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