struct page_frag *alloc_frag = &rq->alloc_frag;
 963         char *buf;
 964         unsigned int xdp_headroom = virtnet_get_headroom(vi);
 965         void *ctx = (void *)(unsigned long)xdp_headroom;
 966         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
 967         int err;
 968
 969         len = SKB_DATA_ALIGN(len) +
 970               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 971         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 972                 return -ENOMEM;
 973
 974         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 975         get_page(alloc_frag->page);
 976         alloc_frag->offset += len;
 977         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
 978                     vi->hdr_len + GOOD_PACKET_LEN);
 979         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
 980         if (err < 0)
 981                 put_page(virt_to_head_page(buf));
 982         return err;
 983 }
 984
 985 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 986                            gfp_t gfp)
 987 {
 988         struct page *first, *list = NULL;
 989         char *p;
 990         int i, err, offset;
 991
 992         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 993
 994         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 995         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 996                 first = get_a_page(rq, gfp);
 997                 if (!first) {
 998                         if (list)
 999                                 give_pages(rq, list);
1000                         return -ENOMEM;
1001                 }
1002                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1003
1004                 /* chain new page in list head to match sg */
1005                 first->private = (unsigned long)list;
1006                 list = first;
1007         }
1008
1009         first = get_a_page(rq, gfp);
1010         if (!first) {
1011                 give_pages(rq, list);
1012                 return -ENOMEM;
1013         }
1014         p = page_address(first);
1015
1016         /* rq->sg[0], rq->sg[1] share the same page */
1017         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1018         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1019
1020         /* rq->sg[1] for data packet, from offset */
1021         offset = sizeof(struct padded_vnet_hdr);
1022         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1023
1024         /* chain first in list head */
1025         first->private = (unsigned long)list;
1026         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1027                                   first, gfp);
1028         if (err < 0)
1029                 give_pages(rq, first);
1030
1031         return err;
1032 }
1033
1034 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1035                                           struct ewma_pkt_len *avg_pkt_len,
1036                                           unsigned int room)
1037 {
1038         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1039         unsigned int len;
1040
1041         if (room)
1042                 return PAGE_SIZE - room;
1043
1044         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1045                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1046
1047         return ALIGN(len, L1_CACHE_BYTES);
1048 }
1049
1050 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1051                                  struct receive_queue *rq, gfp_t gfp)
1052 {
1053         struct page_frag *alloc_frag = &rq->alloc_frag;
1054         unsigned int headroom = virtnet_get_headroom(vi);
1055         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1056         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1057         char *buf;
1058         void *ctx;
1059         int err;
1060         unsigned int len, hole;
1061
1062         /* Extra tailroom is needed to satisfy XDP's assumption. This
1063          * means rx frags coalescing won't work, but consider we've
1064          * disabled GSO for XDP, it won't be a big issue.
1065          */
1066         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1067         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1068                 return -ENOMEM;
1069
1070         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1071         buf += headroom; /* advance address leaving hole at front of pkt */
1072         get_page(alloc_frag->page);
1073         alloc_frag->offset += len + room;
1074         hole = alloc_frag->size - alloc_frag->offset;
1075         if (hole < len + room) {
1076                 /* To avoid internal fragmentation, if there is very likely not
1077                  * enough space for another buffer, add the remaining space to
1078                  * the current buffer.
1079                  */
1080                 len += hole;
1081                 alloc_frag->offset += hole;
1082         }
1083
1084         sg_init_one(rq->sg, buf, len);
1085         ctx = mergeable_len_to_ctx(len, headroom);
1086         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1087         if (err < 0)
1088                 put_page(virt_to_head_page(buf));
1089
1090         return err;
1091 }
1092
1093 /*
1094  * Returns false if we couldn't fill entirely (OOM).
1095  *
1096  * Normally run in the receive path, but can also be run from ndo_open
1097  * before we're receiving packets, or from refill_work which is
1098  * careful to disable receiving (using napi_disable).
1099  */
1100 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1101                           gfp_t gfp)
1102 {
1103         int err;
1104         bool oom;
1105
1106         do {
1107                 if (vi->mergeable_rx_bufs)
1108                         err = add_recvbuf_mergeable(vi, rq, gfp);
1109                 else if (vi->big_packets)
1110                         err = add_recvbuf_big(vi, rq, gfp);
1111                 else
1112                         err = add_recvbuf_small(vi, rq, gfp);
1113
1114                 oom = err == -ENOMEM;
1115                 if (err)
1116                         break;
1117         } while (rq->vq->num_free);
1118         virtqueue_kick(rq->vq);
1119         return !oom;
1120 }
1121
1122 static void skb_recv_done(struct virtqueue *rvq)
1123 {
1124         struct virtnet_info *vi = rvq->vdev->priv;
1125         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1126
1127         virtqueue_napi_schedule(&rq->napi, rvq);
1128 }
1129
1130 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1131 {
1132         napi_enable(napi);
1133
1134         /* If all buffers were filled by other side before we napi_enabled, we
1135          * won't get another interrupt, so process any outstanding packets now.
1136          * Call local_bh_enable after to trigger softIRQ processing.
1137          */
1138         local_bh_disable();
1139         virtqueue_napi_schedule(napi, vq);
1140         local_bh_enable();
1141 }
1142
1143 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1144                                    struct virtqueue *vq,
1145                                    struct napi_struct *napi)
1146 {
1147         if (!napi->weight)
1148                 return;
1149
1150         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1151          * enable the feature if this is likely affine with the transmit path.
1152          */
1153         if (!vi->affinity_hint_set) {
1154                 napi->weight = 0;
1155                 return;
1156         }
1157
1158         return virtnet_napi_enable(vq, napi);
1159 }
1160
1161 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1162 {
1163         if (napi->weight)
1164                 napi_disable(napi);
1165 }
1166
1167 static void refill_work(struct work_struct *work)
1168 {
1169         struct virtnet_info *vi =
1170                 container_of(work, struct virtnet_info, refill.work);
1171         bool still_empty;
1172         int i;
1173
1174         for (i = 0; i < vi->curr_queue_pairs; i++) {
1175                 struct receive_queue *rq = &vi->rq[i];
1176
1177                 napi_disable(&rq->napi);
1178                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1179                 virtnet_napi_enable(rq->vq, &rq->napi);
1180
1181                 /* In theory, this can happen: if we don't get any buffers in
1182                  * we will *never* try to fill again.
1183                  */
1184                 if (still_empty)
1185                         schedule_delayed_work(&vi->refill, HZ/2);
1186         }
1187 }
1188
1189 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit)
1190 {
1191         struct virtnet_info *vi = rq->vq->vdev->priv;
1192         unsigned int len, received = 0, bytes = 0;
1193         void *buf;
1194
1195         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1196                 void *ctx;
1197
1198                 while (received < budget &&
1199                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1200                         bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit);
1201                         received++;
1202                 }
1203         } else {
1204                 while (received < budget &&
1205                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1206                         bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit);
1207                         received++;
1208                 }
1209         }
1210
1211         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1212                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1213                         schedule_delayed_work(&vi->refill, 0);
1214         }
1215
1216         u64_stats_update_begin(&rq->stats.syncp);
1217         rq->stats.bytes += bytes;
1218         rq->stats.packets += received;
1219         u64_stats_update_end(&rq->stats.syncp);
1220
1221         return received;
1222 }
1223
1224 static void free_old_xmit_skbs(struct send_queue *sq)
1225 {
1226         struct sk_buff *skb;
1227         unsigned int len;
1228         unsigned int packets = 0;
1229         unsigned int bytes = 0;
1230
1231         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1232                 pr_debug("Sent skb %p\n", skb);
1233
1234                 bytes += skb->len;
1235                 packets++;
1236
1237                 dev_consume_skb_any(skb);
1238         }
1239
1240         /* Avoid overhead when no packets have been processed
1241          * happens when called speculatively from start_xmit.
1242          */
1243         if (!packets)
1244                 return;
1245
1246         u64_stats_update_begin(&sq->stats.syncp);
1247         sq->stats.bytes += bytes;
1248         sq->stats.packets += packets;
1249         u64_stats_update_end(&sq->stats.syncp);
1250 }
1251
1252 static void virtnet_poll_cleantx(struct receive_queue *rq)
1253 {
1254         struct virtnet_info *vi = rq->vq->vdev->priv;
1255         unsigned int index = vq2rxq(rq->vq);
1256         struct send_queue *sq = &vi->sq[index];
1257         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1258
1259         if (!sq->napi.weight)
1260                 return;
1261
1262         if (__netif_tx_trylock(txq)) {
1263                 free_old_xmit_skbs(sq);
1264                 __netif_tx_unlock(txq);
1265         }
1266
1267         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1268                 netif_tx_wake_queue(txq);
1269 }
1270
1271 static int virtnet_poll(struct napi_struct *napi, int budget)
1272 {
1273         struct receive_queue *rq =
1274                 container_of(napi, struct receive_queue, napi);
1275         struct virtnet_info *vi = rq->vq->vdev->priv;
1276         struct send_queue *sq;
1277         unsigned int received, qp;
1278         bool xdp_xmit = false;
1279
1280         virtnet_poll_cleantx(rq);
1281
1282         received = virtnet_receive(rq, budget, &xdp_xmit);
1283
1284         /* Out of packets? */
1285         if (received < budget)
1286                 virtqueue_napi_complete(napi, rq->vq, received);
1287
1288         if (xdp_xmit) {
1289                 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs +
1290                      smp_processor_id();
1291                 sq = &vi->sq[qp];
1292                 virtqueue_kick(sq->vq);
1293                 xdp_do_flush_map();
1294         }
1295
1296         return received;
1297 }
1298
1299 static int virtnet_open(struct net_device *dev)
1300 {
1301         struct virtnet_info *vi = netdev_priv(dev);
1302         int i, err;
1303
1304         for (i = 0; i < vi->max_queue_pairs; i++) {
1305                 if (i < vi->curr_queue_pairs)
1306                         /* Make sure we have some buffers: if oom use wq. */
1307                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1308                                 schedule_delayed_work(&vi->refill, 0);
1309
1310                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1311                 if (err < 0)
1312                         return err;
1313
1314                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1315                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1316         }
1317
1318         return 0;
1319 }
1320
1321 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1322 {
1323         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1324         struct virtnet_info *vi = sq->vq->vdev->priv;
1325         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1326
1327         __netif_tx_lock(txq, raw_smp_processor_id());
1328         free_old_xmit_skbs(sq);
1329         __netif_tx_unlock(txq);
1330
1331         virtqueue_napi_complete(napi, sq->vq, 0);
1332
1333         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1334                 netif_tx_wake_queue(txq);
1335
1336         return 0;
1337 }
1338
1339 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1340 {
1341         struct virtio_net_hdr_mrg_rxbuf *hdr;
1342         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1343         struct virtnet_info *vi = sq->vq->vdev->priv;
1344         int num_sg;
1345         unsigned hdr_len = vi->hdr_len;
1346         bool can_push;
1347
1348         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1349
1350         can_push = vi->any_header_sg &&
1351                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1352                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1353         /* Even if we can, don't push here yet as this would skew
1354          * csum_start offset below. */
1355         if (can_push)
1356                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1357         else
1358                 hdr = skb_vnet_hdr(skb);
1359
1360         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1361                                     virtio_is_little_endian(vi->vdev), false))
1362                 BUG();
1363
1364         if (vi->mergeable_rx_bufs)
1365                 hdr->num_buffers = 0;
1366
1367         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1368         if (can_push) {
1369                 __skb_push(skb, hdr_len);
1370                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1371                 if (unlikely(num_sg < 0))
1372                         return num_sg;
1373                 /* Pull header back to avoid skew in tx bytes calculations. */
1374                 __skb_pull(skb, hdr_len);
1375         } else {
1376                 sg_set_buf(sq->sg, hdr, hdr_len);
1377                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1378                 if (unlikely(num_sg < 0))
1379                         return num_sg;
1380                 num_sg++;
1381         }
1382         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1383 }
1384
1385 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1386 {
1387         struct virtnet_info *vi = netdev_priv(dev);
1388         int qnum = skb_get_queue_mapping(skb);
1389         struct send_queue *sq = &vi->sq[qnum];
1390         int err;
1391         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1392         bool kick = !skb->xmit_more;
1393         bool use_napi = sq->napi.weight;
1394
1395         /* Free up any pending old buffers before queueing new ones. */
1396         free_old_xmit_skbs(sq);
1397
1398         if (use_napi && kick)
1399                 virtqueue_enable_cb_delayed(sq->vq);
1400
1401         /* timestamp packet in software */
1402         skb_tx_timestamp(skb);
1403
1404         /* Try to transmit */
1405         err = xmit_skb(sq, skb);
1406
1407         /* This should not happen! */
1408         if (unlikely(err)) {
1409                 dev->stats.tx_fifo_errors++;
1410                 if (net_ratelimit())
1411                         dev_warn(&dev->dev,
1412                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1413                 dev->stats.tx_dropped++;
1414                 dev_kfree_skb_any(skb);
1415                 return NETDEV_TX_OK;
1416         }
1417
1418         /* Don't wait up for transmitted skbs to be freed. */
1419         if (!use_napi) {
1420                 skb_orphan(skb);
1421                 nf_reset(skb);
1422         }
1423
1424         /* If running out of space, stop queue to avoid getting packets that we
1425          * are then unable to transmit.
1426          * An alternative would be to force queuing layer to requeue the skb by
1427          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1428          * returned in a normal path of operation: it means that driver is not
1429          * maintaining the TX queue stop/start state properly, and causes
1430          * the stack to do a non-trivial amount of useless work.
1431          * Since most packets only take 1 or 2 ring slots, stopping the queue
1432          * early means 16 slots are typically wasted.
1433          */
1434         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1435                 netif_stop_subqueue(dev, qnum);
1436                 if (!use_napi &&
1437                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1438                         /* More just got used, free them then recheck. */
1439                         free_old_xmit_skbs(sq);
1440                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1441                                 netif_start_subqueue(dev, qnum);
1442                                 virtqueue_disable_cb(sq->vq);
1443                         }
1444                 }
1445         }
1446
1447         if (kick || netif_xmit_stopped(txq))
1448                 virtqueue_kick(sq->vq);
1449
1450         return NETDEV_TX_OK;
1451 }
1452
1453 /*
1454  * Send command via the control virtqueue and check status.  Commands
1455  * supported by the hypervisor, as indicated by feature bits, should
1456  * never fail unless improperly formatted.
1457  */
1458 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1459                                  struct scatterlist *out)
1460 {
1461         struct scatterlist *sgs[4], hdr, stat;
1462         unsigned out_num = 0, tmp;
1463
1464         /* Caller should know better */
1465         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1466
1467         vi->ctrl->status = ~0;
1468         vi->ctrl->hdr.class = class;
1469         vi->ctrl->hdr.cmd = cmd;
1470         /* Add header */
1471         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1472         sgs[out_num++] = &hdr;
1473
1474         if (out)
1475                 sgs[out_num++] = out;
1476
1477         /* Add return status. */
1478         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1479         sgs[out_num] = &stat;
1480
1481         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1482         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1483
1484         if (unlikely(!virtqueue_kick(vi->cvq)))
1485                 return vi->ctrl->status == VIRTIO_NET_OK;
1486
1487         /* Spin for a response, the kick causes an ioport write, trapping
1488          * into the hypervisor, so the request should be handled immediately.
1489          */
1490         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1491                !virtqueue_is_broken(vi->cvq))
1492                 cpu_relax();
1493
1494         return vi->ctrl->status == VIRTIO_NET_OK;
1495 }
1496
1497 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1498 {
1499         struct virtnet_info *vi = netdev_priv(dev);
1500         struct virtio_device *vdev = vi->vdev;
1501         int ret;
1502         struct sockaddr *addr;
1503         struct scatterlist sg;
1504
1505         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1506         if (!addr)
1507                 return -ENOMEM;
1508
1509         ret = eth_prepare_mac_addr_change(dev, addr);
1510         if (ret)
1511                 goto out;
1512
1513         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1514                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1515                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1516                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1517                         dev_warn(&vdev->dev,
1518                                  "Failed to set mac address by vq command.\n");
1519                         ret = -EINVAL;
1520                         goto out;
1521                 }
1522         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1523                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1524                 unsigned int i;
1525
1526                 /* Naturally, this has an atomicity problem. */
1527                 for (i = 0; i < dev->addr_len; i++)
1528                         virtio_cwrite8(vdev,
1529                                        offsetof(struct virtio_net_config, mac) +
1530                                        i, addr->sa_data[i]);
1531         }
1532
1533         eth_commit_mac_addr_change(dev, p);
1534         ret = 0;
1535
1536 out:
1537         kfree(addr);
1538         return ret;
1539 }
1540
1541 static void virtnet_stats(struct net_device *dev,
1542                           struct rtnl_link_stats64 *tot)
1543 {
1544         struct virtnet_info *vi = netdev_priv(dev);
1545         unsigned int start;
1546         int i;
1547
1548         for (i = 0; i < vi->max_queue_pairs; i++) {
1549                 u64 tpackets, tbytes, rpackets, rbytes;
1550                 struct receive_queue *rq = &vi->rq[i];
1551                 struct send_queue *sq = &vi->sq[i];
1552
1553                 do {
1554                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1555                         tpackets = sq->stats.packets;
1556                         tbytes   = sq->stats.bytes;
1557                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1558
1559                 do {
1560                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1561                         rpackets = rq->stats.packets;
1562                         rbytes   = rq->stats.bytes;
1563                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1564
1565                 tot->rx_packets += rpackets;
1566                 tot->tx_packets += tpackets;
1567                 tot->rx_bytes   += rbytes;
1568                 tot->tx_bytes   += tbytes;
1569         }
1570
1571         tot->tx_dropped = dev->stats.tx_dropped;
1572         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1573         tot->rx_dropped = dev->stats.rx_dropped;
1574         tot->rx_length_errors = dev->stats.rx_length_errors;
1575         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1576 }
1577
1578 #ifdef CONFIG_NET_POLL_CONTROLLER
1579 static void virtnet_netpoll(struct net_device *dev)
1580 {
1581         struct virtnet_info *vi = netdev_priv(dev);
1582         int i;
1583
1584         for (i = 0; i < vi->curr_queue_pairs; i++)
1585                 napi_schedule(&vi->rq[i].napi);
1586 }
1587 #endif
1588
1589 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1590 {
1591         rtnl_lock();
1592         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1593                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1594                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1595         rtnl_unlock();
1596 }
1597
1598 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1599 {
1600         struct scatterlist sg;
1601         struct net_device *dev = vi->dev;
1602
1603         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1604                 return 0;
1605
1606         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1607         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1608
1609         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1610                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1611                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1612                          queue_pairs);
1613                 return -EINVAL;
1614         } else {
1615                 vi->curr_queue_pairs = queue_pairs;
1616                 /* virtnet_open() will refill when device is going to up. */
1617                 if (dev->flags & IFF_UP)
1618                         schedule_delayed_work(&vi->refill, 0);
1619         }
1620
1621         return 0;
1622 }
1623
1624 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1625 {
1626         int err;
1627
1628         rtnl_lock();
1629         err = _virtnet_set_queues(vi, queue_pairs);
1630         rtnl_unlock();
1631         return err;
1632 }
1633
1634 static int virtnet_close(struct net_device *dev)
1635 {
1636         struct virtnet_info *vi = netdev_priv(dev);
1637         int i;
1638
1639         /* Make sure refill_work doesn't re-enable napi! */
1640         cancel_delayed_work_sync(&vi->refill);
1641
1642         for (i = 0; i < vi->max_queue_pairs; i++) {
1643                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1644                 napi_disable(&vi->rq[i].napi);
1645                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1646         }
1647
1648         return 0;
1649 }
1650
1651 static void virtnet_set_rx_mode(struct net_device *dev)
1652 {
1653         struct virtnet_info *vi = netdev_priv(dev);
1654         struct scatterlist sg[2];
1655         struct virtio_net_ctrl_mac *mac_data;
1656         struct netdev_hw_addr *ha;
1657         int uc_count;
1658         int mc_count;
1659         void *buf;
1660         int i;
1661
1662         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1663         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1664                 return;
1665
1666         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1667         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1668
1669         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1670
1671         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1672                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1673                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1674                          vi->ctrl->promisc ? "en" : "dis");
1675
1676         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1677
1678         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1679                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1680                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1681                          vi->ctrl->allmulti ? "en" : "dis");
1682
1683         uc_count = netdev_uc_count(dev);
1684         mc_count = netdev_mc_count(dev);
1685         /* MAC filter - use one buffer for both lists */
1686         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1687                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1688         mac_data = buf;
1689         if (!buf)
1690                 return;
1691
1692         sg_init_table(sg, 2);
1693
1694         /* Store the unicast list and count in the front of the buffer */
1695         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1696         i = 0;
1697         netdev_for_each_uc_addr(ha, dev)
1698                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1699
1700         sg_set_buf(&sg[0], mac_data,
1701                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1702
1703         /* multicast list and count fill the end */
1704         mac_data = (void *)&mac_data->macs[uc_count][0];
1705
1706         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1707         i = 0;
1708         netdev_for_each_mc_addr(ha, dev)
1709                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1710
1711         sg_set_buf(&sg[1], mac_data,
1712                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1713
1714         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1715                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1716                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1717
1718         kfree(buf);
1719 }
1720
1721 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1722                                    __be16 proto, u16 vid)
1723 {
1724         struct virtnet_info *vi = netdev_priv(dev);
1725         struct scatterlist sg;
1726
1727         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1728         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1729
1730         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1731                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1732                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1733         return 0;
1734 }
1735
1736 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1737                                     __be16 proto, u16 vid)
1738 {
1739         struct virtnet_info *vi = netdev_priv(dev);
1740         struct scatterlist sg;
1741
1742         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1743         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1744
1745         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1746                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1747                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1748         return 0;
1749 }
1750
1751 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1752 {
1753         int i;
1754
1755         if (vi->affinity_hint_set) {
1756                 for (i = 0; i < vi->max_queue_pairs; i++) {
1757                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1758                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1759                 }
1760
1761                 vi->affinity_hint_set = false;
1762         }
1763 }
1764
1765 static void virtnet_set_affinity(struct virtnet_info *vi)
1766 {
1767         int i;
1768         int cpu;
1769
1770         /* In multiqueue mode, when the number of cpu is equal to the number of
1771          * queue pairs, we let the queue pairs to be private to one cpu by
1772          * setting the affinity hint to eliminate the contention.
1773          */
1774         if (vi->curr_queue_pairs == 1 ||
1775             vi->max_queue_pairs != num_online_cpus()) {
1776                 virtnet_clean_affinity(vi, -1);
1777                 return;
1778         }
1779
1780         i = 0;
1781         for_each_online_cpu(cpu) {
1782                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1783                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1784                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1785                 i++;
1786         }
1787
1788         vi->affinity_hint_set = true;
1789 }
1790
1791 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1792 {
1793         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1794                                                    node);
1795         virtnet_set_affinity(vi);
1796         return 0;
1797 }
1798
1799 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1800 {
1801         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1802                                                    node_dead);
1803         virtnet_set_affinity(vi);
1804         return 0;
1805 }
1806
1807 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1808 {
1809         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1810                                                    node);
1811
1812         virtnet_clean_affinity(vi, cpu);
1813         return 0;
1814 }
1815
1816 static enum cpuhp_state virtionet_online;
1817
1818 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1819 {
1820         int ret;
1821
1822         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1823         if (ret)
1824                 return ret;
1825         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1826                                                &vi->node_dead);
1827         if (!ret)
1828                 return ret;
1829         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1830         return ret;
1831 }
1832
1833 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1834 {
1835         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1836         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1837                                             &vi->node_dead);
1838 }
1839
1840 static void virtnet_get_ringparam(struct net_device *dev,
1841                                 struct ethtool_ringparam *ring)
1842 {
1843         struct virtnet_info *vi = netdev_priv(dev);
1844
1845         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1846         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1847         ring->rx_pending = ring->rx_max_pending;
1848         ring->tx_pending = ring->tx_max_pending;
1849 }
1850
1851
1852 static void virtnet_get_drvinfo(struct net_device *dev,
1853                                 struct ethtool_drvinfo *info)
1854 {
1855         struct virtnet_info *vi = netdev_priv(dev);
1856         struct virtio_device *vdev = vi->vdev;
1857
1858         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1859         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1860         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1861
1862 }
1863
1864 /* TODO: Eliminate OOO packets during switching */
1865 static int virtnet_set_channels(struct net_device *dev,
1866                                 struct ethtool_channels *channels)
1867 {
1868         struct virtnet_info *vi = netdev_priv(dev);
1869         u16 queue_pairs = channels->combined_count;
1870         int err;
1871
1872         /* We don't support separate rx/tx channels.
1873          * We don't allow setting 'other' channels.
1874          */
1875         if (channels->rx_count || channels->tx_count || channels->other_count)
1876                 return -EINVAL;
1877
1878         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1879                 return -EINVAL;
1880
1881         /* For now we don't support modifying channels while XDP is loaded
1882          * also when XDP is loaded all RX queues have XDP programs so we only
1883          * need to check a single RX queue.
1884          */
1885         if (vi->rq[0].xdp_prog)
1886                 return -EINVAL;
1887
1888         get_online_cpus();
1889         err = _virtnet_set_queues(vi, queue_pairs);
1890         if (!err) {
1891                 netif_set_real_num_tx_queues(dev, queue_pairs);
1892                 netif_set_real_num_rx_queues(dev, queue_pairs);
1893
1894                 virtnet_set_affinity(vi);
1895         }
1896         put_online_cpus();
1897
1898         return err;
1899 }
1900
1901 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
1902 {
1903         struct virtnet_info *vi = netdev_priv(dev);
1904         char *p = (char *)data;
1905         unsigned int i, j;
1906
1907         switch (stringset) {
1908         case ETH_SS_STATS:
1909                 for (i = 0; i < vi->curr_queue_pairs; i++) {
1910                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1911                                 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
1912                                          i, virtnet_rq_stats_desc[j].desc);
1913                                 p += ETH_GSTRING_LEN;
1914                         }
1915                 }
1916
1917                 for (i = 0; i < vi->curr_queue_pairs; i++) {
1918                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1919                                 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
1920                                          i, virtnet_sq_stats_desc[j].desc);
1921                                 p += ETH_GSTRING_LEN;
1922                         }
1923                 }
1924                 break;
1925         }
1926 }
1927
1928 static int virtnet_get_sset_count(struct net_device *dev, int sset)
1929 {
1930         struct virtnet_info *vi = netdev_priv(dev);
1931
1932         switch (sset) {
1933         case ETH_SS_STATS:
1934                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
1935                                                VIRTNET_SQ_STATS_LEN);
1936         default:
1937                 return -EOPNOTSUPP;
1938         }
1939 }
1940
1941 static void virtnet_get_ethtool_stats(struct net_device *dev,
1942                                       struct ethtool_stats *stats, u64 *data)
1943 {
1944         struct virtnet_info *vi = netdev_priv(dev);
1945         unsigned int idx = 0, start, i, j;
1946         const u8 *stats_base;
1947         size_t offset;
1948
1949         for (i = 0; i < vi->curr_queue_pairs; i++) {
1950                 struct receive_queue *rq = &vi->rq[i];
1951
1952                 stats_base = (u8 *)&rq->stats;
1953                 do {
1954                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1955                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1956                                 offset = virtnet_rq_stats_desc[j].offset;
1957                                 data[idx + j] = *(u64 *)(stats_base + offset);
1958                         }
1959                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1960                 idx += VIRTNET_RQ_STATS_LEN;
1961         }
1962
1963         for (i = 0; i < vi->curr_queue_pairs; i++) {
1964                 struct send_queue *sq = &vi->sq[i];
1965
1966                 stats_base = (u8 *)&sq->stats;
1967                 do {
1968                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1969                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1970                                 offset = virtnet_sq_stats_desc[j].offset;
1971                                 data[idx + j] = *(u64 *)(stats_base + offset);
1972                         }
1973                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1974                 idx += VIRTNET_SQ_STATS_LEN;
1975         }
1976 }
1977
1978 static void virtnet_get_channels(struct net_device *dev,
1979                                  struct ethtool_channels *channels)
1980 {
1981         struct virtnet_info *vi = netdev_priv(dev);
1982
1983         channels->combined_count = vi->curr_queue_pairs;
1984         channels->max_combined = vi->max_queue_pairs;
1985         channels->max_other = 0;
1986         channels->rx_count = 0;
1987         channels->tx_count = 0;
1988         channels->other_count = 0;
1989 }
1990
1991 /* Check if the user is trying to change anything besides speed/duplex */
1992 static bool
1993 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1994 {
1995         struct ethtool_link_ksettings diff1 = *cmd;
1996         struct ethtool_link_ksettings diff2 = {};
1997
1998         /* cmd is always set so we need to clear it, validate the port type
1999          * and also without autonegotiation we can ignore advertising
2000          */
2001         diff1.base.speed = 0;
2002         diff2.base.port = PORT_OTHER;
2003         ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
2004         diff1.base.duplex = 0;
2005         diff1.base.cmd = 0;
2006         diff1.base.link_mode_masks_nwords = 0;
2007
2008         return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2009                 bitmap_empty(diff1.link_modes.supported,
2010                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2011                 bitmap_empty(diff1.link_modes.advertising,
2012                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2013                 bitmap_empty(diff1.link_modes.lp_advertising,
2014                              __ETHTOOL_LINK_MODE_MASK_NBITS);
2015 }
2016
2017 static int virtnet_set_link_ksettings(struct net_device *dev,
2018                                       const struct ethtool_link_ksettings *cmd)
2019 {
2020         struct virtnet_info *vi = netdev_priv(dev);
2021         u32 speed;
2022
2023         speed = cmd->base.speed;
2024         /* don't allow custom speed and duplex */
2025         if (!ethtool_validate_speed(speed) ||
2026             !ethtool_validate_duplex(cmd->base.duplex) ||
2027             !virtnet_validate_ethtool_cmd(cmd))
2028                 return -EINVAL;
2029         vi->speed = speed;
2030         vi->duplex = cmd->base.duplex;
2031
2032         return 0;
2033 }
2034
2035 static int virtnet_get_link_ksettings(struct net_device *dev,
2036                                       struct ethtool_link_ksettings *cmd)
2037 {
2038         struct virtnet_info *vi = netdev_priv(dev);
2039
2040         cmd->base.speed = vi->speed;
2041         cmd->base.duplex = vi->duplex;
2042         cmd->base.port = PORT_OTHER;
2043
2044         return 0;
2045 }
2046
2047 static void virtnet_init_settings(struct net_device *dev)
2048 {
2049         struct virtnet_info *vi = netdev_priv(dev);
2050
2051         vi->speed = SPEED_UNKNOWN;
2052         vi->duplex = DUPLEX_UNKNOWN;
2053 }
2054
2055 static void virtnet_update_settings(struct virtnet_info *vi)
2056 {
2057         u32 speed;
2058         u8 duplex;
2059
2060         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2061                 return;
2062
2063         speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2064                                                   speed));
2065         if (ethtool_validate_speed(speed))
2066                 vi->speed = speed;
2067         duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2068                                                   duplex));
2069         if (ethtool_validate_duplex(duplex))
2070                 vi->duplex = duplex;
2071 }
2072
2073 static const struct ethtool_ops virtnet_ethtool_ops = {
2074         .get_drvinfo = virtnet_get_drvinfo,
2075         .get_link = ethtool_op_get_link,
2076         .get_ringparam = virtnet_get_ringparam,
2077         .get_strings = virtnet_get_strings,
2078         .get_sset_count = virtnet_get_sset_count,
2079         .get_ethtool_stats = virtnet_get_ethtool_stats,
2080         .set_channels = virtnet_set_channels,
2081         .get_channels = virtnet_get_channels,
2082         .get_ts_info = ethtool_op_get_ts_info,
2083         .get_link_ksettings = virtnet_get_link_ksettings,
2084         .set_link_ksettings = virtnet_set_link_ksettings,
2085 };
2086
2087 static void virtnet_freeze_down(struct virtio_device *vdev)
2088 {
2089         struct virtnet_info *vi = vdev->priv;
2090         int i;
2091
2092         /* Make sure no work handler is accessing the device */
2093         flush_work(&vi->config_work);
2094
2095         netif_device_detach(vi->dev);
2096         netif_tx_disable(vi->dev);
2097         cancel_delayed_work_sync(&vi->refill);
2098
2099         if (netif_running(vi->dev)) {
2100                 for (i = 0; i < vi->max_queue_pairs; i++) {
2101                         napi_disable(&vi->rq[i].napi);
2102                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2103                 }
2104         }
2105 }
2106
2107 static int init_vqs(struct virtnet_info *vi);
2108
2109 static int virtnet_restore_up(struct virtio_device *vdev)
2110 {
2111         struct virtnet_info *vi = vdev->priv;
2112         int err, i;
2113
2114         err = init_vqs(vi);
2115         if (err)
2116                 return err;
2117
2118         virtio_device_ready(vdev);
2119
2120         if (netif_running(vi->dev)) {
2121                 for (i = 0; i < vi->curr_queue_pairs; i++)
2122                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2123                                 schedule_delayed_work(&vi->refill, 0);
2124
2125                 for (i = 0; i < vi->max_queue_pairs; i++) {
2126                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2127                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2128                                                &vi->sq[i].napi);
2129                 }
2130         }
2131
2132         netif_device_attach(vi->dev);
2133         return err;
2134 }
2135
2136 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2137 {
2138         struct scatterlist sg;
2139         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2140
2141         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2142
2143         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2144                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2145                 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2146                 return -EINVAL;
2147         }
2148
2149         return 0;
2150 }
2151
2152 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2153 {
2154         u64 offloads = 0;
2155
2156         if (!vi->guest_offloads)
2157                 return 0;
2158
2159         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2160                 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2161
2162         return virtnet_set_guest_offloads(vi, offloads);
2163 }
2164
2165 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2166 {
2167         u64 offloads = vi->guest_offloads;
2168
2169         if (!vi->guest_offloads)
2170                 return 0;
2171         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2172                 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2173
2174         return virtnet_set_guest_offloads(vi, offloads);
2175 }
2176
2177 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2178                            struct netlink_ext_ack *extack)
2179 {
2180         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2181         struct virtnet_info *vi = netdev_priv(dev);
2182         struct bpf_prog *old_prog;
2183         u16 xdp_qp = 0, curr_qp;
2184         int i, err;
2185
2186         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2187             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2188                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2189                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2190                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
2191                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
2192                 return -EOPNOTSUPP;
2193         }
2194
2195         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2196                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2197                 return -EINVAL;
2198         }
2199
2200         if (dev->mtu > max_sz) {
2201                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2202                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2203                 return -EINVAL;
2204         }
2205
2206         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2207         if (prog)
2208                 xdp_qp = nr_cpu_ids;
2209
2210         /* XDP requires extra queues for XDP_TX */
2211         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2212                 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2213                 netdev_warn(dev, "request %i queues but max is %i\n",
2214                             curr_qp + xdp_qp, vi->max_queue_pairs);
2215                 return -ENOMEM;
2216         }
2217
2218         if (prog) {
2219                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2220                 if (IS_ERR(prog))
2221                         return PTR_ERR(prog);
2222         }
2223
2224         /* Make sure NAPI is not using any XDP TX queues for RX. */
2225         if (netif_running(dev))
2226                 for (i = 0; i < vi->max_queue_pairs; i++)
2227                         napi_disable(&vi->rq[i].napi);
2228
2229         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2230         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2231         if (err)
2232                 goto err;
2233         vi->xdp_queue_pairs = xdp_qp;
2234
2235         for (i = 0; i < vi->max_queue_pairs; i++) {
2236                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2237                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2238                 if (i == 0) {
2239                         if (!old_prog)
2240                                 virtnet_clear_guest_offloads(vi);
2241                         if (!prog)
2242                                 virtnet_restore_guest_offloads(vi);
2243                 }
2244                 if (old_prog)
2245                         bpf_prog_put(old_prog);
2246                 if (netif_running(dev))
2247                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2248         }
2249
2250         return 0;
2251
2252 err:
2253         for (i = 0; i < vi->max_queue_pairs; i++)
2254                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2255         if (prog)
2256                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2257         return err;
2258 }
2259
2260 static u32 virtnet_xdp_query(struct net_device *dev)
2261 {
2262         struct virtnet_info *vi = netdev_priv(dev);
2263         const struct bpf_prog *xdp_prog;
2264         int i;
2265
2266         for (i = 0; i < vi->max_queue_pairs; i++) {
2267                 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2268                 if (xdp_prog)
2269                         return xdp_prog->aux->id;
2270         }
2271         return 0;
2272 }
2273
2274 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2275 {
2276         switch (xdp->command) {
2277         case XDP_SETUP_PROG:
2278                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2279         case XDP_QUERY_PROG:
2280                 xdp->prog_id = virtnet_xdp_query(dev);
2281                 xdp->prog_attached = !!xdp->prog_id;
2282                 return 0;
2283         default:
2284                 return -EINVAL;
2285         }
2286 }
2287
2288 static const struct net_device_ops virtnet_netdev = {
2289         .ndo_open            = virtnet_open,
2290         .ndo_stop            = virtnet_close,
2291         .ndo_start_xmit      = start_xmit,
2292         .ndo_validate_addr   = eth_validate_addr,
2293         .ndo_set_mac_address = virtnet_set_mac_address,
2294         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2295         .ndo_get_stats64     = virtnet_stats,
2296         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2297         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2298 #ifdef CONFIG_NET_POLL_CONTROLLER
2299         .ndo_poll_controller = virtnet_netpoll,
2300 #endif
2301         .ndo_bpf                = virtnet_xdp,
2302         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2303         .ndo_xdp_flush          = virtnet_xdp_flush,
2304         .ndo_features_check     = passthru_features_check,
2305 };
2306
2307 static void virtnet_config_changed_work(struct work_struct *work)
2308 {
2309         struct virtnet_info *vi =
2310                 container_of(work, struct virtnet_info, config_work);
2311         u16 v;
2312
2313         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2314                                  struct virtio_net_config, status, &v) < 0)
2315                 return;
2316
2317         if (v & VIRTIO_NET_S_ANNOUNCE) {
2318                 netdev_notify_peers(vi->dev);
2319                 virtnet_ack_link_announce(vi);
2320         }
2321
2322         /* Ignore unknown (future) status bits */
2323         v &= VIRTIO_NET_S_LINK_UP;
2324
2325         if (vi->status == v)
2326                 return;
2327
2328         vi->status = v;
2329
2330         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2331                 virtnet_update_settings(vi);
2332                 netif_carrier_on(vi->dev);
2333                 netif_tx_wake_all_queues(vi->dev);
2334         } else {
2335                 netif_carrier_off(vi->dev);
2336                 netif_tx_stop_all_queues(vi->dev);
2337         }
2338 }
2339
2340 static void virtnet_config_changed(struct virtio_device *vdev)
2341 {
2342         struct virtnet_info *vi = vdev->priv;
2343
2344         schedule_work(&vi->config_work);
2345 }
2346
2347 static void virtnet_free_queues(struct virtnet_info *vi)
2348 {
2349         int i;
2350
2351         for (i = 0; i < vi->max_queue_pairs; i++) {
2352                 napi_hash_del(&vi->rq[i].napi);
2353                 netif_napi_del(&vi->rq[i].napi);
2354                 netif_napi_del(&vi->sq[i].napi);
2355         }
2356
2357         /* We called napi_hash_del() before netif_napi_del(),
2358          * we need to respect an RCU grace period before freeing vi->rq
2359          */
2360         synchronize_net();
2361
2362         kfree(vi->rq);
2363         kfree(vi->sq);
2364         kfree(vi->ctrl);
2365 }
2366
2367 static void _free_receive_bufs(struct virtnet_info *vi)
2368 {
2369         struct bpf_prog *old_prog;
2370         int i;
2371
2372         for (i = 0; i < vi->max_queue_pairs; i++) {
2373                 while (vi->rq[i].pages)
2374                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2375
2376                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2377                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2378                 if (old_prog)
2379                         bpf_prog_put(old_prog);
2380         }
2381 }
2382
2383 static void free_receive_bufs(struct virtnet_info *vi)
2384 {
2385         rtnl_lock();
2386         _free_receive_bufs(vi);
2387         rtnl_unlock();
2388 }
2389
2390 static void free_receive_page_frags(struct virtnet_info *vi)
2391 {
2392         int i;
2393         for (i = 0; i < vi->max_queue_pairs; i++)
2394                 if (vi->rq[i].alloc_frag.page)
2395                         put_page(vi->rq[i].alloc_frag.page);
2396 }
2397
2398 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2399 {
2400         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2401                 return false;
2402         else if (q < vi->curr_queue_pairs)
2403                 return true;
2404         else
2405                 return false;
2406 }
2407
2408 static void free_unused_bufs(struct virtnet_info *vi)
2409 {
2410         void *buf;
2411         int i;
2412
2413         for (i = 0; i < vi->max_queue_pairs; i++) {
2414                 struct virtqueue *vq = vi->sq[i].vq;
2415                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2416                         if (!is_xdp_raw_buffer_queue(vi, i))
2417                                 dev_kfree_skb(buf);
2418                         else
2419                                 put_page(virt_to_head_page(buf));
2420                 }
2421         }
2422
2423         for (i = 0; i < vi->max_queue_pairs; i++) {
2424                 struct virtqueue *vq = vi->rq[i].vq;
2425
2426                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2427                         if (vi->mergeable_rx_bufs) {
2428                                 put_page(virt_to_head_page(buf));
2429                         } else if (vi->big_packets) {
2430                                 give_pages(&vi->rq[i], buf);
2431                         } else {
2432                                 put_page(virt_to_head_page(buf));
2433                         }
2434                 }
2435         }
2436 }
2437
2438 static void virtnet_del_vqs(struct virtnet_info *vi)
2439 {
2440         struct virtio_device *vdev = vi->vdev;
2441
2442         virtnet_clean_affinity(vi, -1);
2443
2444         vdev->config->del_vqs(vdev);
2445
2446         virtnet_free_queues(vi);
2447 }
2448
2449 /* How large should a single buffer be so a queue full of these can fit at
2450  * least one full packet?
2451  * Logic below assumes the mergeable buffer header is used.
2452  */
2453 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2454 {
2455         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2456         unsigned int rq_size = virtqueue_get_vring_size(vq);
2457         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2458         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2459         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2460
2461         return max(max(min_buf_len, hdr_len) - hdr_len,
2462                    (unsigned int)GOOD_PACKET_LEN);
2463 }
2464
2465 static int virtnet_find_vqs(struct virtnet_info *vi)
2466 {
2467         vq_callback_t **callbacks;
2468         struct virtqueue **vqs;
2469         int ret = -ENOMEM;
2470         int i, total_vqs;
2471         const char **names;
2472         bool *ctx;
2473
2474         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2475          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2476          * possible control vq.
2477          */
2478         total_vqs = vi->max_queue_pairs * 2 +
2479                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2480
2481         /* Allocate space for find_vqs parameters */
2482         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2483         if (!vqs)
2484                 goto err_vq;
2485         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2486         if (!callbacks)
2487                 goto err_callback;
2488         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2489         if (!names)
2490                 goto err_names;
2491         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2492                 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2493                 if (!ctx)
2494                         goto err_ctx;
2495         } else {
2496                 ctx = NULL;
2497         }
2498
2499         /* Parameters for control virtqueue, if any */
2500         if (vi->has_cvq) {
2501                 callbacks[total_vqs - 1] = NULL;
2502                 names[total_vqs - 1] = "control";
2503         }
2504
2505         /* Allocate/initialize parameters for send/receive virtqueues */
2506         for (i = 0; i < vi->max_queue_pairs; i++) {
2507                 callbacks[rxq2vq(i)] = skb_recv_done;
2508                 callbacks[txq2vq(i)] = skb_xmit_done;
2509                 sprintf(vi->rq[i].name, "input.%d", i);
2510                 sprintf(vi->sq[i].name, "output.%d", i);
2511                 names[rxq2vq(i)] = vi->rq[i].name;
2512                 names[txq2vq(i)] = vi->sq[i].name;
2513                 if (ctx)
2514                         ctx[rxq2vq(i)] = true;
2515         }
2516
2517         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2518                                          names, ctx, NULL);
2519         if (ret)
2520                 goto err_find;
2521
2522         if (vi->has_cvq) {
2523                 vi->cvq = vqs[total_vqs - 1];
2524                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2525                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2526         }
2527
2528         for (i = 0; i < vi->max_queue_pairs; i++) {
2529                 vi->rq[i].vq = vqs[rxq2vq(i)];
2530                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2531                 vi->sq[i].vq = vqs[txq2vq(i)];
2532         }
2533
2534         kfree(names);
2535         kfree(callbacks);
2536         kfree(vqs);
2537         kfree(ctx);
2538
2539         return 0;
2540
2541 err_find:
2542         kfree(ctx);
2543 err_ctx:
2544         kfree(names);
2545 err_names:
2546         kfree(callbacks);
2547 err_callback:
2548         kfree(vqs);
2549 err_vq:
2550         return ret;
2551 }
2552
2553 static int virtnet_alloc_queues(struct virtnet_info *vi)
2554 {
2555         int i;
2556
2557         vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2558         if (!vi->ctrl)
2559                 goto err_ctrl;
2560         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2561         if (!vi->sq)
2562                 goto err_sq;
2563         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2564         if (!vi->rq)
2565                 goto err_rq;
2566
2567         INIT_DELAYED_WORK(&vi->refill, refill_work);
2568         for (i = 0; i < vi->max_queue_pairs; i++) {
2569                 vi->rq[i].pages = NULL;
2570                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2571                                napi_weight);
2572                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2573                                   napi_tx ? napi_weight : 0);
2574
2575                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2576                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2577                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2578
2579                 u64_stats_init(&vi->rq[i].stats.syncp);
2580                 u64_stats_init(&vi->sq[i].stats.syncp);
2581         }
2582
2583         return 0;
2584
2585 err_rq:
2586         kfree(vi->sq);
2587 err_sq:
2588         kfree(vi->ctrl);
2589 err_ctrl:
2590         return -ENOMEM;
2591 }
2592
2593 static int init_vqs(struct virtnet_info *vi)
2594 {
2595         int ret;
2596
2597         /* Allocate send & receive queues */
2598         ret = virtnet_alloc_queues(vi);
2599         if (ret)
2600                 goto err;
2601
2602         ret = virtnet_find_vqs(vi);
2603         if (ret)
2604                 goto err_free;
2605
2606         get_online_cpus();
2607         virtnet_set_affinity(vi);
2608         put_online_cpus();
2609
2610         return 0;
2611
2612 err_free:
2613         virtnet_free_queues(vi);
2614 err:
2615         return ret;
2616 }
2617
2618 #ifdef CONFIG_SYSFS
2619 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2620                 char *buf)
2621 {
2622         struct virtnet_info *vi = netdev_priv(queue->dev);
2623         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2624         unsigned int headroom = virtnet_get_headroom(vi);
2625         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2626         struct ewma_pkt_len *avg;
2627
2628         BUG_ON(queue_index >= vi->max_queue_pairs);
2629         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2630         return sprintf(buf, "%u\n",
2631                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
2632                                        SKB_DATA_ALIGN(headroom + tailroom)));
2633 }
2634
2635 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2636         __ATTR_RO(mergeable_rx_buffer_size);
2637
2638 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2639         &mergeable_rx_buffer_size_attribute.attr,
2640         NULL
2641 };
2642
2643 static const struct attribute_group virtio_net_mrg_rx_group = {
2644         .name = "virtio_net",
2645         .attrs = virtio_net_mrg_rx_attrs
2646 };
2647 #endif
2648
2649 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2650                                     unsigned int fbit,
2651                                     const char *fname, const char *dname)
2652 {
2653         if (!virtio_has_feature(vdev, fbit))
2654                 return false;
2655
2656         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2657                 fname, dname);
2658
2659         return true;
2660 }
2661
2662 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2663         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2664
2665 static bool virtnet_validate_features(struct virtio_device *vdev)
2666 {
2667         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2668             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2669                              "VIRTIO_NET_F_CTRL_VQ") ||
2670              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2671                              "VIRTIO_NET_F_CTRL_VQ") ||
2672              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2673                              "VIRTIO_NET_F_CTRL_VQ") ||
2674              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2675              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2676                              "VIRTIO_NET_F_CTRL_VQ"))) {
2677                 return false;
2678         }
2679
2680         return true;
2681 }
2682
2683 #define MIN_MTU ETH_MIN_MTU
2684 #define MAX_MTU ETH_MAX_MTU
2685
2686 static int virtnet_validate(struct virtio_device *vdev)
2687 {
2688         if (!vdev->config->get) {
2689                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2690                         __func__);
2691                 return -EINVAL;
2692         }
2693
2694         if (!virtnet_validate_features(vdev))
2695                 return -EINVAL;
2696
2697         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2698                 int mtu = virtio_cread16(vdev,
2699                                          offsetof(struct virtio_net_config,
2700                                                   mtu));
2701                 if (mtu < MIN_MTU)
2702                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2703         }
2704
2705         return 0;
2706 }
2707
2708 static int virtnet_probe(struct virtio_device *vdev)
2709 {
2710         int i, err = -ENOMEM;
2711         struct net_device *dev;
2712         struct virtnet_info *vi;
2713         u16 max_queue_pairs;
2714         int mtu;
2715
2716         /* Find if host supports multiqueue virtio_net device */
2717         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2718                                    struct virtio_net_config,
2719                                    max_virtqueue_pairs, &max_queue_pairs);
2720
2721         /* We need at least 2 queue's */
2722         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2723             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2724             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2725                 max_queue_pairs = 1;
2726
2727         /* Allocate ourselves a network device with room for our info */
2728         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2729         if (!dev)
2730                 return -ENOMEM;
2731
2732         /* Set up network device as normal. */
2733         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2734         dev->netdev_ops = &virtnet_netdev;
2735         dev->features = NETIF_F_HIGHDMA;
2736
2737         dev->ethtool_ops = &virtnet_ethtool_ops;
2738         SET_NETDEV_DEV(dev, &vdev->dev);
2739
2740         /* Do we support "hardware" checksums? */
2741         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2742                 /* This opens up the world of extra features. */
2743                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2744                 if (csum)
2745                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2746
2747                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2748                         dev->hw_features |= NETIF_F_TSO
2749                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2750                 }
2751                 /* Individual feature bits: what can host handle? */
2752                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2753                         dev->hw_features |= NETIF_F_TSO;
2754                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2755                         dev->hw_features |= NETIF_F_TSO6;
2756                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2757                         dev->hw_features |= NETIF_F_TSO_ECN;
2758
2759                 dev->features |= NETIF_F_GSO_ROBUST;
2760
2761                 if (gso)
2762                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2763                 /* (!csum && gso) case will be fixed by register_netdev() */
2764         }
2765         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2766                 dev->features |= NETIF_F_RXCSUM;
2767
2768         dev->vlan_features = dev->features;
2769
2770         /* MTU range: 68 - 65535 */
2771         dev->min_mtu = MIN_MTU;
2772         dev->max_mtu = MAX_MTU;
2773
2774         /* Configuration may specify what MAC to use.  Otherwise random. */
2775         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2776                 virtio_cread_bytes(vdev,
2777                                    offsetof(struct virtio_net_config, mac),
2778                                    dev->dev_addr, dev->addr_len);
2779         else
2780                 eth_hw_addr_random(dev);
2781
2782         /* Set up our device-specific information */
2783         vi = netdev_priv(dev);
2784         vi->dev = dev;
2785         vi->vdev = vdev;
2786         vdev->priv = vi;
2787
2788         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2789
2790         /* If we can receive ANY GSO packets, we must allocate large ones. */
2791         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2792             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2793             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2794             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2795                 vi->big_packets = true;
2796
2797         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2798                 vi->mergeable_rx_bufs = true;
2799
2800         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2801             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2802                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2803         else
2804                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2805
2806         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2807             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2808                 vi->any_header_sg = true;
2809
2810         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2811                 vi->has_cvq = true;
2812
2813         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2814                 mtu = virtio_cread16(vdev,
2815                                      offsetof(struct virtio_net_config,
2816                                               mtu));
2817                 if (mtu < dev->min_mtu) {
2818                         /* Should never trigger: MTU was previously validated
2819                          * in virtnet_validate.
2820                          */
2821                         dev_err(&vdev->dev, "device MTU appears to have changed "
2822                                 "it is now %d < %d", mtu, dev->min_mtu);
2823                         goto free;
2824                 }
2825
2826                 dev->mtu = mtu;
2827                 dev->max_mtu = mtu;
2828
2829                 /* TODO: size buffers correctly in this case. */
2830                 if (dev->mtu > ETH_DATA_LEN)
2831                         vi->big_packets = true;
2832         }
2833
2834         if (vi->any_header_sg)
2835                 dev->needed_headroom = vi->hdr_len;
2836
2837         /* Enable multiqueue by default */
2838         if (num_online_cpus() >= max_queue_pairs)
2839                 vi->curr_queue_pairs = max_queue_pairs;
2840         else
2841                 vi->curr_queue_pairs = num_online_cpus();
2842         vi->max_queue_pairs = max_queue_pairs;
2843
2844         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2845         err = init_vqs(vi);
2846         if (err)
2847                 goto free;
2848
2849 #ifdef CONFIG_SYSFS
2850         if (vi->mergeable_rx_bufs)
2851                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2852 #endif
2853         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2854         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2855
2856         virtnet_init_settings(dev);
2857
2858         err = register_netdev(dev);
2859         if (err) {
2860                 pr_debug("virtio_net: registering device failed\n");
2861                 goto free_vqs;
2862         }
2863
2864         virtio_device_ready(vdev);
2865
2866         err = virtnet_cpu_notif_add(vi);
2867         if (err) {
2868                 pr_debug("virtio_net: registering cpu notifier failed\n");
2869                 goto free_unregister_netdev;
2870         }
2871
2872         virtnet_set_queues(vi, vi->curr_queue_pairs);
2873
2874         /* Assume link up if device can't report link status,
2875            otherwise get link status from config. */
2876         netif_carrier_off(dev);
2877         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2878                 schedule_work(&vi->config_work);
2879         } else {
2880                 vi->status = VIRTIO_NET_S_LINK_UP;
2881                 virtnet_update_settings(vi);
2882                 netif_carrier_on(dev);
2883         }
2884
2885         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2886                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2887                         set_bit(guest_offloads[i], &vi->guest_offloads);
2888
2889         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2890                  dev->name, max_queue_pairs);
2891
2892         return 0;
2893
2894 free_unregister_netdev:
2895         vi->vdev->config->reset(vdev);
2896
2897         unregister_netdev(dev);
2898 free_vqs:
2899         cancel_delayed_work_sync(&vi->refill);
2900         free_receive_page_frags(vi);
2901         virtnet_del_vqs(vi);
2902 free:
2903         free_netdev(dev);
2904         return err;
2905 }
2906
2907 static void remove_vq_common(struct virtnet_info *vi)
2908 {
2909         vi->vdev->config->reset(vi->vdev);
2910
2911         /* Free unused buffers in both send and recv, if any. */
2912         free_unused_bufs(vi);
2913
2914         free_receive_bufs(vi);
2915
2916         free_receive_page_frags(vi);
2917
2918         virtnet_del_vqs(vi);
2919 }
2920
2921 static void virtnet_remove(struct virtio_device *vdev)
2922 {
2923         struct virtnet_info *vi = vdev->priv;
2924
2925         virtnet_cpu_notif_remove(vi);
2926
2927         /* Make sure no work handler is accessing the device. */
2928         flush_work(&vi->config_work);
2929
2930         unregister_netdev(vi->dev);
2931
2932         remove_vq_common(vi);
2933
2934         free_netdev(vi->dev);
2935 }
2936
2937 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
2938 {
2939         struct virtnet_info *vi = vdev->priv;
2940
2941         virtnet_cpu_notif_remove(vi);
2942         virtnet_freeze_down(vdev);
2943         remove_vq_common(vi);
2944
2945         return 0;
2946 }
2947
2948 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
2949 {
2950         struct virtnet_info *vi = vdev->priv;
2951         int err;
2952
2953         err = virtnet_restore_up(vdev);
2954         if (err)
2955                 return err;
2956         virtnet_set_queues(vi, vi->curr_queue_pairs);
2957
2958         err = virtnet_cpu_notif_add(vi);
2959         if (err)
2960                 return err;
2961
2962         return 0;
2963 }
2964
2965 static struct virtio_device_id id_table[] = {
2966         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2967         { 0 },
2968 };
2969
2970 #define VIRTNET_FEATURES \
2971         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2972         VIRTIO_NET_F_MAC, \
2973         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2974         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2975         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2976         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2977         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2978         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2979         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2980         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
2981         VIRTIO_NET_F_SPEED_DUPLEX
2982
2983 static unsigned int features[] = {
2984         VIRTNET_FEATURES,
2985 };
2986
2987 static unsigned int features_legacy[] = {
2988         VIRTNET_FEATURES,
2989         VIRTIO_NET_F_GSO,
2990         VIRTIO_F_ANY_LAYOUT,
2991 };
2992
2993 static struct virtio_driver virtio_net_driver = {
2994         .feature_table = features,
2995         .feature_table_size = ARRAY_SIZE(features),
2996         .feature_table_legacy = features_legacy,
2997         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2998         .driver.name =  KBUILD_MODNAME,
2999         .driver.owner = THIS_MODULE,
3000         .id_table =     id_table,
3001         .validate =     virtnet_validate,
3002         .probe =        virtnet_probe,
3003         .remove =       virtnet_remove,
3004         .config_changed = virtnet_config_changed,
3005 #ifdef CONFIG_PM_SLEEP
3006         .freeze =       virtnet_freeze,
3007         .restore =      virtnet_restore,
3008 #endif
3009 };
3010
3011 static __init int virtio_net_driver_init(void)
3012 {
3013         int ret;
3014
3015         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3016                                       virtnet_cpu_online,
3017                                       virtnet_cpu_down_prep);
3018         if (ret < 0)
3019                 goto out;
3020         virtionet_online = ret;
3021         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3022                                       NULL, virtnet_cpu_dead);
3023         if (ret)
3024                 goto err_dead;
3025
3026         ret = register_virtio_driver(&virtio_net_driver);
3027         if (ret)
3028                 goto err_virtio;
3029         return 0;
3030 err_virtio:
3031         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3032 err_dead:
3033         cpuhp_remove_multi_state(virtionet_online);
3034 out:
3035         return ret;
3036 }
3037 module_init(virtio_net_driver_init);
3038
3039 static __exit void virtio_net_driver_exit(void)
3040 {
3041         unregister_virtio_driver(&virtio_net_driver);
3042         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3043         cpuhp_remove_multi_state(virtionet_online);
3044 }
3045 module_exit(virtio_net_driver_exit);
3046
3047 MODULE_DEVICE_TABLE(virtio, id_table);
3048 MODULE_DESCRIPTION("Virtio network driver");
3049 MODULE_LICENSE("GPL");