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