drivers/net/virtio_net.c

   1 /* A network driver using virtio.
   2  *
   3  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation; either version 2 of the License, or
   8  * (at your option) any later version.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, see <http://www.gnu.org/licenses/>.
  17  */
  18 //#define DEBUG
  19 #include <linux/netdevice.h>
  20 #include <linux/etherdevice.h>
  21 #include <linux/ethtool.h>
  22 #include <linux/module.h>
  23 #include <linux/virtio.h>
  24 #include <linux/virtio_net.h>
  25 #include <linux/bpf.h>
  26 #include <linux/bpf_trace.h>
  27 #include <linux/scatterlist.h>
  28 #include <linux/if_vlan.h>
  29 #include <linux/slab.h>
  30 #include <linux/cpu.h>
  31 #include <linux/average.h>
  32 #include <linux/filter.h>
  33 #include <net/route.h>
  34 #include <net/xdp.h>
  35
  36 static int napi_weight = NAPI_POLL_WEIGHT;
  37 module_param(napi_weight, int, 0444);
  38
  39 static bool csum = true, gso = true, napi_tx;
  40 module_param(csum, bool, 0444);
  41 module_param(gso, bool, 0444);
  42 module_param(napi_tx, bool, 0644);
  43
  44 /* FIXME: MTU in config. */
  45 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  46 #define GOOD_COPY_LEN   128
  47
  48 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  49
  50 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  51 #define VIRTIO_XDP_HEADROOM 256
  52
  53 /* RX packet size EWMA. The average packet size is used to determine the packet
  54  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  55  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  56  * term, transient changes in packet size.
  57  */
  58 DECLARE_EWMA(pkt_len, 0, 64)
  59
  60 #define VIRTNET_DRIVER_VERSION "1.0.0"
  61
  62 static const unsigned long guest_offloads[] = {
  63         VIRTIO_NET_F_GUEST_TSO4,
  64         VIRTIO_NET_F_GUEST_TSO6,
  65         VIRTIO_NET_F_GUEST_ECN,
  66         VIRTIO_NET_F_GUEST_UFO
  67 };
  68
  69 struct virtnet_stat_desc {
  70         char desc[ETH_GSTRING_LEN];
  71         size_t offset;
  72 };
  73
  74 struct virtnet_sq_stats {
  75         struct u64_stats_sync syncp;
  76         u64 packets;
  77         u64 bytes;
  78 };
  79
  80 struct virtnet_rq_stats {
  81         struct u64_stats_sync syncp;
  82         u64 packets;
  83         u64 bytes;
  84 };
  85
  86 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
  87 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
  88
  89 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
  90         { "packets",    VIRTNET_SQ_STAT(packets) },
  91         { "bytes",      VIRTNET_SQ_STAT(bytes) },
  92 };
  93
  94 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
  95         { "packets",    VIRTNET_RQ_STAT(packets) },
  96         { "bytes",      VIRTNET_RQ_STAT(bytes) },
  97 };
  98
  99 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 100 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 101
 102 /* Internal representation of a send virtqueue */
 103 struct send_queue {
 104         /* Virtqueue associated with this send _queue */
 105         struct virtqueue *vq;
 106
 107         /* TX: fragments + linear part + virtio header */
 108         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 109
 110         /* Name of the send queue: output.$index */
 111         char name[40];
 112
 113         struct virtnet_sq_stats stats;
 114
 115         struct napi_struct napi;
 116 };
 117
 118 /* Internal representation of a receive virtqueue */
 119 struct receive_queue {
 120         /* Virtqueue associated with this receive_queue */
 121         struct virtqueue *vq;
 122
 123         struct napi_struct napi;
 124
 125         struct bpf_prog __rcu *xdp_prog;
 126
 127         struct virtnet_rq_stats stats;
 128
 129         /* Chain pages by the private ptr. */
 130         struct page *pages;
 131
 132         /* Average packet length for mergeable receive buffers. */
 133         struct ewma_pkt_len mrg_avg_pkt_len;
 134
 135         /* Page frag for packet buffer allocation. */
 136         struct page_frag alloc_frag;
 137
 138         /* RX: fragments + linear part + virtio header */
 139         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 140
 141         /* Min single buffer size for mergeable buffers case. */
 142         unsigned int min_buf_len;
 143
 144         /* Name of this receive queue: input.$index */
 145         char name[40];
 146
 147         struct xdp_rxq_info xdp_rxq;
 148 };
 149
 150 struct virtnet_info {
 151         struct virtio_device *vdev;
 152         struct virtqueue *cvq;
 153         struct net_device *dev;
 154         struct send_queue *sq;
 155         struct receive_queue *rq;
 156         unsigned int status;
 157
 158         /* Max # of queue pairs supported by the device */
 159         u16 max_queue_pairs;
 160
 161         /* # of queue pairs currently used by the driver */
 162         u16 curr_queue_pairs;
 163
 164         /* # of XDP queue pairs currently used by the driver */
 165         u16 xdp_queue_pairs;
 166
 167         /* I like... big packets and I cannot lie! */
 168         bool big_packets;
 169
 170         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 171         bool mergeable_rx_bufs;
 172
 173         /* Has control virtqueue */
 174         bool has_cvq;
 175
 176         /* Host can handle any s/g split between our header and packet data */
 177         bool any_header_sg;
 178
 179         /* Packet virtio header size */
 180         u8 hdr_len;
 181
 182         /* Work struct for refilling if we run low on memory. */
 183         struct delayed_work refill;
 184
 185         /* Work struct for config space updates */
 186         struct work_struct config_work;
 187
 188         /* Does the affinity hint is set for virtqueues? */
 189         bool affinity_hint_set;
 190
 191         /* CPU hotplug instances for online & dead */
 192         struct hlist_node node;
 193         struct hlist_node node_dead;
 194
 195         /* Control VQ buffers: protected by the rtnl lock */
 196         struct virtio_net_ctrl_hdr ctrl_hdr;
 197         virtio_net_ctrl_ack ctrl_status;
 198         struct virtio_net_ctrl_mq ctrl_mq;
 199         u8 ctrl_promisc;
 200         u8 ctrl_allmulti;
 201         u16 ctrl_vid;
 202         u64 ctrl_offloads;
 203
 204         /* Ethtool settings */
 205         u8 duplex;
 206         u32 speed;
 207
 208         unsigned long guest_offloads;
 209 };
 210
 211 struct padded_vnet_hdr {
 212         struct virtio_net_hdr_mrg_rxbuf hdr;
 213         /*
 214          * hdr is in a separate sg buffer, and data sg buffer shares same page
 215          * with this header sg. This padding makes next sg 16 byte aligned
 216          * after the header.
 217          */
 218         char padding[4];
 219 };
 220
 221 /* Converting between virtqueue no. and kernel tx/rx queue no.
 222  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 223  */
 224 static int vq2txq(struct virtqueue *vq)
 225 {
 226         return (vq->index - 1) / 2;
 227 }
 228
 229 static int txq2vq(int txq)
 230 {
 231         return txq * 2 + 1;
 232 }
 233
 234 static int vq2rxq(struct virtqueue *vq)
 235 {
 236         return vq->index / 2;
 237 }
 238
 239 static int rxq2vq(int rxq)
 240 {
 241         return rxq * 2;
 242 }
 243
 244 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 245 {
 246         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 247 }
 248
 249 /*
 250  * private is used to chain pages for big packets, put the whole
 251  * most recent used list in the beginning for reuse
 252  */
 253 static void give_pages(struct receive_queue *rq, struct page *page)
 254 {
 255         struct page *end;
 256
 257         /* Find end of list, sew whole thing into vi->rq.pages. */
 258         for (end = page; end->private; end = (struct page *)end->private);
 259         end->private = (unsigned long)rq->pages;
 260         rq->pages = page;
 261 }
 262
 263 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 264 {
 265         struct page *p = rq->pages;
 266
 267         if (p) {
 268                 rq->pages = (struct page *)p->private;
 269                 /* clear private here, it is used to chain pages */
 270                 p->private = 0;
 271         } else
 272                 p = alloc_page(gfp_mask);
 273         return p;
 274 }
 275
 276 static void virtqueue_napi_schedule(struct napi_struct *napi,
 277                                     struct virtqueue *vq)
 278 {
 279         if (napi_schedule_prep(napi)) {
 280                 virtqueue_disable_cb(vq);
 281                 __napi_schedule(napi);
 282         }
 283 }
 284
 285 static void virtqueue_napi_complete(struct napi_struct *napi,
 286                                     struct virtqueue *vq, int processed)
 287 {
 288         int opaque;
 289
 290         opaque = virtqueue_enable_cb_prepare(vq);
 291         if (napi_complete_done(napi, processed)) {
 292                 if (unlikely(virtqueue_poll(vq, opaque)))
 293                         virtqueue_napi_schedule(napi, vq);
 294         } else {
 295                 virtqueue_disable_cb(vq);
 296         }
 297 }
 298
 299 static void skb_xmit_done(struct virtqueue *vq)
 300 {
 301         struct virtnet_info *vi = vq->vdev->priv;
 302         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 303
 304         /* Suppress further interrupts. */
 305         virtqueue_disable_cb(vq);
 306
 307         if (napi->weight)
 308                 virtqueue_napi_schedule(napi, vq);
 309         else
 310                 /* We were probably waiting for more output buffers. */
 311                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 312 }
 313
 314 #define MRG_CTX_HEADER_SHIFT 22
 315 static void *mergeable_len_to_ctx(unsigned int truesize,
 316                                   unsigned int headroom)
 317 {
 318         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 319 }
 320
 321 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 322 {
 323         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 324 }
 325
 326 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 327 {
 328         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 329 }
 330
 331 /* Called from bottom half context */
 332 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 333                                    struct receive_queue *rq,
 334                                    struct page *page, unsigned int offset,
 335                                    unsigned int len, unsigned int truesize)
 336 {
 337         struct sk_buff *skb;
 338         struct virtio_net_hdr_mrg_rxbuf *hdr;
 339         unsigned int copy, hdr_len, hdr_padded_len;
 340         char *p;
 341
 342         p = page_address(page) + offset;
 343
 344         /* copy small packet so we can reuse these pages for small data */
 345         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 346         if (unlikely(!skb))
 347                 return NULL;
 348
 349         hdr = skb_vnet_hdr(skb);
 350
 351         hdr_len = vi->hdr_len;
 352         if (vi->mergeable_rx_bufs)
 353                 hdr_padded_len = sizeof(*hdr);
 354         else
 355                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 356
 357         memcpy(hdr, p, hdr_len);
 358
 359         len -= hdr_len;
 360         offset += hdr_padded_len;
 361         p += hdr_padded_len;
 362
 363         copy = len;
 364         if (copy > skb_tailroom(skb))
 365                 copy = skb_tailroom(skb);
 366         skb_put_data(skb, p, copy);
 367
 368         len -= copy;
 369         offset += copy;
 370
 371         if (vi->mergeable_rx_bufs) {
 372                 if (len)
 373                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 374                 else
 375                         put_page(page);
 376                 return skb;
 377         }
 378
 379         /*
 380          * Verify that we can indeed put this data into a skb.
 381          * This is here to handle cases when the device erroneously
 382          * tries to receive more than is possible. This is usually
 383          * the case of a broken device.
 384          */
 385         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 386                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 387                 dev_kfree_skb(skb);
 388                 return NULL;
 389         }
 390         BUG_ON(offset >= PAGE_SIZE);
 391         while (len) {
 392                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 393                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 394                                 frag_size, truesize);
 395                 len -= frag_size;
 396                 page = (struct page *)page->private;
 397                 offset = 0;
 398         }
 399
 400         if (page)
 401                 give_pages(rq, page);
 402
 403         return skb;
 404 }
 405
 406 static void virtnet_xdp_flush(struct net_device *dev)
 407 {
 408         struct virtnet_info *vi = netdev_priv(dev);
 409         struct send_queue *sq;
 410         unsigned int qp;
 411
 412         qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
 413         sq = &vi->sq[qp];
 414
 415         virtqueue_kick(sq->vq);
 416 }
 417
 418 static int __virtnet_xdp_xmit(struct virtnet_info *vi,
 419                               struct xdp_buff *xdp)
 420 {
 421         struct virtio_net_hdr_mrg_rxbuf *hdr;
 422         struct xdp_frame *xdpf, *xdpf_sent;
 423         struct send_queue *sq;
 424         unsigned int len;
 425         unsigned int qp;
 426         int err;
 427
 428         qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
 429         sq = &vi->sq[qp];
 430
 431         /* Free up any pending old buffers before queueing new ones. */
 432         while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL)
 433                 xdp_return_frame(xdpf_sent->data, &xdpf_sent->mem);
 434
 435         xdpf = convert_to_xdp_frame(xdp);
 436         if (unlikely(!xdpf))
 437                 return -EOVERFLOW;
 438
 439         /* virtqueue want to use data area in-front of packet */
 440         if (unlikely(xdpf->metasize > 0))
 441                 return -EOPNOTSUPP;
 442
 443         if (unlikely(xdpf->headroom < vi->hdr_len))
 444                 return -EOVERFLOW;
 445
 446         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 447         xdpf->data -= vi->hdr_len;
 448         /* Zero header and leave csum up to XDP layers */
 449         hdr = xdpf->data;
 450         memset(hdr, 0, vi->hdr_len);
 451         xdpf->len   += vi->hdr_len;
 452
 453         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 454
 455         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdpf, GFP_ATOMIC);
 456         if (unlikely(err))
 457                 return -ENOSPC; /* Caller handle free/refcnt */
 458
 459         return 0;
 460 }
 461
 462 static int virtnet_xdp_xmit(struct net_device *dev, struct xdp_buff *xdp)
 463 {
 464         struct virtnet_info *vi = netdev_priv(dev);
 465         struct receive_queue *rq = vi->rq;
 466         struct bpf_prog *xdp_prog;
 467
 468         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 469          * indicate XDP resources have been successfully allocated.
 470          */
 471         xdp_prog = rcu_dereference(rq->xdp_prog);
 472         if (!xdp_prog)
 473                 return -ENXIO;
 474
 475         return __virtnet_xdp_xmit(vi, xdp);
 476 }
 477
 478 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 479 {
 480         return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
 481 }
 482
 483 /* We copy the packet for XDP in the following cases:
 484  *
 485  * 1) Packet is scattered across multiple rx buffers.
 486  * 2) Headroom space is insufficient.
 487  *
 488  * This is inefficient but it's a temporary condition that
 489  * we hit right after XDP is enabled and until queue is refilled
 490  * with large buffers with sufficient headroom - so it should affect
 491  * at most queue size packets.
 492  * Afterwards, the conditions to enable
 493  * XDP should preclude the underlying device from sending packets
 494  * across multiple buffers (num_buf > 1), and we make sure buffers
 495  * have enough headroom.
 496  */
 497 static struct page *xdp_linearize_page(struct receive_queue *rq,
 498                                        u16 *num_buf,
 499                                        struct page *p,
 500                                        int offset,
 501                                        int page_off,
 502                                        unsigned int *len)
 503 {
 504         struct page *page = alloc_page(GFP_ATOMIC);
 505
 506         if (!page)
 507                 return NULL;
 508
 509         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 510         page_off += *len;
 511
 512         while (--*num_buf) {
 513                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 514                 unsigned int buflen;
 515                 void *buf;
 516                 int off;
 517
 518                 buf = virtqueue_get_buf(rq->vq, &buflen);
 519                 if (unlikely(!buf))
 520                         goto err_buf;
 521
 522                 p = virt_to_head_page(buf);
 523                 off = buf - page_address(p);
 524
 525                 /* guard against a misconfigured or uncooperative backend that
 526                  * is sending packet larger than the MTU.
 527                  */
 528                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 529                         put_page(p);
 530                         goto err_buf;
 531                 }
 532
 533                 memcpy(page_address(page) + page_off,
 534                        page_address(p) + off, buflen);
 535                 page_off += buflen;
 536                 put_page(p);
 537         }
 538
 539         /* Headroom does not contribute to packet length */
 540         *len = page_off - VIRTIO_XDP_HEADROOM;
 541         return page;
 542 err_buf:
 543         __free_pages(page, 0);
 544         return NULL;
 545 }
 546
 547 static struct sk_buff *receive_small(struct net_device *dev,
 548                                      struct virtnet_info *vi,
 549                                      struct receive_queue *rq,
 550                                      void *buf, void *ctx,
 551                                      unsigned int len,
 552                                      bool *xdp_xmit)
 553 {
 554         struct sk_buff *skb;
 555         struct bpf_prog *xdp_prog;
 556         unsigned int xdp_headroom = (unsigned long)ctx;
 557         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 558         unsigned int headroom = vi->hdr_len + header_offset;
 559         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 560                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 561         struct page *page = virt_to_head_page(buf);
 562         unsigned int delta = 0;
 563         struct page *xdp_page;
 564         int err;
 565
 566         len -= vi->hdr_len;
 567
 568         rcu_read_lock();
 569         xdp_prog = rcu_dereference(rq->xdp_prog);
 570         if (xdp_prog) {
 571                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 572                 struct xdp_buff xdp;
 573                 void *orig_data;
 574                 u32 act;
 575
 576                 if (unlikely(hdr->hdr.gso_type))
 577                         goto err_xdp;
 578
 579                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 580                         int offset = buf - page_address(page) + header_offset;
 581                         unsigned int tlen = len + vi->hdr_len;
 582                         u16 num_buf = 1;
 583
 584                         xdp_headroom = virtnet_get_headroom(vi);
 585                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 586                         headroom = vi->hdr_len + header_offset;
 587                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 588                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 589                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 590                                                       offset, header_offset,
 591                                                       &tlen);
 592                         if (!xdp_page)
 593                                 goto err_xdp;
 594
 595                         buf = page_address(xdp_page);
 596                         put_page(page);
 597                         page = xdp_page;
 598                 }
 599
 600                 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
 601                 xdp.data = xdp.data_hard_start + xdp_headroom;
 602                 xdp_set_data_meta_invalid(&xdp);
 603                 xdp.data_end = xdp.data + len;
 604                 xdp.rxq = &rq->xdp_rxq;
 605                 orig_data = xdp.data;
 606                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 607
 608                 switch (act) {
 609                 case XDP_PASS:
 610                         /* Recalculate length in case bpf program changed it */
 611                         delta = orig_data - xdp.data;
 612                         break;
 613                 case XDP_TX:
 614                         err = __virtnet_xdp_xmit(vi, &xdp);
 615                         if (unlikely(err)) {
 616                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 617                                 goto err_xdp;
 618                         }
 619                         *xdp_xmit = true;
 620                         rcu_read_unlock();
 621                         goto xdp_xmit;
 622                 case XDP_REDIRECT:
 623                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 624                         if (err)
 625                                 goto err_xdp;
 626                         *xdp_xmit = true;
 627                         rcu_read_unlock();
 628                         goto xdp_xmit;
 629                 default:
 630                         bpf_warn_invalid_xdp_action(act);
 631                 case XDP_ABORTED:
 632                         trace_xdp_exception(vi->dev, xdp_prog, act);
 633                 case XDP_DROP:
 634                         goto err_xdp;
 635                 }
 636         }
 637         rcu_read_unlock();
 638
 639         skb = build_skb(buf, buflen);
 640         if (!skb) {
 641                 put_page(page);
 642                 goto err;
 643         }
 644         skb_reserve(skb, headroom - delta);
 645         skb_put(skb, len + delta);
 646         if (!delta) {
 647                 buf += header_offset;
 648                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 649         } /* keep zeroed vnet hdr since packet was changed by bpf */
 650
 651 err:
 652         return skb;
 653
 654 err_xdp:
 655         rcu_read_unlock();
 656         dev->stats.rx_dropped++;
 657         put_page(page);
 658 xdp_xmit:
 659         return NULL;
 660 }
 661
 662 static struct sk_buff *receive_big(struct net_device *dev,
 663                                    struct virtnet_info *vi,
 664                                    struct receive_queue *rq,
 665                                    void *buf,
 666                                    unsigned int len)
 667 {
 668         struct page *page = buf;
 669         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 670
 671         if (unlikely(!skb))
 672                 goto err;
 673
 674         return skb;
 675
 676 err:
 677         dev->stats.rx_dropped++;
 678         give_pages(rq, page);
 679         return NULL;
 680 }
 681
 682 static struct sk_buff *receive_mergeable(struct net_device *dev,
 683                                          struct virtnet_info *vi,
 684                                          struct receive_queue *rq,
 685                                          void *buf,
 686                                          void *ctx,
 687                                          unsigned int len,
 688                                          bool *xdp_xmit)
 689 {
 690         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 691         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 692         struct page *page = virt_to_head_page(buf);
 693         int offset = buf - page_address(page);
 694         struct sk_buff *head_skb, *curr_skb;
 695         struct bpf_prog *xdp_prog;
 696         unsigned int truesize;
 697         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 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                 /* This happens when rx buffer size is underestimated
 711                  * or headroom is not enough because of the buffer
 712                  * was refilled before XDP is set. This should only
 713                  * happen for the first several packets, so we don't
 714                  * care much about its performance.
 715                  */
 716                 if (unlikely(num_buf > 1 ||
 717                              headroom < virtnet_get_headroom(vi))) {
 718                         /* linearize data for XDP */
 719                         xdp_page = xdp_linearize_page(rq, &num_buf,
 720                                                       page, offset,
 721                                                       VIRTIO_XDP_HEADROOM,
 722                                                       &len);
 723                         if (!xdp_page)
 724                                 goto err_xdp;
 725                         offset = VIRTIO_XDP_HEADROOM;
 726                 } else {
 727                         xdp_page = page;
 728                 }
 729
 730                 /* Transient failure which in theory could occur if
 731                  * in-flight packets from before XDP was enabled reach
 732                  * the receive path after XDP is loaded. In practice I
 733                  * was not able to create this condition.
 734                  */
 735                 if (unlikely(hdr->hdr.gso_type))
 736                         goto err_xdp;
 737
 738                 /* Allow consuming headroom but reserve enough space to push
 739                  * the descriptor on if we get an XDP_TX return code.
 740                  */
 741                 data = page_address(xdp_page) + offset;
 742                 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
 743                 xdp.data = data + vi->hdr_len;
 744                 xdp_set_data_meta_invalid(&xdp);
 745                 xdp.data_end = xdp.data + (len - vi->hdr_len);
 746                 xdp.rxq = &rq->xdp_rxq;
 747
 748                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 749
 750                 switch (act) {
 751                 case XDP_PASS:
 752                         /* recalculate offset to account for any header
 753                          * adjustments. Note other cases do not build an
 754                          * skb and avoid using offset
 755                          */
 756                         offset = xdp.data -
 757                                         page_address(xdp_page) - vi->hdr_len;
 758
 759                         /* We can only create skb based on xdp_page. */
 760                         if (unlikely(xdp_page != page)) {
 761                                 rcu_read_unlock();
 762                                 put_page(page);
 763                                 head_skb = page_to_skb(vi, rq, xdp_page,
 764                                                        offset, len, PAGE_SIZE);
 765                                 return head_skb;
 766                         }
 767                         break;
 768                 case XDP_TX:
 769                         err = __virtnet_xdp_xmit(vi, &xdp);
 770                         if (unlikely(err)) {
 771                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 772                                 if (unlikely(xdp_page != page))
 773                                         put_page(xdp_page);
 774                                 goto err_xdp;
 775                         }
 776                         *xdp_xmit = true;
 777                         if (unlikely(xdp_page != page))
 778                                 goto err_xdp;
 779                         rcu_read_unlock();
 780                         goto xdp_xmit;
 781                 case XDP_REDIRECT:
 782                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 783                         if (err) {
 784                                 if (unlikely(xdp_page != page))
 785                                         put_page(xdp_page);
 786                                 goto err_xdp;
 787                         }
 788                         *xdp_xmit = true;
 789                         if (unlikely(xdp_page != page))
 790                                 goto err_xdp;
 791                         rcu_read_unlock();
 792                         goto xdp_xmit;
 793                 default:
 794                         bpf_warn_invalid_xdp_action(act);
 795                 case XDP_ABORTED:
 796                         trace_xdp_exception(vi->dev, xdp_prog, act);
 797                 case XDP_DROP:
 798                         if (unlikely(xdp_page != page))
 799                                 __free_pages(xdp_page, 0);
 800                         goto err_xdp;
 801                 }
 802         }
 803         rcu_read_unlock();
 804
 805         truesize = mergeable_ctx_to_truesize(ctx);
 806         if (unlikely(len > truesize)) {
 807                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 808                          dev->name, len, (unsigned long)ctx);
 809                 dev->stats.rx_length_errors++;
 810                 goto err_skb;
 811         }
 812
 813         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
 814         curr_skb = head_skb;
 815
 816         if (unlikely(!curr_skb))
 817                 goto err_skb;
 818         while (--num_buf) {
 819                 int num_skb_frags;
 820
 821                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
 822                 if (unlikely(!buf)) {
 823                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 824                                  dev->name, num_buf,
 825                                  virtio16_to_cpu(vi->vdev,
 826                                                  hdr->num_buffers));
 827                         dev->stats.rx_length_errors++;
 828                         goto err_buf;
 829                 }
 830
 831                 page = virt_to_head_page(buf);
 832
 833                 truesize = mergeable_ctx_to_truesize(ctx);
 834                 if (unlikely(len > truesize)) {
 835                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 836                                  dev->name, len, (unsigned long)ctx);
 837                         dev->stats.rx_length_errors++;
 838                         goto err_skb;
 839                 }
 840
 841                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 842                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 843                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 844
 845                         if (unlikely(!nskb))
 846                                 goto err_skb;
 847                         if (curr_skb == head_skb)
 848                                 skb_shinfo(curr_skb)->frag_list = nskb;
 849                         else
 850                                 curr_skb->next = nskb;
 851                         curr_skb = nskb;
 852                         head_skb->truesize += nskb->truesize;
 853                         num_skb_frags = 0;
 854                 }
 855                 if (curr_skb != head_skb) {
 856                         head_skb->data_len += len;
 857                         head_skb->len += len;
 858                         head_skb->truesize += truesize;
 859                 }
 860                 offset = buf - page_address(page);
 861                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 862                         put_page(page);
 863                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 864                                              len, truesize);
 865                 } else {
 866                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 867                                         offset, len, truesize);
 868                 }
 869         }
 870
 871         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 872         return head_skb;
 873
 874 err_xdp:
 875         rcu_read_unlock();
 876 err_skb:
 877         put_page(page);
 878         while (--num_buf) {
 879                 buf = virtqueue_get_buf(rq->vq, &len);
 880                 if (unlikely(!buf)) {
 881                         pr_debug("%s: rx error: %d buffers missing\n",
 882                                  dev->name, num_buf);
 883                         dev->stats.rx_length_errors++;
 884                         break;
 885                 }
 886                 page = virt_to_head_page(buf);
 887                 put_page(page);
 888         }
 889 err_buf:
 890         dev->stats.rx_dropped++;
 891         dev_kfree_skb(head_skb);
 892 xdp_xmit:
 893         return NULL;
 894 }
 895
 896 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 897                        void *buf, unsigned int len, void **ctx, bool *xdp_xmit)
 898 {
 899         struct net_device *dev = vi->dev;
 900         struct sk_buff *skb;
 901         struct virtio_net_hdr_mrg_rxbuf *hdr;
 902         int ret;
 903
 904         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 905                 pr_debug("%s: short packet %i\n", dev->name, len);
 906                 dev->stats.rx_length_errors++;
 907                 if (vi->mergeable_rx_bufs) {
 908                         put_page(virt_to_head_page(buf));
 909                 } else if (vi->big_packets) {
 910                         give_pages(rq, buf);
 911                 } else {
 912                         put_page(virt_to_head_page(buf));
 913                 }
 914                 return 0;
 915         }
 916
 917         if (vi->mergeable_rx_bufs)
 918                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit);
 919         else if (vi->big_packets)
 920                 skb = receive_big(dev, vi, rq, buf, len);
 921         else
 922                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit);
 923
 924         if (unlikely(!skb))
 925                 return 0;
 926
 927         hdr = skb_vnet_hdr(skb);
 928
 929         ret = skb->len;
 930
 931         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 932                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 933
 934         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 935                                   virtio_is_little_endian(vi->vdev))) {
 936                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 937                                      dev->name, hdr->hdr.gso_type,
 938                                      hdr->hdr.gso_size);
 939                 goto frame_err;
 940         }
 941
 942         skb->protocol = eth_type_trans(skb, dev);
 943         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 944                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 945
 946         napi_gro_receive(&rq->napi, skb);
 947         return ret;
 948
 949 frame_err:
 950         dev->stats.rx_frame_errors++;
 951         dev_kfree_skb(skb);
 952         return 0;
 953 }
 954
 955 /* Unlike mergeable buffers, all buffers are allocated to the
 956  * same size, except for the headroom. For this reason we do
 957  * not need to use  mergeable_len_to_ctx here - it is enough
 958  * to store the headroom as the context ignoring the truesize.
 959  */
 960 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 961                              gfp_t gfp)
 962 {
 963         struct page_frag *alloc_frag = &rq->alloc_frag;
 964         char *buf;
 965         unsigned int xdp_headroom = virtnet_get_headroom(vi);
 966         void *ctx = (void *)(unsigned long)xdp_headroom;
 967         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
 968         int err;
 969
 970         len = SKB_DATA_ALIGN(len) +
 971               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 972         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 973                 return -ENOMEM;
 974
 975         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 976         get_page(alloc_frag->page);
 977         alloc_frag->offset += len;
 978         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
 979                     vi->hdr_len + GOOD_PACKET_LEN);
 980         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
 981         if (err < 0)
 982                 put_page(virt_to_head_page(buf));
 983         return err;
 984 }
 985
 986 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 987                            gfp_t gfp)
 988 {
 989         struct page *first, *list = NULL;
 990         char *p;
 991         int i, err, offset;
 992
 993         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 994
 995         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 996         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 997                 first = get_a_page(rq, gfp);
 998                 if (!first) {
 999                         if (list)
1000                                 give_pages(rq, list);
1001                         return -ENOMEM;
1002                 }
1003                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1004
1005                 /* chain new page in list head to match sg */
1006                 first->private = (unsigned long)list;
1007                 list = first;
1008         }
1009
1010         first = get_a_page(rq, gfp);
1011         if (!first) {
1012                 give_pages(rq, list);
1013                 return -ENOMEM;
1014         }
1015         p = page_address(first);
1016
1017         /* rq->sg[0], rq->sg[1] share the same page */
1018         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1019         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1020
1021         /* rq->sg[1] for data packet, from offset */
1022         offset = sizeof(struct padded_vnet_hdr);
1023         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1024
1025         /* chain first in list head */
1026         first->private = (unsigned long)list;
1027         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1028                                   first, gfp);
1029         if (err < 0)
1030                 give_pages(rq, first);
1031
1032         return err;
1033 }
1034
1035 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1036                                           struct ewma_pkt_len *avg_pkt_len,
1037                                           unsigned int room)
1038 {
1039         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1040         unsigned int len;
1041
1042         if (room)
1043                 return PAGE_SIZE - room;
1044
1045         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1046                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1047
1048         return ALIGN(len, L1_CACHE_BYTES);
1049 }
1050
1051 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1052                                  struct receive_queue *rq, gfp_t gfp)
1053 {
1054         struct page_frag *alloc_frag = &rq->alloc_frag;
1055         unsigned int headroom = virtnet_get_headroom(vi);
1056         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1057         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1058         char *buf;
1059         void *ctx;
1060         int err;
1061         unsigned int len, hole;
1062
1063         /* Extra tailroom is needed to satisfy XDP's assumption. This
1064          * means rx frags coalescing won't work, but consider we've
1065          * disabled GSO for XDP, it won't be a big issue.
1066          */
1067         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1068         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1069                 return -ENOMEM;
1070
1071         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1072         buf += headroom; /* advance address leaving hole at front of pkt */
1073         get_page(alloc_frag->page);
1074         alloc_frag->offset += len + room;
1075         hole = alloc_frag->size - alloc_frag->offset;
1076         if (hole < len + room) {
1077                 /* To avoid internal fragmentation, if there is very likely not
1078                  * enough space for another buffer, add the remaining space to
1079                  * the current buffer.
1080                  */
1081                 len += hole;
1082                 alloc_frag->offset += hole;
1083         }
1084
1085         sg_init_one(rq->sg, buf, len);
1086         ctx = mergeable_len_to_ctx(len, headroom);
1087         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1088         if (err < 0)
1089                 put_page(virt_to_head_page(buf));
1090
1091         return err;
1092 }
1093
1094 /*
1095  * Returns false if we couldn't fill entirely (OOM).
1096  *
1097  * Normally run in the receive path, but can also be run from ndo_open
1098  * before we're receiving packets, or from refill_work which is
1099  * careful to disable receiving (using napi_disable).
1100  */
1101 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1102                           gfp_t gfp)
1103 {
1104         int err;
1105         bool oom;
1106
1107         do {
1108                 if (vi->mergeable_rx_bufs)
1109                         err = add_recvbuf_mergeable(vi, rq, gfp);
1110                 else if (vi->big_packets)
1111                         err = add_recvbuf_big(vi, rq, gfp);
1112                 else
1113                         err = add_recvbuf_small(vi, rq, gfp);
1114
1115                 oom = err == -ENOMEM;
1116                 if (err)
1117                         break;
1118         } while (rq->vq->num_free);
1119         virtqueue_kick(rq->vq);
1120         return !oom;
1121 }
1122
1123 static void skb_recv_done(struct virtqueue *rvq)
1124 {
1125         struct virtnet_info *vi = rvq->vdev->priv;
1126         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1127
1128         virtqueue_napi_schedule(&rq->napi, rvq);
1129 }
1130
1131 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1132 {
1133         napi_enable(napi);
1134
1135         /* If all buffers were filled by other side before we napi_enabled, we
1136          * won't get another interrupt, so process any outstanding packets now.
1137          * Call local_bh_enable after to trigger softIRQ processing.
1138          */
1139         local_bh_disable();
1140         virtqueue_napi_schedule(napi, vq);
1141         local_bh_enable();
1142 }
1143
1144 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1145                                    struct virtqueue *vq,
1146                                    struct napi_struct *napi)
1147 {
1148         if (!napi->weight)
1149                 return;
1150
1151         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1152          * enable the feature if this is likely affine with the transmit path.
1153          */
1154         if (!vi->affinity_hint_set) {
1155                 napi->weight = 0;
1156                 return;
1157         }
1158
1159         return virtnet_napi_enable(vq, napi);
1160 }
1161
1162 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1163 {
1164         if (napi->weight)
1165                 napi_disable(napi);
1166 }
1167
1168 static void refill_work(struct work_struct *work)
1169 {
1170         struct virtnet_info *vi =
1171                 container_of(work, struct virtnet_info, refill.work);
1172         bool still_empty;
1173         int i;
1174
1175         for (i = 0; i < vi->curr_queue_pairs; i++) {
1176                 struct receive_queue *rq = &vi->rq[i];
1177
1178                 napi_disable(&rq->napi);
1179                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1180                 virtnet_napi_enable(rq->vq, &rq->napi);
1181
1182                 /* In theory, this can happen: if we don't get any buffers in
1183                  * we will *never* try to fill again.
1184                  */
1185                 if (still_empty)
1186                         schedule_delayed_work(&vi->refill, HZ/2);
1187         }
1188 }
1189
1190 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit)
1191 {
1192         struct virtnet_info *vi = rq->vq->vdev->priv;
1193         unsigned int len, received = 0, bytes = 0;
1194         void *buf;
1195
1196         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1197                 void *ctx;
1198
1199                 while (received < budget &&
1200                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1201                         bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit);
1202                         received++;
1203                 }
1204         } else {
1205                 while (received < budget &&
1206                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1207                         bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit);
1208                         received++;
1209                 }
1210         }
1211
1212         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1213                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1214                         schedule_delayed_work(&vi->refill, 0);
1215         }
1216
1217         u64_stats_update_begin(&rq->stats.syncp);
1218         rq->stats.bytes += bytes;
1219         rq->stats.packets += received;
1220         u64_stats_update_end(&rq->stats.syncp);
1221
1222         return received;
1223 }
1224
1225 static void free_old_xmit_skbs(struct send_queue *sq)
1226 {
1227         struct sk_buff *skb;
1228         unsigned int len;
1229         unsigned int packets = 0;
1230         unsigned int bytes = 0;
1231
1232         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1233                 pr_debug("Sent skb %p\n", skb);
1234
1235                 bytes += skb->len;
1236                 packets++;
1237
1238                 dev_consume_skb_any(skb);
1239         }
1240
1241         /* Avoid overhead when no packets have been processed
1242          * happens when called speculatively from start_xmit.
1243          */
1244         if (!packets)
1245                 return;
1246
1247         u64_stats_update_begin(&sq->stats.syncp);
1248         sq->stats.bytes += bytes;
1249         sq->stats.packets += packets;
1250         u64_stats_update_end(&sq->stats.syncp);
1251 }
1252
1253 static void virtnet_poll_cleantx(struct receive_queue *rq)
1254 {
1255         struct virtnet_info *vi = rq->vq->vdev->priv;
1256         unsigned int index = vq2rxq(rq->vq);
1257         struct send_queue *sq = &vi->sq[index];
1258         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1259
1260         if (!sq->napi.weight)
1261                 return;
1262
1263         if (__netif_tx_trylock(txq)) {
1264                 free_old_xmit_skbs(sq);
1265                 __netif_tx_unlock(txq);
1266         }
1267
1268         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1269                 netif_tx_wake_queue(txq);
1270 }
1271
1272 static int virtnet_poll(struct napi_struct *napi, int budget)
1273 {
1274         struct receive_queue *rq =
1275                 container_of(napi, struct receive_queue, napi);
1276         unsigned int received;
1277         bool xdp_xmit = false;
1278
1279         virtnet_poll_cleantx(rq);
1280
1281         received = virtnet_receive(rq, budget, &xdp_xmit);
1282
1283         /* Out of packets? */
1284         if (received < budget)
1285                 virtqueue_napi_complete(napi, rq->vq, received);
1286
1287         if (xdp_xmit)
1288                 xdp_do_flush_map();
1289
1290         return received;
1291 }
1292
1293 static int virtnet_open(struct net_device *dev)
1294 {
1295         struct virtnet_info *vi = netdev_priv(dev);
1296         int i, err;
1297
1298         for (i = 0; i < vi->max_queue_pairs; i++) {
1299                 if (i < vi->curr_queue_pairs)
1300                         /* Make sure we have some buffers: if oom use wq. */
1301                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1302                                 schedule_delayed_work(&vi->refill, 0);
1303
1304                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1305                 if (err < 0)
1306                         return err;
1307
1308                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1309                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1310         }
1311
1312         return 0;
1313 }
1314
1315 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1316 {
1317         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1318         struct virtnet_info *vi = sq->vq->vdev->priv;
1319         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1320
1321         __netif_tx_lock(txq, raw_smp_processor_id());
1322         free_old_xmit_skbs(sq);
1323         __netif_tx_unlock(txq);
1324
1325         virtqueue_napi_complete(napi, sq->vq, 0);
1326
1327         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1328                 netif_tx_wake_queue(txq);
1329
1330         return 0;
1331 }
1332
1333 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1334 {
1335         struct virtio_net_hdr_mrg_rxbuf *hdr;
1336         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1337         struct virtnet_info *vi = sq->vq->vdev->priv;
1338         int num_sg;
1339         unsigned hdr_len = vi->hdr_len;
1340         bool can_push;
1341
1342         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1343
1344         can_push = vi->any_header_sg &&
1345                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1346                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1347         /* Even if we can, don't push here yet as this would skew
1348          * csum_start offset below. */
1349         if (can_push)
1350                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1351         else
1352                 hdr = skb_vnet_hdr(skb);
1353
1354         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1355                                     virtio_is_little_endian(vi->vdev), false))
1356                 BUG();
1357
1358         if (vi->mergeable_rx_bufs)
1359                 hdr->num_buffers = 0;
1360
1361         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1362         if (can_push) {
1363                 __skb_push(skb, hdr_len);
1364                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1365                 if (unlikely(num_sg < 0))
1366                         return num_sg;
1367                 /* Pull header back to avoid skew in tx bytes calculations. */
1368                 __skb_pull(skb, hdr_len);
1369         } else {
1370                 sg_set_buf(sq->sg, hdr, hdr_len);
1371                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1372                 if (unlikely(num_sg < 0))
1373                         return num_sg;
1374                 num_sg++;
1375         }
1376         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1377 }
1378
1379 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1380 {
1381         struct virtnet_info *vi = netdev_priv(dev);
1382         int qnum = skb_get_queue_mapping(skb);
1383         struct send_queue *sq = &vi->sq[qnum];
1384         int err;
1385         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1386         bool kick = !skb->xmit_more;
1387         bool use_napi = sq->napi.weight;
1388
1389         /* Free up any pending old buffers before queueing new ones. */
1390         free_old_xmit_skbs(sq);
1391
1392         if (use_napi && kick)
1393                 virtqueue_enable_cb_delayed(sq->vq);
1394
1395         /* timestamp packet in software */
1396         skb_tx_timestamp(skb);
1397
1398         /* Try to transmit */
1399         err = xmit_skb(sq, skb);
1400
1401         /* This should not happen! */
1402         if (unlikely(err)) {
1403                 dev->stats.tx_fifo_errors++;
1404                 if (net_ratelimit())
1405                         dev_warn(&dev->dev,
1406                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1407                 dev->stats.tx_dropped++;
1408                 dev_kfree_skb_any(skb);
1409                 return NETDEV_TX_OK;
1410         }
1411
1412         /* Don't wait up for transmitted skbs to be freed. */
1413         if (!use_napi) {
1414                 skb_orphan(skb);
1415                 nf_reset(skb);
1416         }
1417
1418         /* If running out of space, stop queue to avoid getting packets that we
1419          * are then unable to transmit.
1420          * An alternative would be to force queuing layer to requeue the skb by
1421          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1422          * returned in a normal path of operation: it means that driver is not
1423          * maintaining the TX queue stop/start state properly, and causes
1424          * the stack to do a non-trivial amount of useless work.
1425          * Since most packets only take 1 or 2 ring slots, stopping the queue
1426          * early means 16 slots are typically wasted.
1427          */
1428         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1429                 netif_stop_subqueue(dev, qnum);
1430                 if (!use_napi &&
1431                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1432                         /* More just got used, free them then recheck. */
1433                         free_old_xmit_skbs(sq);
1434                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1435                                 netif_start_subqueue(dev, qnum);
1436                                 virtqueue_disable_cb(sq->vq);
1437                         }
1438                 }
1439         }
1440
1441         if (kick || netif_xmit_stopped(txq))
1442                 virtqueue_kick(sq->vq);
1443
1444         return NETDEV_TX_OK;
1445 }
1446
1447 /*
1448  * Send command via the control virtqueue and check status.  Commands
1449  * supported by the hypervisor, as indicated by feature bits, should
1450  * never fail unless improperly formatted.
1451  */
1452 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1453                                  struct scatterlist *out)
1454 {
1455         struct scatterlist *sgs[4], hdr, stat;
1456         unsigned out_num = 0, tmp;
1457
1458         /* Caller should know better */
1459         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1460
1461         vi->ctrl_status = ~0;
1462         vi->ctrl_hdr.class = class;
1463         vi->ctrl_hdr.cmd = cmd;
1464         /* Add header */
1465         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1466         sgs[out_num++] = &hdr;
1467
1468         if (out)
1469                 sgs[out_num++] = out;
1470
1471         /* Add return status. */
1472         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1473         sgs[out_num] = &stat;
1474
1475         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1476         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1477
1478         if (unlikely(!virtqueue_kick(vi->cvq)))
1479                 return vi->ctrl_status == VIRTIO_NET_OK;
1480
1481         /* Spin for a response, the kick causes an ioport write, trapping
1482          * into the hypervisor, so the request should be handled immediately.
1483          */
1484         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1485                !virtqueue_is_broken(vi->cvq))
1486                 cpu_relax();
1487
1488         return vi->ctrl_status == VIRTIO_NET_OK;
1489 }
1490
1491 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1492 {
1493         struct virtnet_info *vi = netdev_priv(dev);
1494         struct virtio_device *vdev = vi->vdev;
1495         int ret;
1496         struct sockaddr *addr;
1497         struct scatterlist sg;
1498
1499         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1500         if (!addr)
1501                 return -ENOMEM;
1502
1503         ret = eth_prepare_mac_addr_change(dev, addr);
1504         if (ret)
1505                 goto out;
1506
1507         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1508                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1509                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1510                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1511                         dev_warn(&vdev->dev,
1512                                  "Failed to set mac address by vq command.\n");
1513                         ret = -EINVAL;
1514                         goto out;
1515                 }
1516         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1517                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1518                 unsigned int i;
1519
1520                 /* Naturally, this has an atomicity problem. */
1521                 for (i = 0; i < dev->addr_len; i++)
1522                         virtio_cwrite8(vdev,
1523                                        offsetof(struct virtio_net_config, mac) +
1524                                        i, addr->sa_data[i]);
1525         }
1526
1527         eth_commit_mac_addr_change(dev, p);
1528         ret = 0;
1529
1530 out:
1531         kfree(addr);
1532         return ret;
1533 }
1534
1535 static void virtnet_stats(struct net_device *dev,
1536                           struct rtnl_link_stats64 *tot)
1537 {
1538         struct virtnet_info *vi = netdev_priv(dev);
1539         unsigned int start;
1540         int i;
1541
1542         for (i = 0; i < vi->max_queue_pairs; i++) {
1543                 u64 tpackets, tbytes, rpackets, rbytes;
1544                 struct receive_queue *rq = &vi->rq[i];
1545                 struct send_queue *sq = &vi->sq[i];
1546
1547                 do {
1548                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1549                         tpackets = sq->stats.packets;
1550                         tbytes   = sq->stats.bytes;
1551                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1552
1553                 do {
1554                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1555                         rpackets = rq->stats.packets;
1556                         rbytes   = rq->stats.bytes;
1557                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1558
1559                 tot->rx_packets += rpackets;
1560                 tot->tx_packets += tpackets;
1561                 tot->rx_bytes   += rbytes;
1562                 tot->tx_bytes   += tbytes;
1563         }
1564
1565         tot->tx_dropped = dev->stats.tx_dropped;
1566         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1567         tot->rx_dropped = dev->stats.rx_dropped;
1568         tot->rx_length_errors = dev->stats.rx_length_errors;
1569         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1570 }
1571
1572 #ifdef CONFIG_NET_POLL_CONTROLLER
1573 static void virtnet_netpoll(struct net_device *dev)
1574 {
1575         struct virtnet_info *vi = netdev_priv(dev);
1576         int i;
1577
1578         for (i = 0; i < vi->curr_queue_pairs; i++)
1579                 napi_schedule(&vi->rq[i].napi);
1580 }
1581 #endif
1582
1583 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1584 {
1585         rtnl_lock();
1586         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1587                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1588                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1589         rtnl_unlock();
1590 }
1591
1592 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1593 {
1594         struct scatterlist sg;
1595         struct net_device *dev = vi->dev;
1596
1597         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1598                 return 0;
1599
1600         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1601         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1602
1603         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1604                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1605                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1606                          queue_pairs);
1607                 return -EINVAL;
1608         } else {
1609                 vi->curr_queue_pairs = queue_pairs;
1610                 /* virtnet_open() will refill when device is going to up. */
1611                 if (dev->flags & IFF_UP)
1612                         schedule_delayed_work(&vi->refill, 0);
1613         }
1614
1615         return 0;
1616 }
1617
1618 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1619 {
1620         int err;
1621
1622         rtnl_lock();
1623         err = _virtnet_set_queues(vi, queue_pairs);
1624         rtnl_unlock();
1625         return err;
1626 }
1627
1628 static int virtnet_close(struct net_device *dev)
1629 {
1630         struct virtnet_info *vi = netdev_priv(dev);
1631         int i;
1632
1633         /* Make sure refill_work doesn't re-enable napi! */
1634         cancel_delayed_work_sync(&vi->refill);
1635
1636         for (i = 0; i < vi->max_queue_pairs; i++) {
1637                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1638                 napi_disable(&vi->rq[i].napi);
1639                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1640         }
1641
1642         return 0;
1643 }
1644
1645 static void virtnet_set_rx_mode(struct net_device *dev)
1646 {
1647         struct virtnet_info *vi = netdev_priv(dev);
1648         struct scatterlist sg[2];
1649         struct virtio_net_ctrl_mac *mac_data;
1650         struct netdev_hw_addr *ha;
1651         int uc_count;
1652         int mc_count;
1653         void *buf;
1654         int i;
1655
1656         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1657         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1658                 return;
1659
1660         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1661         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1662
1663         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1664
1665         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1666                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1667                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1668                          vi->ctrl_promisc ? "en" : "dis");
1669
1670         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1671
1672         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1673                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1674                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1675                          vi->ctrl_allmulti ? "en" : "dis");
1676
1677         uc_count = netdev_uc_count(dev);
1678         mc_count = netdev_mc_count(dev);
1679         /* MAC filter - use one buffer for both lists */
1680         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1681                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1682         mac_data = buf;
1683         if (!buf)
1684                 return;
1685
1686         sg_init_table(sg, 2);
1687
1688         /* Store the unicast list and count in the front of the buffer */
1689         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1690         i = 0;
1691         netdev_for_each_uc_addr(ha, dev)
1692                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1693
1694         sg_set_buf(&sg[0], mac_data,
1695                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1696
1697         /* multicast list and count fill the end */
1698         mac_data = (void *)&mac_data->macs[uc_count][0];
1699
1700         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1701         i = 0;
1702         netdev_for_each_mc_addr(ha, dev)
1703                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1704
1705         sg_set_buf(&sg[1], mac_data,
1706                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1707
1708         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1709                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1710                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1711
1712         kfree(buf);
1713 }
1714
1715 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1716                                    __be16 proto, u16 vid)
1717 {
1718         struct virtnet_info *vi = netdev_priv(dev);
1719         struct scatterlist sg;
1720
1721         vi->ctrl_vid = vid;
1722         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1723
1724         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1725                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1726                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1727         return 0;
1728 }
1729
1730 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1731                                     __be16 proto, u16 vid)
1732 {
1733         struct virtnet_info *vi = netdev_priv(dev);
1734         struct scatterlist sg;
1735
1736         vi->ctrl_vid = vid;
1737         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1738
1739         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1740                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1741                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1742         return 0;
1743 }
1744
1745 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1746 {
1747         int i;
1748
1749         if (vi->affinity_hint_set) {
1750                 for (i = 0; i < vi->max_queue_pairs; i++) {
1751                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1752                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1753                 }
1754
1755                 vi->affinity_hint_set = false;
1756         }
1757 }
1758
1759 static void virtnet_set_affinity(struct virtnet_info *vi)
1760 {
1761         int i;
1762         int cpu;
1763
1764         /* In multiqueue mode, when the number of cpu is equal to the number of
1765          * queue pairs, we let the queue pairs to be private to one cpu by
1766          * setting the affinity hint to eliminate the contention.
1767          */
1768         if (vi->curr_queue_pairs == 1 ||
1769             vi->max_queue_pairs != num_online_cpus()) {
1770                 virtnet_clean_affinity(vi, -1);
1771                 return;
1772         }
1773
1774         i = 0;
1775         for_each_online_cpu(cpu) {
1776                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1777                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1778                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1779                 i++;
1780         }
1781
1782         vi->affinity_hint_set = true;
1783 }
1784
1785 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1786 {
1787         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1788                                                    node);
1789         virtnet_set_affinity(vi);
1790         return 0;
1791 }
1792
1793 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1794 {
1795         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1796                                                    node_dead);
1797         virtnet_set_affinity(vi);
1798         return 0;
1799 }
1800
1801 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1802 {
1803         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1804                                                    node);
1805
1806         virtnet_clean_affinity(vi, cpu);
1807         return 0;
1808 }
1809
1810 static enum cpuhp_state virtionet_online;
1811
1812 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1813 {
1814         int ret;
1815
1816         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1817         if (ret)
1818                 return ret;
1819         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1820                                                &vi->node_dead);
1821         if (!ret)
1822                 return ret;
1823         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1824         return ret;
1825 }
1826
1827 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1828 {
1829         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1830         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1831                                             &vi->node_dead);
1832 }
1833
1834 static void virtnet_get_ringparam(struct net_device *dev,
1835                                 struct ethtool_ringparam *ring)
1836 {
1837         struct virtnet_info *vi = netdev_priv(dev);
1838
1839         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1840         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1841         ring->rx_pending = ring->rx_max_pending;
1842         ring->tx_pending = ring->tx_max_pending;
1843 }
1844
1845
1846 static void virtnet_get_drvinfo(struct net_device *dev,
1847                                 struct ethtool_drvinfo *info)
1848 {
1849         struct virtnet_info *vi = netdev_priv(dev);
1850         struct virtio_device *vdev = vi->vdev;
1851
1852         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1853         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1854         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1855
1856 }
1857
1858 /* TODO: Eliminate OOO packets during switching */
1859 static int virtnet_set_channels(struct net_device *dev,
1860                                 struct ethtool_channels *channels)
1861 {
1862         struct virtnet_info *vi = netdev_priv(dev);
1863         u16 queue_pairs = channels->combined_count;
1864         int err;
1865
1866         /* We don't support separate rx/tx channels.
1867          * We don't allow setting 'other' channels.
1868          */
1869         if (channels->rx_count || channels->tx_count || channels->other_count)
1870                 return -EINVAL;
1871
1872         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1873                 return -EINVAL;
1874
1875         /* For now we don't support modifying channels while XDP is loaded
1876          * also when XDP is loaded all RX queues have XDP programs so we only
1877          * need to check a single RX queue.
1878          */
1879         if (vi->rq[0].xdp_prog)
1880                 return -EINVAL;
1881
1882         get_online_cpus();
1883         err = _virtnet_set_queues(vi, queue_pairs);
1884         if (!err) {
1885                 netif_set_real_num_tx_queues(dev, queue_pairs);
1886                 netif_set_real_num_rx_queues(dev, queue_pairs);
1887
1888                 virtnet_set_affinity(vi);
1889         }
1890         put_online_cpus();
1891
1892         return err;
1893 }
1894
1895 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
1896 {
1897         struct virtnet_info *vi = netdev_priv(dev);
1898         char *p = (char *)data;
1899         unsigned int i, j;
1900
1901         switch (stringset) {
1902         case ETH_SS_STATS:
1903                 for (i = 0; i < vi->curr_queue_pairs; i++) {
1904                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1905                                 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
1906                                          i, virtnet_rq_stats_desc[j].desc);
1907                                 p += ETH_GSTRING_LEN;
1908                         }
1909                 }
1910
1911                 for (i = 0; i < vi->curr_queue_pairs; i++) {
1912                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1913                                 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
1914                                          i, virtnet_sq_stats_desc[j].desc);
1915                                 p += ETH_GSTRING_LEN;
1916                         }
1917                 }
1918                 break;
1919         }
1920 }
1921
1922 static int virtnet_get_sset_count(struct net_device *dev, int sset)
1923 {
1924         struct virtnet_info *vi = netdev_priv(dev);
1925
1926         switch (sset) {
1927         case ETH_SS_STATS:
1928                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
1929                                                VIRTNET_SQ_STATS_LEN);
1930         default:
1931                 return -EOPNOTSUPP;
1932         }
1933 }
1934
1935 static void virtnet_get_ethtool_stats(struct net_device *dev,
1936                                       struct ethtool_stats *stats, u64 *data)
1937 {
1938         struct virtnet_info *vi = netdev_priv(dev);
1939         unsigned int idx = 0, start, i, j;
1940         const u8 *stats_base;
1941         size_t offset;
1942
1943         for (i = 0; i < vi->curr_queue_pairs; i++) {
1944                 struct receive_queue *rq = &vi->rq[i];
1945
1946                 stats_base = (u8 *)&rq->stats;
1947                 do {
1948                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1949                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1950                                 offset = virtnet_rq_stats_desc[j].offset;
1951                                 data[idx + j] = *(u64 *)(stats_base + offset);
1952                         }
1953                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1954                 idx += VIRTNET_RQ_STATS_LEN;
1955         }
1956
1957         for (i = 0; i < vi->curr_queue_pairs; i++) {
1958                 struct send_queue *sq = &vi->sq[i];
1959
1960                 stats_base = (u8 *)&sq->stats;
1961                 do {
1962                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1963                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1964                                 offset = virtnet_sq_stats_desc[j].offset;
1965                                 data[idx + j] = *(u64 *)(stats_base + offset);
1966                         }
1967                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1968                 idx += VIRTNET_SQ_STATS_LEN;
1969         }
1970 }
1971
1972 static void virtnet_get_channels(struct net_device *dev,
1973                                  struct ethtool_channels *channels)
1974 {
1975         struct virtnet_info *vi = netdev_priv(dev);
1976
1977         channels->combined_count = vi->curr_queue_pairs;
1978         channels->max_combined = vi->max_queue_pairs;
1979         channels->max_other = 0;
1980         channels->rx_count = 0;
1981         channels->tx_count = 0;
1982         channels->other_count = 0;
1983 }
1984
1985 /* Check if the user is trying to change anything besides speed/duplex */
1986 static bool
1987 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1988 {
1989         struct ethtool_link_ksettings diff1 = *cmd;
1990         struct ethtool_link_ksettings diff2 = {};
1991
1992         /* cmd is always set so we need to clear it, validate the port type
1993          * and also without autonegotiation we can ignore advertising
1994          */
1995         diff1.base.speed = 0;
1996         diff2.base.port = PORT_OTHER;
1997         ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
1998         diff1.base.duplex = 0;
1999         diff1.base.cmd = 0;
2000         diff1.base.link_mode_masks_nwords = 0;
2001
2002         return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2003                 bitmap_empty(diff1.link_modes.supported,
2004                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2005                 bitmap_empty(diff1.link_modes.advertising,
2006                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2007                 bitmap_empty(diff1.link_modes.lp_advertising,
2008                              __ETHTOOL_LINK_MODE_MASK_NBITS);
2009 }
2010
2011 static int virtnet_set_link_ksettings(struct net_device *dev,
2012                                       const struct ethtool_link_ksettings *cmd)
2013 {
2014         struct virtnet_info *vi = netdev_priv(dev);
2015         u32 speed;
2016
2017         speed = cmd->base.speed;
2018         /* don't allow custom speed and duplex */
2019         if (!ethtool_validate_speed(speed) ||
2020             !ethtool_validate_duplex(cmd->base.duplex) ||
2021             !virtnet_validate_ethtool_cmd(cmd))
2022                 return -EINVAL;
2023         vi->speed = speed;
2024         vi->duplex = cmd->base.duplex;
2025
2026         return 0;
2027 }
2028
2029 static int virtnet_get_link_ksettings(struct net_device *dev,
2030                                       struct ethtool_link_ksettings *cmd)
2031 {
2032         struct virtnet_info *vi = netdev_priv(dev);
2033
2034         cmd->base.speed = vi->speed;
2035         cmd->base.duplex = vi->duplex;
2036         cmd->base.port = PORT_OTHER;
2037
2038         return 0;
2039 }
2040
2041 static void virtnet_init_settings(struct net_device *dev)
2042 {
2043         struct virtnet_info *vi = netdev_priv(dev);
2044
2045         vi->speed = SPEED_UNKNOWN;
2046         vi->duplex = DUPLEX_UNKNOWN;
2047 }
2048
2049 static void virtnet_update_settings(struct virtnet_info *vi)
2050 {
2051         u32 speed;
2052         u8 duplex;
2053
2054         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2055                 return;
2056
2057         speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2058                                                   speed));
2059         if (ethtool_validate_speed(speed))
2060                 vi->speed = speed;
2061         duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2062                                                   duplex));
2063         if (ethtool_validate_duplex(duplex))
2064                 vi->duplex = duplex;
2065 }
2066
2067 static const struct ethtool_ops virtnet_ethtool_ops = {
2068         .get_drvinfo = virtnet_get_drvinfo,
2069         .get_link = ethtool_op_get_link,
2070         .get_ringparam = virtnet_get_ringparam,
2071         .get_strings = virtnet_get_strings,
2072         .get_sset_count = virtnet_get_sset_count,
2073         .get_ethtool_stats = virtnet_get_ethtool_stats,
2074         .set_channels = virtnet_set_channels,
2075         .get_channels = virtnet_get_channels,
2076         .get_ts_info = ethtool_op_get_ts_info,
2077         .get_link_ksettings = virtnet_get_link_ksettings,
2078         .set_link_ksettings = virtnet_set_link_ksettings,
2079 };
2080
2081 static void virtnet_freeze_down(struct virtio_device *vdev)
2082 {
2083         struct virtnet_info *vi = vdev->priv;
2084         int i;
2085
2086         /* Make sure no work handler is accessing the device */
2087         flush_work(&vi->config_work);
2088
2089         netif_device_detach(vi->dev);
2090         netif_tx_disable(vi->dev);
2091         cancel_delayed_work_sync(&vi->refill);
2092
2093         if (netif_running(vi->dev)) {
2094                 for (i = 0; i < vi->max_queue_pairs; i++) {
2095                         napi_disable(&vi->rq[i].napi);
2096                         virtnet_napi_tx_disable(&vi->sq[i].napi);
2097                 }
2098         }
2099 }
2100
2101 static int init_vqs(struct virtnet_info *vi);
2102
2103 static int virtnet_restore_up(struct virtio_device *vdev)
2104 {
2105         struct virtnet_info *vi = vdev->priv;
2106         int err, i;
2107
2108         err = init_vqs(vi);
2109         if (err)
2110                 return err;
2111
2112         virtio_device_ready(vdev);
2113
2114         if (netif_running(vi->dev)) {
2115                 for (i = 0; i < vi->curr_queue_pairs; i++)
2116                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2117                                 schedule_delayed_work(&vi->refill, 0);
2118
2119                 for (i = 0; i < vi->max_queue_pairs; i++) {
2120                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2121                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2122                                                &vi->sq[i].napi);
2123                 }
2124         }
2125
2126         netif_device_attach(vi->dev);
2127         return err;
2128 }
2129
2130 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2131 {
2132         struct scatterlist sg;
2133         vi->ctrl_offloads = cpu_to_virtio64(vi->vdev, offloads);
2134
2135         sg_init_one(&sg, &vi->ctrl_offloads, sizeof(vi->ctrl_offloads));
2136
2137         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2138                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2139                 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2140                 return -EINVAL;
2141         }
2142
2143         return 0;
2144 }
2145
2146 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2147 {
2148         u64 offloads = 0;
2149
2150         if (!vi->guest_offloads)
2151                 return 0;
2152
2153         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2154                 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2155
2156         return virtnet_set_guest_offloads(vi, offloads);
2157 }
2158
2159 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2160 {
2161         u64 offloads = vi->guest_offloads;
2162
2163         if (!vi->guest_offloads)
2164                 return 0;
2165         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2166                 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2167
2168         return virtnet_set_guest_offloads(vi, offloads);
2169 }
2170
2171 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2172                            struct netlink_ext_ack *extack)
2173 {
2174         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2175         struct virtnet_info *vi = netdev_priv(dev);
2176         struct bpf_prog *old_prog;
2177         u16 xdp_qp = 0, curr_qp;
2178         int i, err;
2179
2180         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2181             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2182                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2183                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2184                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
2185                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
2186                 return -EOPNOTSUPP;
2187         }
2188
2189         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2190                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2191                 return -EINVAL;
2192         }
2193
2194         if (dev->mtu > max_sz) {
2195                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2196                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2197                 return -EINVAL;
2198         }
2199
2200         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2201         if (prog)
2202                 xdp_qp = nr_cpu_ids;
2203
2204         /* XDP requires extra queues for XDP_TX */
2205         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2206                 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2207                 netdev_warn(dev, "request %i queues but max is %i\n",
2208                             curr_qp + xdp_qp, vi->max_queue_pairs);
2209                 return -ENOMEM;
2210         }
2211
2212         if (prog) {
2213                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2214                 if (IS_ERR(prog))
2215                         return PTR_ERR(prog);
2216         }
2217
2218         /* Make sure NAPI is not using any XDP TX queues for RX. */
2219         if (netif_running(dev))
2220                 for (i = 0; i < vi->max_queue_pairs; i++)
2221                         napi_disable(&vi->rq[i].napi);
2222
2223         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2224         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2225         if (err)
2226                 goto err;
2227         vi->xdp_queue_pairs = xdp_qp;
2228
2229         for (i = 0; i < vi->max_queue_pairs; i++) {
2230                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2231                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2232                 if (i == 0) {
2233                         if (!old_prog)
2234                                 virtnet_clear_guest_offloads(vi);
2235                         if (!prog)
2236                                 virtnet_restore_guest_offloads(vi);
2237                 }
2238                 if (old_prog)
2239                         bpf_prog_put(old_prog);
2240                 if (netif_running(dev))
2241                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2242         }
2243
2244         return 0;
2245
2246 err:
2247         for (i = 0; i < vi->max_queue_pairs; i++)
2248                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2249         if (prog)
2250                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2251         return err;
2252 }
2253
2254 static u32 virtnet_xdp_query(struct net_device *dev)
2255 {
2256         struct virtnet_info *vi = netdev_priv(dev);
2257         const struct bpf_prog *xdp_prog;
2258         int i;
2259
2260         for (i = 0; i < vi->max_queue_pairs; i++) {
2261                 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2262                 if (xdp_prog)
2263                         return xdp_prog->aux->id;
2264         }
2265         return 0;
2266 }
2267
2268 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2269 {
2270         switch (xdp->command) {
2271         case XDP_SETUP_PROG:
2272                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2273         case XDP_QUERY_PROG:
2274                 xdp->prog_id = virtnet_xdp_query(dev);
2275                 xdp->prog_attached = !!xdp->prog_id;
2276                 return 0;
2277         default:
2278                 return -EINVAL;
2279         }
2280 }
2281
2282 static const struct net_device_ops virtnet_netdev = {
2283         .ndo_open            = virtnet_open,
2284         .ndo_stop            = virtnet_close,
2285         .ndo_start_xmit      = start_xmit,
2286         .ndo_validate_addr   = eth_validate_addr,
2287         .ndo_set_mac_address = virtnet_set_mac_address,
2288         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2289         .ndo_get_stats64     = virtnet_stats,
2290         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2291         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2292 #ifdef CONFIG_NET_POLL_CONTROLLER
2293         .ndo_poll_controller = virtnet_netpoll,
2294 #endif
2295         .ndo_bpf                = virtnet_xdp,
2296         .ndo_xdp_xmit           = virtnet_xdp_xmit,
2297         .ndo_xdp_flush          = virtnet_xdp_flush,
2298         .ndo_features_check     = passthru_features_check,
2299 };
2300
2301 static void virtnet_config_changed_work(struct work_struct *work)
2302 {
2303         struct virtnet_info *vi =
2304                 container_of(work, struct virtnet_info, config_work);
2305         u16 v;
2306
2307         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2308                                  struct virtio_net_config, status, &v) < 0)
2309                 return;
2310
2311         if (v & VIRTIO_NET_S_ANNOUNCE) {
2312                 netdev_notify_peers(vi->dev);
2313                 virtnet_ack_link_announce(vi);
2314         }
2315
2316         /* Ignore unknown (future) status bits */
2317         v &= VIRTIO_NET_S_LINK_UP;
2318
2319         if (vi->status == v)
2320                 return;
2321
2322         vi->status = v;
2323
2324         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2325                 virtnet_update_settings(vi);
2326                 netif_carrier_on(vi->dev);
2327                 netif_tx_wake_all_queues(vi->dev);
2328         } else {
2329                 netif_carrier_off(vi->dev);
2330                 netif_tx_stop_all_queues(vi->dev);
2331         }
2332 }
2333
2334 static void virtnet_config_changed(struct virtio_device *vdev)
2335 {
2336         struct virtnet_info *vi = vdev->priv;
2337
2338         schedule_work(&vi->config_work);
2339 }
2340
2341 static void virtnet_free_queues(struct virtnet_info *vi)
2342 {
2343         int i;
2344
2345         for (i = 0; i < vi->max_queue_pairs; i++) {
2346                 napi_hash_del(&vi->rq[i].napi);
2347                 netif_napi_del(&vi->rq[i].napi);
2348                 netif_napi_del(&vi->sq[i].napi);
2349         }
2350
2351         /* We called napi_hash_del() before netif_napi_del(),
2352          * we need to respect an RCU grace period before freeing vi->rq
2353          */
2354         synchronize_net();
2355
2356         kfree(vi->rq);
2357         kfree(vi->sq);
2358 }
2359
2360 static void _free_receive_bufs(struct virtnet_info *vi)
2361 {
2362         struct bpf_prog *old_prog;
2363         int i;
2364
2365         for (i = 0; i < vi->max_queue_pairs; i++) {
2366                 while (vi->rq[i].pages)
2367                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2368
2369                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2370                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2371                 if (old_prog)
2372                         bpf_prog_put(old_prog);
2373         }
2374 }
2375
2376 static void free_receive_bufs(struct virtnet_info *vi)
2377 {
2378         rtnl_lock();
2379         _free_receive_bufs(vi);
2380         rtnl_unlock();
2381 }
2382
2383 static void free_receive_page_frags(struct virtnet_info *vi)
2384 {
2385         int i;
2386         for (i = 0; i < vi->max_queue_pairs; i++)
2387                 if (vi->rq[i].alloc_frag.page)
2388                         put_page(vi->rq[i].alloc_frag.page);
2389 }
2390
2391 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2392 {
2393         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2394                 return false;
2395         else if (q < vi->curr_queue_pairs)
2396                 return true;
2397         else
2398                 return false;
2399 }
2400
2401 static void free_unused_bufs(struct virtnet_info *vi)
2402 {
2403         void *buf;
2404         int i;
2405
2406         for (i = 0; i < vi->max_queue_pairs; i++) {
2407                 struct virtqueue *vq = vi->sq[i].vq;
2408                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2409                         if (!is_xdp_raw_buffer_queue(vi, i))
2410                                 dev_kfree_skb(buf);
2411                         else
2412                                 put_page(virt_to_head_page(buf));
2413                 }
2414         }
2415
2416         for (i = 0; i < vi->max_queue_pairs; i++) {
2417                 struct virtqueue *vq = vi->rq[i].vq;
2418
2419                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2420                         if (vi->mergeable_rx_bufs) {
2421                                 put_page(virt_to_head_page(buf));
2422                         } else if (vi->big_packets) {
2423                                 give_pages(&vi->rq[i], buf);
2424                         } else {
2425                                 put_page(virt_to_head_page(buf));
2426                         }
2427                 }
2428         }
2429 }
2430
2431 static void virtnet_del_vqs(struct virtnet_info *vi)
2432 {
2433         struct virtio_device *vdev = vi->vdev;
2434
2435         virtnet_clean_affinity(vi, -1);
2436
2437         vdev->config->del_vqs(vdev);
2438
2439         virtnet_free_queues(vi);
2440 }
2441
2442 /* How large should a single buffer be so a queue full of these can fit at
2443  * least one full packet?
2444  * Logic below assumes the mergeable buffer header is used.
2445  */
2446 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2447 {
2448         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2449         unsigned int rq_size = virtqueue_get_vring_size(vq);
2450         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2451         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2452         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2453
2454         return max(max(min_buf_len, hdr_len) - hdr_len,
2455                    (unsigned int)GOOD_PACKET_LEN);
2456 }
2457
2458 static int virtnet_find_vqs(struct virtnet_info *vi)
2459 {
2460         vq_callback_t **callbacks;
2461         struct virtqueue **vqs;
2462         int ret = -ENOMEM;
2463         int i, total_vqs;
2464         const char **names;
2465         bool *ctx;
2466
2467         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2468          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2469          * possible control vq.
2470          */
2471         total_vqs = vi->max_queue_pairs * 2 +
2472                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2473
2474         /* Allocate space for find_vqs parameters */
2475         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2476         if (!vqs)
2477                 goto err_vq;
2478         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2479         if (!callbacks)
2480                 goto err_callback;
2481         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2482         if (!names)
2483                 goto err_names;
2484         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2485                 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2486                 if (!ctx)
2487                         goto err_ctx;
2488         } else {
2489                 ctx = NULL;
2490         }
2491
2492         /* Parameters for control virtqueue, if any */
2493         if (vi->has_cvq) {
2494                 callbacks[total_vqs - 1] = NULL;
2495                 names[total_vqs - 1] = "control";
2496         }
2497
2498         /* Allocate/initialize parameters for send/receive virtqueues */
2499         for (i = 0; i < vi->max_queue_pairs; i++) {
2500                 callbacks[rxq2vq(i)] = skb_recv_done;
2501                 callbacks[txq2vq(i)] = skb_xmit_done;
2502                 sprintf(vi->rq[i].name, "input.%d", i);
2503                 sprintf(vi->sq[i].name, "output.%d", i);
2504                 names[rxq2vq(i)] = vi->rq[i].name;
2505                 names[txq2vq(i)] = vi->sq[i].name;
2506                 if (ctx)
2507                         ctx[rxq2vq(i)] = true;
2508         }
2509
2510         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2511                                          names, ctx, NULL);
2512         if (ret)
2513                 goto err_find;
2514
2515         if (vi->has_cvq) {
2516                 vi->cvq = vqs[total_vqs - 1];
2517                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2518                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2519         }
2520
2521         for (i = 0; i < vi->max_queue_pairs; i++) {
2522                 vi->rq[i].vq = vqs[rxq2vq(i)];
2523                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2524                 vi->sq[i].vq = vqs[txq2vq(i)];
2525         }
2526
2527         kfree(names);
2528         kfree(callbacks);
2529         kfree(vqs);
2530         kfree(ctx);
2531
2532         return 0;
2533
2534 err_find:
2535         kfree(ctx);
2536 err_ctx:
2537         kfree(names);
2538 err_names:
2539         kfree(callbacks);
2540 err_callback:
2541         kfree(vqs);
2542 err_vq:
2543         return ret;
2544 }
2545
2546 static int virtnet_alloc_queues(struct virtnet_info *vi)
2547 {
2548         int i;
2549
2550         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2551         if (!vi->sq)
2552                 goto err_sq;
2553         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2554         if (!vi->rq)
2555                 goto err_rq;
2556
2557         INIT_DELAYED_WORK(&vi->refill, refill_work);
2558         for (i = 0; i < vi->max_queue_pairs; i++) {
2559                 vi->rq[i].pages = NULL;
2560                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2561                                napi_weight);
2562                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2563                                   napi_tx ? napi_weight : 0);
2564
2565                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2566                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2567                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2568
2569                 u64_stats_init(&vi->rq[i].stats.syncp);
2570                 u64_stats_init(&vi->sq[i].stats.syncp);
2571         }
2572
2573         return 0;
2574
2575 err_rq:
2576         kfree(vi->sq);
2577 err_sq:
2578         return -ENOMEM;
2579 }
2580
2581 static int init_vqs(struct virtnet_info *vi)
2582 {
2583         int ret;
2584
2585         /* Allocate send & receive queues */
2586         ret = virtnet_alloc_queues(vi);
2587         if (ret)
2588                 goto err;
2589
2590         ret = virtnet_find_vqs(vi);
2591         if (ret)
2592                 goto err_free;
2593
2594         get_online_cpus();
2595         virtnet_set_affinity(vi);
2596         put_online_cpus();
2597
2598         return 0;
2599
2600 err_free:
2601         virtnet_free_queues(vi);
2602 err:
2603         return ret;
2604 }
2605
2606 #ifdef CONFIG_SYSFS
2607 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2608                 char *buf)
2609 {
2610         struct virtnet_info *vi = netdev_priv(queue->dev);
2611         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2612         unsigned int headroom = virtnet_get_headroom(vi);
2613         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2614         struct ewma_pkt_len *avg;
2615
2616         BUG_ON(queue_index >= vi->max_queue_pairs);
2617         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2618         return sprintf(buf, "%u\n",
2619                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
2620                                        SKB_DATA_ALIGN(headroom + tailroom)));
2621 }
2622
2623 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2624         __ATTR_RO(mergeable_rx_buffer_size);
2625
2626 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2627         &mergeable_rx_buffer_size_attribute.attr,
2628         NULL
2629 };
2630
2631 static const struct attribute_group virtio_net_mrg_rx_group = {
2632         .name = "virtio_net",
2633         .attrs = virtio_net_mrg_rx_attrs
2634 };
2635 #endif
2636
2637 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2638                                     unsigned int fbit,
2639                                     const char *fname, const char *dname)
2640 {
2641         if (!virtio_has_feature(vdev, fbit))
2642                 return false;
2643
2644         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2645                 fname, dname);
2646
2647         return true;
2648 }
2649
2650 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2651         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2652
2653 static bool virtnet_validate_features(struct virtio_device *vdev)
2654 {
2655         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2656             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2657                              "VIRTIO_NET_F_CTRL_VQ") ||
2658              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2659                              "VIRTIO_NET_F_CTRL_VQ") ||
2660              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2661                              "VIRTIO_NET_F_CTRL_VQ") ||
2662              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2663              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2664                              "VIRTIO_NET_F_CTRL_VQ"))) {
2665                 return false;
2666         }
2667
2668         return true;
2669 }
2670
2671 #define MIN_MTU ETH_MIN_MTU
2672 #define MAX_MTU ETH_MAX_MTU
2673
2674 static int virtnet_validate(struct virtio_device *vdev)
2675 {
2676         if (!vdev->config->get) {
2677                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2678                         __func__);
2679                 return -EINVAL;
2680         }
2681
2682         if (!virtnet_validate_features(vdev))
2683                 return -EINVAL;
2684
2685         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2686                 int mtu = virtio_cread16(vdev,
2687                                          offsetof(struct virtio_net_config,
2688                                                   mtu));
2689                 if (mtu < MIN_MTU)
2690                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2691         }
2692
2693         return 0;
2694 }
2695
2696 static int virtnet_probe(struct virtio_device *vdev)
2697 {
2698         int i, err = -ENOMEM;
2699         struct net_device *dev;
2700         struct virtnet_info *vi;
2701         u16 max_queue_pairs;
2702         int mtu;
2703
2704         /* Find if host supports multiqueue virtio_net device */
2705         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2706                                    struct virtio_net_config,
2707                                    max_virtqueue_pairs, &max_queue_pairs);
2708
2709         /* We need at least 2 queue's */
2710         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2711             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2712             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2713                 max_queue_pairs = 1;
2714
2715         /* Allocate ourselves a network device with room for our info */
2716         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2717         if (!dev)
2718                 return -ENOMEM;
2719
2720         /* Set up network device as normal. */
2721         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2722         dev->netdev_ops = &virtnet_netdev;
2723         dev->features = NETIF_F_HIGHDMA;
2724
2725         dev->ethtool_ops = &virtnet_ethtool_ops;
2726         SET_NETDEV_DEV(dev, &vdev->dev);
2727
2728         /* Do we support "hardware" checksums? */
2729         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2730                 /* This opens up the world of extra features. */
2731                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2732                 if (csum)
2733                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2734
2735                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2736                         dev->hw_features |= NETIF_F_TSO
2737                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2738                 }
2739                 /* Individual feature bits: what can host handle? */
2740                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2741                         dev->hw_features |= NETIF_F_TSO;
2742                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2743                         dev->hw_features |= NETIF_F_TSO6;
2744                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2745                         dev->hw_features |= NETIF_F_TSO_ECN;
2746
2747                 dev->features |= NETIF_F_GSO_ROBUST;
2748
2749                 if (gso)
2750                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2751                 /* (!csum && gso) case will be fixed by register_netdev() */
2752         }
2753         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2754                 dev->features |= NETIF_F_RXCSUM;
2755
2756         dev->vlan_features = dev->features;
2757
2758         /* MTU range: 68 - 65535 */
2759         dev->min_mtu = MIN_MTU;
2760         dev->max_mtu = MAX_MTU;
2761
2762         /* Configuration may specify what MAC to use.  Otherwise random. */
2763         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2764                 virtio_cread_bytes(vdev,
2765                                    offsetof(struct virtio_net_config, mac),
2766                                    dev->dev_addr, dev->addr_len);
2767         else
2768                 eth_hw_addr_random(dev);
2769
2770         /* Set up our device-specific information */
2771         vi = netdev_priv(dev);
2772         vi->dev = dev;
2773         vi->vdev = vdev;
2774         vdev->priv = vi;
2775
2776         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2777
2778         /* If we can receive ANY GSO packets, we must allocate large ones. */
2779         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2780             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2781             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2782             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2783                 vi->big_packets = true;
2784
2785         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2786                 vi->mergeable_rx_bufs = true;
2787
2788         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2789             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2790                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2791         else
2792                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2793
2794         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2795             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2796                 vi->any_header_sg = true;
2797
2798         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2799                 vi->has_cvq = true;
2800
2801         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2802                 mtu = virtio_cread16(vdev,
2803                                      offsetof(struct virtio_net_config,
2804                                               mtu));
2805                 if (mtu < dev->min_mtu) {
2806                         /* Should never trigger: MTU was previously validated
2807                          * in virtnet_validate.
2808                          */
2809                         dev_err(&vdev->dev, "device MTU appears to have changed "
2810                                 "it is now %d < %d", mtu, dev->min_mtu);
2811                         goto free;
2812                 }
2813
2814                 dev->mtu = mtu;
2815                 dev->max_mtu = mtu;
2816
2817                 /* TODO: size buffers correctly in this case. */
2818                 if (dev->mtu > ETH_DATA_LEN)
2819                         vi->big_packets = true;
2820         }
2821
2822         if (vi->any_header_sg)
2823                 dev->needed_headroom = vi->hdr_len;
2824
2825         /* Enable multiqueue by default */
2826         if (num_online_cpus() >= max_queue_pairs)
2827                 vi->curr_queue_pairs = max_queue_pairs;
2828         else
2829                 vi->curr_queue_pairs = num_online_cpus();
2830         vi->max_queue_pairs = max_queue_pairs;
2831
2832         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2833         err = init_vqs(vi);
2834         if (err)
2835                 goto free;
2836
2837 #ifdef CONFIG_SYSFS
2838         if (vi->mergeable_rx_bufs)
2839                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2840 #endif
2841         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2842         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2843
2844         virtnet_init_settings(dev);
2845
2846         err = register_netdev(dev);
2847         if (err) {
2848                 pr_debug("virtio_net: registering device failed\n");
2849                 goto free_vqs;
2850         }
2851
2852         virtio_device_ready(vdev);
2853
2854         err = virtnet_cpu_notif_add(vi);
2855         if (err) {
2856                 pr_debug("virtio_net: registering cpu notifier failed\n");
2857                 goto free_unregister_netdev;
2858         }
2859
2860         virtnet_set_queues(vi, vi->curr_queue_pairs);
2861
2862         /* Assume link up if device can't report link status,
2863            otherwise get link status from config. */
2864         netif_carrier_off(dev);
2865         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2866                 schedule_work(&vi->config_work);
2867         } else {
2868                 vi->status = VIRTIO_NET_S_LINK_UP;
2869                 virtnet_update_settings(vi);
2870                 netif_carrier_on(dev);
2871         }
2872
2873         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2874                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2875                         set_bit(guest_offloads[i], &vi->guest_offloads);
2876
2877         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2878                  dev->name, max_queue_pairs);
2879
2880         return 0;
2881
2882 free_unregister_netdev:
2883         vi->vdev->config->reset(vdev);
2884
2885         unregister_netdev(dev);
2886 free_vqs:
2887         cancel_delayed_work_sync(&vi->refill);
2888         free_receive_page_frags(vi);
2889         virtnet_del_vqs(vi);
2890 free:
2891         free_netdev(dev);
2892         return err;
2893 }
2894
2895 static void remove_vq_common(struct virtnet_info *vi)
2896 {
2897         vi->vdev->config->reset(vi->vdev);
2898
2899         /* Free unused buffers in both send and recv, if any. */
2900         free_unused_bufs(vi);
2901
2902         free_receive_bufs(vi);
2903
2904         free_receive_page_frags(vi);
2905
2906         virtnet_del_vqs(vi);
2907 }
2908
2909 static void virtnet_remove(struct virtio_device *vdev)
2910 {
2911         struct virtnet_info *vi = vdev->priv;
2912
2913         virtnet_cpu_notif_remove(vi);
2914
2915         /* Make sure no work handler is accessing the device. */
2916         flush_work(&vi->config_work);
2917
2918         unregister_netdev(vi->dev);
2919
2920         remove_vq_common(vi);
2921
2922         free_netdev(vi->dev);
2923 }
2924
2925 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
2926 {
2927         struct virtnet_info *vi = vdev->priv;
2928
2929         virtnet_cpu_notif_remove(vi);
2930         virtnet_freeze_down(vdev);
2931         remove_vq_common(vi);
2932
2933         return 0;
2934 }
2935
2936 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
2937 {
2938         struct virtnet_info *vi = vdev->priv;
2939         int err;
2940
2941         err = virtnet_restore_up(vdev);
2942         if (err)
2943                 return err;
2944         virtnet_set_queues(vi, vi->curr_queue_pairs);
2945
2946         err = virtnet_cpu_notif_add(vi);
2947         if (err)
2948                 return err;
2949
2950         return 0;
2951 }
2952
2953 static struct virtio_device_id id_table[] = {
2954         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2955         { 0 },
2956 };
2957
2958 #define VIRTNET_FEATURES \
2959         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2960         VIRTIO_NET_F_MAC, \
2961         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2962         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2963         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2964         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2965         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2966         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2967         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2968         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
2969         VIRTIO_NET_F_SPEED_DUPLEX
2970
2971 static unsigned int features[] = {
2972         VIRTNET_FEATURES,
2973 };
2974
2975 static unsigned int features_legacy[] = {
2976         VIRTNET_FEATURES,
2977         VIRTIO_NET_F_GSO,
2978         VIRTIO_F_ANY_LAYOUT,
2979 };
2980
2981 static struct virtio_driver virtio_net_driver = {
2982         .feature_table = features,
2983         .feature_table_size = ARRAY_SIZE(features),
2984         .feature_table_legacy = features_legacy,
2985         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2986         .driver.name =  KBUILD_MODNAME,
2987         .driver.owner = THIS_MODULE,
2988         .id_table =     id_table,
2989         .validate =     virtnet_validate,
2990         .probe =        virtnet_probe,
2991         .remove =       virtnet_remove,
2992         .config_changed = virtnet_config_changed,
2993 #ifdef CONFIG_PM_SLEEP
2994         .freeze =       virtnet_freeze,
2995         .restore =      virtnet_restore,
2996 #endif
2997 };
2998
2999 static __init int virtio_net_driver_init(void)
3000 {
3001         int ret;
3002
3003         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3004                                       virtnet_cpu_online,
3005                                       virtnet_cpu_down_prep);
3006         if (ret < 0)
3007                 goto out;
3008         virtionet_online = ret;
3009         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3010                                       NULL, virtnet_cpu_dead);
3011         if (ret)
3012                 goto err_dead;
3013
3014         ret = register_virtio_driver(&virtio_net_driver);
3015         if (ret)
3016                 goto err_virtio;
3017         return 0;
3018 err_virtio:
3019         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3020 err_dead:
3021         cpuhp_remove_multi_state(virtionet_online);
3022 out:
3023         return ret;
3024 }
3025 module_init(virtio_net_driver_init);
3026
3027 static __exit void virtio_net_driver_exit(void)
3028 {
3029         unregister_virtio_driver(&virtio_net_driver);
3030         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3031         cpuhp_remove_multi_state(virtionet_online);
3032 }
3033 module_exit(virtio_net_driver_exit);
3034
3035 MODULE_DEVICE_TABLE(virtio, id_table);
3036 MODULE_DESCRIPTION("Virtio network driver");
3037 MODULE_LICENSE("GPL");