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