drivers/net/virtio_net.c

   1 /* A network driver using virtio.
   2  *
   3  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation; either version 2 of the License, or
   8  * (at your option) any later version.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, see <http://www.gnu.org/licenses/>.
  17  */
  18 //#define DEBUG
  19 #include <linux/netdevice.h>
  20 #include <linux/etherdevice.h>
  21 #include <linux/ethtool.h>
  22 #include <linux/module.h>
  23 #include <linux/virtio.h>
  24 #include <linux/virtio_net.h>
  25 #include <linux/bpf.h>
  26 #include <linux/bpf_trace.h>
  27 #include <linux/scatterlist.h>
  28 #include <linux/if_vlan.h>
  29 #include <linux/slab.h>
  30 #include <linux/cpu.h>
  31 #include <linux/average.h>
  32 #include <net/route.h>
  33
  34 static int napi_weight = NAPI_POLL_WEIGHT;
  35 module_param(napi_weight, int, 0444);
  36
  37 static bool csum = true, gso = true, napi_tx;
  38 module_param(csum, bool, 0444);
  39 module_param(gso, bool, 0444);
  40 module_param(napi_tx, bool, 0644);
  41
  42 /* FIXME: MTU in config. */
  43 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  44 #define GOOD_COPY_LEN   128
  45
  46 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  47
  48 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  49 #define VIRTIO_XDP_HEADROOM 256
  50
  51 /* RX packet size EWMA. The average packet size is used to determine the packet
  52  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  53  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  54  * term, transient changes in packet size.
  55  */
  56 DECLARE_EWMA(pkt_len, 0, 64)
  57
  58 #define VIRTNET_DRIVER_VERSION "1.0.0"
  59
  60 struct virtnet_stats {
  61         struct u64_stats_sync tx_syncp;
  62         struct u64_stats_sync rx_syncp;
  63         u64 tx_bytes;
  64         u64 tx_packets;
  65
  66         u64 rx_bytes;
  67         u64 rx_packets;
  68 };
  69
  70 /* Internal representation of a send virtqueue */
  71 struct send_queue {
  72         /* Virtqueue associated with this send _queue */
  73         struct virtqueue *vq;
  74
  75         /* TX: fragments + linear part + virtio header */
  76         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  77
  78         /* Name of the send queue: output.$index */
  79         char name[40];
  80
  81         struct napi_struct napi;
  82 };
  83
  84 /* Internal representation of a receive virtqueue */
  85 struct receive_queue {
  86         /* Virtqueue associated with this receive_queue */
  87         struct virtqueue *vq;
  88
  89         struct napi_struct napi;
  90
  91         struct bpf_prog __rcu *xdp_prog;
  92
  93         /* Chain pages by the private ptr. */
  94         struct page *pages;
  95
  96         /* Average packet length for mergeable receive buffers. */
  97         struct ewma_pkt_len mrg_avg_pkt_len;
  98
  99         /* Page frag for packet buffer allocation. */
 100         struct page_frag alloc_frag;
 101
 102         /* RX: fragments + linear part + virtio header */
 103         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 104
 105         /* Min single buffer size for mergeable buffers case. */
 106         unsigned int min_buf_len;
 107
 108         /* Name of this receive queue: input.$index */
 109         char name[40];
 110 };
 111
 112 struct virtnet_info {
 113         struct virtio_device *vdev;
 114         struct virtqueue *cvq;
 115         struct net_device *dev;
 116         struct send_queue *sq;
 117         struct receive_queue *rq;
 118         unsigned int status;
 119
 120         /* Max # of queue pairs supported by the device */
 121         u16 max_queue_pairs;
 122
 123         /* # of queue pairs currently used by the driver */
 124         u16 curr_queue_pairs;
 125
 126         /* # of XDP queue pairs currently used by the driver */
 127         u16 xdp_queue_pairs;
 128
 129         /* I like... big packets and I cannot lie! */
 130         bool big_packets;
 131
 132         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 133         bool mergeable_rx_bufs;
 134
 135         /* Has control virtqueue */
 136         bool has_cvq;
 137
 138         /* Host can handle any s/g split between our header and packet data */
 139         bool any_header_sg;
 140
 141         /* Packet virtio header size */
 142         u8 hdr_len;
 143
 144         /* Active statistics */
 145         struct virtnet_stats __percpu *stats;
 146
 147         /* Work struct for refilling if we run low on memory. */
 148         struct delayed_work refill;
 149
 150         /* Work struct for config space updates */
 151         struct work_struct config_work;
 152
 153         /* Does the affinity hint is set for virtqueues? */
 154         bool affinity_hint_set;
 155
 156         /* CPU hotplug instances for online & dead */
 157         struct hlist_node node;
 158         struct hlist_node node_dead;
 159
 160         /* Control VQ buffers: protected by the rtnl lock */
 161         struct virtio_net_ctrl_hdr ctrl_hdr;
 162         virtio_net_ctrl_ack ctrl_status;
 163         struct virtio_net_ctrl_mq ctrl_mq;
 164         u8 ctrl_promisc;
 165         u8 ctrl_allmulti;
 166         u16 ctrl_vid;
 167
 168         /* Ethtool settings */
 169         u8 duplex;
 170         u32 speed;
 171 };
 172
 173 struct padded_vnet_hdr {
 174         struct virtio_net_hdr_mrg_rxbuf hdr;
 175         /*
 176          * hdr is in a separate sg buffer, and data sg buffer shares same page
 177          * with this header sg. This padding makes next sg 16 byte aligned
 178          * after the header.
 179          */
 180         char padding[4];
 181 };
 182
 183 /* Converting between virtqueue no. and kernel tx/rx queue no.
 184  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 185  */
 186 static int vq2txq(struct virtqueue *vq)
 187 {
 188         return (vq->index - 1) / 2;
 189 }
 190
 191 static int txq2vq(int txq)
 192 {
 193         return txq * 2 + 1;
 194 }
 195
 196 static int vq2rxq(struct virtqueue *vq)
 197 {
 198         return vq->index / 2;
 199 }
 200
 201 static int rxq2vq(int rxq)
 202 {
 203         return rxq * 2;
 204 }
 205
 206 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 207 {
 208         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 209 }
 210
 211 /*
 212  * private is used to chain pages for big packets, put the whole
 213  * most recent used list in the beginning for reuse
 214  */
 215 static void give_pages(struct receive_queue *rq, struct page *page)
 216 {
 217         struct page *end;
 218
 219         /* Find end of list, sew whole thing into vi->rq.pages. */
 220         for (end = page; end->private; end = (struct page *)end->private);
 221         end->private = (unsigned long)rq->pages;
 222         rq->pages = page;
 223 }
 224
 225 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 226 {
 227         struct page *p = rq->pages;
 228
 229         if (p) {
 230                 rq->pages = (struct page *)p->private;
 231                 /* clear private here, it is used to chain pages */
 232                 p->private = 0;
 233         } else
 234                 p = alloc_page(gfp_mask);
 235         return p;
 236 }
 237
 238 static void virtqueue_napi_schedule(struct napi_struct *napi,
 239                                     struct virtqueue *vq)
 240 {
 241         if (napi_schedule_prep(napi)) {
 242                 virtqueue_disable_cb(vq);
 243                 __napi_schedule(napi);
 244         }
 245 }
 246
 247 static void virtqueue_napi_complete(struct napi_struct *napi,
 248                                     struct virtqueue *vq, int processed)
 249 {
 250         int opaque;
 251
 252         opaque = virtqueue_enable_cb_prepare(vq);
 253         if (napi_complete_done(napi, processed) &&
 254             unlikely(virtqueue_poll(vq, opaque)))
 255                 virtqueue_napi_schedule(napi, vq);
 256 }
 257
 258 static void skb_xmit_done(struct virtqueue *vq)
 259 {
 260         struct virtnet_info *vi = vq->vdev->priv;
 261         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 262
 263         /* Suppress further interrupts. */
 264         virtqueue_disable_cb(vq);
 265
 266         if (napi->weight)
 267                 virtqueue_napi_schedule(napi, vq);
 268         else
 269                 /* We were probably waiting for more output buffers. */
 270                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 271 }
 272
 273 #define MRG_CTX_HEADER_SHIFT 22
 274 static void *mergeable_len_to_ctx(unsigned int truesize,
 275                                   unsigned int headroom)
 276 {
 277         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 278 }
 279
 280 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 281 {
 282         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 283 }
 284
 285 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 286 {
 287         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 288 }
 289
 290 /* Called from bottom half context */
 291 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 292                                    struct receive_queue *rq,
 293                                    struct page *page, unsigned int offset,
 294                                    unsigned int len, unsigned int truesize)
 295 {
 296         struct sk_buff *skb;
 297         struct virtio_net_hdr_mrg_rxbuf *hdr;
 298         unsigned int copy, hdr_len, hdr_padded_len;
 299         char *p;
 300
 301         p = page_address(page) + offset;
 302
 303         /* copy small packet so we can reuse these pages for small data */
 304         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 305         if (unlikely(!skb))
 306                 return NULL;
 307
 308         hdr = skb_vnet_hdr(skb);
 309
 310         hdr_len = vi->hdr_len;
 311         if (vi->mergeable_rx_bufs)
 312                 hdr_padded_len = sizeof *hdr;
 313         else
 314                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 315
 316         memcpy(hdr, p, hdr_len);
 317
 318         len -= hdr_len;
 319         offset += hdr_padded_len;
 320         p += hdr_padded_len;
 321
 322         copy = len;
 323         if (copy > skb_tailroom(skb))
 324                 copy = skb_tailroom(skb);
 325         skb_put_data(skb, p, copy);
 326
 327         len -= copy;
 328         offset += copy;
 329
 330         if (vi->mergeable_rx_bufs) {
 331                 if (len)
 332                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 333                 else
 334                         put_page(page);
 335                 return skb;
 336         }
 337
 338         /*
 339          * Verify that we can indeed put this data into a skb.
 340          * This is here to handle cases when the device erroneously
 341          * tries to receive more than is possible. This is usually
 342          * the case of a broken device.
 343          */
 344         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 345                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 346                 dev_kfree_skb(skb);
 347                 return NULL;
 348         }
 349         BUG_ON(offset >= PAGE_SIZE);
 350         while (len) {
 351                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 352                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 353                                 frag_size, truesize);
 354                 len -= frag_size;
 355                 page = (struct page *)page->private;
 356                 offset = 0;
 357         }
 358
 359         if (page)
 360                 give_pages(rq, page);
 361
 362         return skb;
 363 }
 364
 365 static bool virtnet_xdp_xmit(struct virtnet_info *vi,
 366                              struct receive_queue *rq,
 367                              struct xdp_buff *xdp)
 368 {
 369         struct virtio_net_hdr_mrg_rxbuf *hdr;
 370         unsigned int len;
 371         struct send_queue *sq;
 372         unsigned int qp;
 373         void *xdp_sent;
 374         int err;
 375
 376         qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
 377         sq = &vi->sq[qp];
 378
 379         /* Free up any pending old buffers before queueing new ones. */
 380         while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 381                 struct page *sent_page = virt_to_head_page(xdp_sent);
 382
 383                 put_page(sent_page);
 384         }
 385
 386         xdp->data -= vi->hdr_len;
 387         /* Zero header and leave csum up to XDP layers */
 388         hdr = xdp->data;
 389         memset(hdr, 0, vi->hdr_len);
 390
 391         sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
 392
 393         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp->data, GFP_ATOMIC);
 394         if (unlikely(err)) {
 395                 struct page *page = virt_to_head_page(xdp->data);
 396
 397                 put_page(page);
 398                 return false;
 399         }
 400
 401         virtqueue_kick(sq->vq);
 402         return true;
 403 }
 404
 405 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 406 {
 407         return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
 408 }
 409
 410 static struct sk_buff *receive_small(struct net_device *dev,
 411                                      struct virtnet_info *vi,
 412                                      struct receive_queue *rq,
 413                                      void *buf, void *ctx,
 414                                      unsigned int len)
 415 {
 416         struct sk_buff *skb;
 417         struct bpf_prog *xdp_prog;
 418         unsigned int xdp_headroom = virtnet_get_headroom(vi);
 419         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 420         unsigned int headroom = vi->hdr_len + header_offset;
 421         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 422                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 423         unsigned int delta = 0;
 424         len -= vi->hdr_len;
 425
 426         rcu_read_lock();
 427         xdp_prog = rcu_dereference(rq->xdp_prog);
 428         if (xdp_prog) {
 429                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 430                 struct xdp_buff xdp;
 431                 void *orig_data;
 432                 u32 act;
 433
 434                 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
 435                         goto err_xdp;
 436
 437                 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
 438                 xdp.data = xdp.data_hard_start + xdp_headroom;
 439                 xdp.data_end = xdp.data + len;
 440                 orig_data = xdp.data;
 441                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 442
 443                 switch (act) {
 444                 case XDP_PASS:
 445                         /* Recalculate length in case bpf program changed it */
 446                         delta = orig_data - xdp.data;
 447                         break;
 448                 case XDP_TX:
 449                         if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
 450                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 451                         rcu_read_unlock();
 452                         goto xdp_xmit;
 453                 default:
 454                         bpf_warn_invalid_xdp_action(act);
 455                 case XDP_ABORTED:
 456                         trace_xdp_exception(vi->dev, xdp_prog, act);
 457                 case XDP_DROP:
 458                         goto err_xdp;
 459                 }
 460         }
 461         rcu_read_unlock();
 462
 463         skb = build_skb(buf, buflen);
 464         if (!skb) {
 465                 put_page(virt_to_head_page(buf));
 466                 goto err;
 467         }
 468         skb_reserve(skb, headroom - delta);
 469         skb_put(skb, len + delta);
 470         if (!delta) {
 471                 buf += header_offset;
 472                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 473         } /* keep zeroed vnet hdr since packet was changed by bpf */
 474
 475 err:
 476         return skb;
 477
 478 err_xdp:
 479         rcu_read_unlock();
 480         dev->stats.rx_dropped++;
 481         put_page(virt_to_head_page(buf));
 482 xdp_xmit:
 483         return NULL;
 484 }
 485
 486 static struct sk_buff *receive_big(struct net_device *dev,
 487                                    struct virtnet_info *vi,
 488                                    struct receive_queue *rq,
 489                                    void *buf,
 490                                    unsigned int len)
 491 {
 492         struct page *page = buf;
 493         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 494
 495         if (unlikely(!skb))
 496                 goto err;
 497
 498         return skb;
 499
 500 err:
 501         dev->stats.rx_dropped++;
 502         give_pages(rq, page);
 503         return NULL;
 504 }
 505
 506 /* The conditions to enable XDP should preclude the underlying device from
 507  * sending packets across multiple buffers (num_buf > 1). However per spec
 508  * it does not appear to be illegal to do so but rather just against convention.
 509  * So in order to avoid making a system unresponsive the packets are pushed
 510  * into a page and the XDP program is run. This will be extremely slow and we
 511  * push a warning to the user to fix this as soon as possible. Fixing this may
 512  * require resolving the underlying hardware to determine why multiple buffers
 513  * are being received or simply loading the XDP program in the ingress stack
 514  * after the skb is built because there is no advantage to running it here
 515  * anymore.
 516  */
 517 static struct page *xdp_linearize_page(struct receive_queue *rq,
 518                                        u16 *num_buf,
 519                                        struct page *p,
 520                                        int offset,
 521                                        unsigned int *len)
 522 {
 523         struct page *page = alloc_page(GFP_ATOMIC);
 524         unsigned int page_off = VIRTIO_XDP_HEADROOM;
 525
 526         if (!page)
 527                 return NULL;
 528
 529         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 530         page_off += *len;
 531
 532         while (--*num_buf) {
 533                 unsigned int buflen;
 534                 void *buf;
 535                 int off;
 536
 537                 buf = virtqueue_get_buf(rq->vq, &buflen);
 538                 if (unlikely(!buf))
 539                         goto err_buf;
 540
 541                 p = virt_to_head_page(buf);
 542                 off = buf - page_address(p);
 543
 544                 /* guard against a misconfigured or uncooperative backend that
 545                  * is sending packet larger than the MTU.
 546                  */
 547                 if ((page_off + buflen) > PAGE_SIZE) {
 548                         put_page(p);
 549                         goto err_buf;
 550                 }
 551
 552                 memcpy(page_address(page) + page_off,
 553                        page_address(p) + off, buflen);
 554                 page_off += buflen;
 555                 put_page(p);
 556         }
 557
 558         /* Headroom does not contribute to packet length */
 559         *len = page_off - VIRTIO_XDP_HEADROOM;
 560         return page;
 561 err_buf:
 562         __free_pages(page, 0);
 563         return NULL;
 564 }
 565
 566 static struct sk_buff *receive_mergeable(struct net_device *dev,
 567                                          struct virtnet_info *vi,
 568                                          struct receive_queue *rq,
 569                                          void *buf,
 570                                          void *ctx,
 571                                          unsigned int len)
 572 {
 573         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 574         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 575         struct page *page = virt_to_head_page(buf);
 576         int offset = buf - page_address(page);
 577         struct sk_buff *head_skb, *curr_skb;
 578         struct bpf_prog *xdp_prog;
 579         unsigned int truesize;
 580
 581         head_skb = NULL;
 582
 583         rcu_read_lock();
 584         xdp_prog = rcu_dereference(rq->xdp_prog);
 585         if (xdp_prog) {
 586                 struct page *xdp_page;
 587                 struct xdp_buff xdp;
 588                 void *data;
 589                 u32 act;
 590
 591                 /* This happens when rx buffer size is underestimated */
 592                 if (unlikely(num_buf > 1)) {
 593                         /* linearize data for XDP */
 594                         xdp_page = xdp_linearize_page(rq, &num_buf,
 595                                                       page, offset, &len);
 596                         if (!xdp_page)
 597                                 goto err_xdp;
 598                         offset = VIRTIO_XDP_HEADROOM;
 599                 } else {
 600                         xdp_page = page;
 601                 }
 602
 603                 /* Transient failure which in theory could occur if
 604                  * in-flight packets from before XDP was enabled reach
 605                  * the receive path after XDP is loaded. In practice I
 606                  * was not able to create this condition.
 607                  */
 608                 if (unlikely(hdr->hdr.gso_type))
 609                         goto err_xdp;
 610
 611                 /* Allow consuming headroom but reserve enough space to push
 612                  * the descriptor on if we get an XDP_TX return code.
 613                  */
 614                 data = page_address(xdp_page) + offset;
 615                 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
 616                 xdp.data = data + vi->hdr_len;
 617                 xdp.data_end = xdp.data + (len - vi->hdr_len);
 618                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 619
 620                 switch (act) {
 621                 case XDP_PASS:
 622                         /* recalculate offset to account for any header
 623                          * adjustments. Note other cases do not build an
 624                          * skb and avoid using offset
 625                          */
 626                         offset = xdp.data -
 627                                         page_address(xdp_page) - vi->hdr_len;
 628
 629                         /* We can only create skb based on xdp_page. */
 630                         if (unlikely(xdp_page != page)) {
 631                                 rcu_read_unlock();
 632                                 put_page(page);
 633                                 head_skb = page_to_skb(vi, rq, xdp_page,
 634                                                        offset, len, PAGE_SIZE);
 635                                 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 636                                 return head_skb;
 637                         }
 638                         break;
 639                 case XDP_TX:
 640                         if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
 641                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 642                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 643                         if (unlikely(xdp_page != page))
 644                                 goto err_xdp;
 645                         rcu_read_unlock();
 646                         goto xdp_xmit;
 647                 default:
 648                         bpf_warn_invalid_xdp_action(act);
 649                 case XDP_ABORTED:
 650                         trace_xdp_exception(vi->dev, xdp_prog, act);
 651                 case XDP_DROP:
 652                         if (unlikely(xdp_page != page))
 653                                 __free_pages(xdp_page, 0);
 654                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 655                         goto err_xdp;
 656                 }
 657         }
 658         rcu_read_unlock();
 659
 660         truesize = mergeable_ctx_to_truesize(ctx);
 661         if (unlikely(len > truesize)) {
 662                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 663                          dev->name, len, (unsigned long)ctx);
 664                 dev->stats.rx_length_errors++;
 665                 goto err_skb;
 666         }
 667
 668         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
 669         curr_skb = head_skb;
 670
 671         if (unlikely(!curr_skb))
 672                 goto err_skb;
 673         while (--num_buf) {
 674                 int num_skb_frags;
 675
 676                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
 677                 if (unlikely(!ctx)) {
 678                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 679                                  dev->name, num_buf,
 680                                  virtio16_to_cpu(vi->vdev,
 681                                                  hdr->num_buffers));
 682                         dev->stats.rx_length_errors++;
 683                         goto err_buf;
 684                 }
 685
 686                 page = virt_to_head_page(buf);
 687
 688                 truesize = mergeable_ctx_to_truesize(ctx);
 689                 if (unlikely(len > truesize)) {
 690                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 691                                  dev->name, len, (unsigned long)ctx);
 692                         dev->stats.rx_length_errors++;
 693                         goto err_skb;
 694                 }
 695
 696                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 697                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 698                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 699
 700                         if (unlikely(!nskb))
 701                                 goto err_skb;
 702                         if (curr_skb == head_skb)
 703                                 skb_shinfo(curr_skb)->frag_list = nskb;
 704                         else
 705                                 curr_skb->next = nskb;
 706                         curr_skb = nskb;
 707                         head_skb->truesize += nskb->truesize;
 708                         num_skb_frags = 0;
 709                 }
 710                 if (curr_skb != head_skb) {
 711                         head_skb->data_len += len;
 712                         head_skb->len += len;
 713                         head_skb->truesize += truesize;
 714                 }
 715                 offset = buf - page_address(page);
 716                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 717                         put_page(page);
 718                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 719                                              len, truesize);
 720                 } else {
 721                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 722                                         offset, len, truesize);
 723                 }
 724         }
 725
 726         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 727         return head_skb;
 728
 729 err_xdp:
 730         rcu_read_unlock();
 731 err_skb:
 732         put_page(page);
 733         while (--num_buf) {
 734                 buf = virtqueue_get_buf(rq->vq, &len);
 735                 if (unlikely(!buf)) {
 736                         pr_debug("%s: rx error: %d buffers missing\n",
 737                                  dev->name, num_buf);
 738                         dev->stats.rx_length_errors++;
 739                         break;
 740                 }
 741                 page = virt_to_head_page(buf);
 742                 put_page(page);
 743         }
 744 err_buf:
 745         dev->stats.rx_dropped++;
 746         dev_kfree_skb(head_skb);
 747 xdp_xmit:
 748         return NULL;
 749 }
 750
 751 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 752                        void *buf, unsigned int len, void **ctx)
 753 {
 754         struct net_device *dev = vi->dev;
 755         struct sk_buff *skb;
 756         struct virtio_net_hdr_mrg_rxbuf *hdr;
 757         int ret;
 758
 759         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 760                 pr_debug("%s: short packet %i\n", dev->name, len);
 761                 dev->stats.rx_length_errors++;
 762                 if (vi->mergeable_rx_bufs) {
 763                         put_page(virt_to_head_page(buf));
 764                 } else if (vi->big_packets) {
 765                         give_pages(rq, buf);
 766                 } else {
 767                         put_page(virt_to_head_page(buf));
 768                 }
 769                 return 0;
 770         }
 771
 772         if (vi->mergeable_rx_bufs)
 773                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len);
 774         else if (vi->big_packets)
 775                 skb = receive_big(dev, vi, rq, buf, len);
 776         else
 777                 skb = receive_small(dev, vi, rq, buf, ctx, len);
 778
 779         if (unlikely(!skb))
 780                 return 0;
 781
 782         hdr = skb_vnet_hdr(skb);
 783
 784         ret = skb->len;
 785
 786         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 787                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 788
 789         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 790                                   virtio_is_little_endian(vi->vdev))) {
 791                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 792                                      dev->name, hdr->hdr.gso_type,
 793                                      hdr->hdr.gso_size);
 794                 goto frame_err;
 795         }
 796
 797         skb->protocol = eth_type_trans(skb, dev);
 798         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 799                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 800
 801         napi_gro_receive(&rq->napi, skb);
 802         return ret;
 803
 804 frame_err:
 805         dev->stats.rx_frame_errors++;
 806         dev_kfree_skb(skb);
 807         return 0;
 808 }
 809
 810 /* Unlike mergeable buffers, all buffers are allocated to the
 811  * same size, except for the headroom. For this reason we do
 812  * not need to use  mergeable_len_to_ctx here - it is enough
 813  * to store the headroom as the context ignoring the truesize.
 814  */
 815 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 816                              gfp_t gfp)
 817 {
 818         struct page_frag *alloc_frag = &rq->alloc_frag;
 819         char *buf;
 820         unsigned int xdp_headroom = virtnet_get_headroom(vi);
 821         void *ctx = (void *)(unsigned long)xdp_headroom;
 822         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
 823         int err;
 824
 825         len = SKB_DATA_ALIGN(len) +
 826               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 827         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 828                 return -ENOMEM;
 829
 830         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 831         get_page(alloc_frag->page);
 832         alloc_frag->offset += len;
 833         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
 834                     vi->hdr_len + GOOD_PACKET_LEN);
 835         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
 836         if (err < 0)
 837                 put_page(virt_to_head_page(buf));
 838
 839         return err;
 840 }
 841
 842 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 843                            gfp_t gfp)
 844 {
 845         struct page *first, *list = NULL;
 846         char *p;
 847         int i, err, offset;
 848
 849         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 850
 851         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 852         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 853                 first = get_a_page(rq, gfp);
 854                 if (!first) {
 855                         if (list)
 856                                 give_pages(rq, list);
 857                         return -ENOMEM;
 858                 }
 859                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 860
 861                 /* chain new page in list head to match sg */
 862                 first->private = (unsigned long)list;
 863                 list = first;
 864         }
 865
 866         first = get_a_page(rq, gfp);
 867         if (!first) {
 868                 give_pages(rq, list);
 869                 return -ENOMEM;
 870         }
 871         p = page_address(first);
 872
 873         /* rq->sg[0], rq->sg[1] share the same page */
 874         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 875         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 876
 877         /* rq->sg[1] for data packet, from offset */
 878         offset = sizeof(struct padded_vnet_hdr);
 879         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 880
 881         /* chain first in list head */
 882         first->private = (unsigned long)list;
 883         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 884                                   first, gfp);
 885         if (err < 0)
 886                 give_pages(rq, first);
 887
 888         return err;
 889 }
 890
 891 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
 892                                           struct ewma_pkt_len *avg_pkt_len)
 893 {
 894         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 895         unsigned int len;
 896
 897         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
 898                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
 899         return ALIGN(len, L1_CACHE_BYTES);
 900 }
 901
 902 static int add_recvbuf_mergeable(struct virtnet_info *vi,
 903                                  struct receive_queue *rq, gfp_t gfp)
 904 {
 905         struct page_frag *alloc_frag = &rq->alloc_frag;
 906         unsigned int headroom = virtnet_get_headroom(vi);
 907         char *buf;
 908         void *ctx;
 909         int err;
 910         unsigned int len, hole;
 911
 912         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len);
 913         if (unlikely(!skb_page_frag_refill(len + headroom, alloc_frag, gfp)))
 914                 return -ENOMEM;
 915
 916         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 917         buf += headroom; /* advance address leaving hole at front of pkt */
 918         ctx = mergeable_len_to_ctx(len, headroom);
 919         get_page(alloc_frag->page);
 920         alloc_frag->offset += len + headroom;
 921         hole = alloc_frag->size - alloc_frag->offset;
 922         if (hole < len + headroom) {
 923                 /* To avoid internal fragmentation, if there is very likely not
 924                  * enough space for another buffer, add the remaining space to
 925                  * the current buffer. This extra space is not included in
 926                  * the truesize stored in ctx.
 927                  */
 928                 len += hole;
 929                 alloc_frag->offset += hole;
 930         }
 931
 932         sg_init_one(rq->sg, buf, len);
 933         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
 934         if (err < 0)
 935                 put_page(virt_to_head_page(buf));
 936
 937         return err;
 938 }
 939
 940 /*
 941  * Returns false if we couldn't fill entirely (OOM).
 942  *
 943  * Normally run in the receive path, but can also be run from ndo_open
 944  * before we're receiving packets, or from refill_work which is
 945  * careful to disable receiving (using napi_disable).
 946  */
 947 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 948                           gfp_t gfp)
 949 {
 950         int err;
 951         bool oom;
 952
 953         gfp |= __GFP_COLD;
 954         do {
 955                 if (vi->mergeable_rx_bufs)
 956                         err = add_recvbuf_mergeable(vi, rq, gfp);
 957                 else if (vi->big_packets)
 958                         err = add_recvbuf_big(vi, rq, gfp);
 959                 else
 960                         err = add_recvbuf_small(vi, rq, gfp);
 961
 962                 oom = err == -ENOMEM;
 963                 if (err)
 964                         break;
 965         } while (rq->vq->num_free);
 966         virtqueue_kick(rq->vq);
 967         return !oom;
 968 }
 969
 970 static void skb_recv_done(struct virtqueue *rvq)
 971 {
 972         struct virtnet_info *vi = rvq->vdev->priv;
 973         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 974
 975         virtqueue_napi_schedule(&rq->napi, rvq);
 976 }
 977
 978 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
 979 {
 980         napi_enable(napi);
 981
 982         /* If all buffers were filled by other side before we napi_enabled, we
 983          * won't get another interrupt, so process any outstanding packets now.
 984          * Call local_bh_enable after to trigger softIRQ processing.
 985          */
 986         local_bh_disable();
 987         virtqueue_napi_schedule(napi, vq);
 988         local_bh_enable();
 989 }
 990
 991 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
 992                                    struct virtqueue *vq,
 993                                    struct napi_struct *napi)
 994 {
 995         if (!napi->weight)
 996                 return;
 997
 998         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
 999          * enable the feature if this is likely affine with the transmit path.
1000          */
1001         if (!vi->affinity_hint_set) {
1002                 napi->weight = 0;
1003                 return;
1004         }
1005
1006         return virtnet_napi_enable(vq, napi);
1007 }
1008
1009 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1010 {
1011         if (napi->weight)
1012                 napi_disable(napi);
1013 }
1014
1015 static void refill_work(struct work_struct *work)
1016 {
1017         struct virtnet_info *vi =
1018                 container_of(work, struct virtnet_info, refill.work);
1019         bool still_empty;
1020         int i;
1021
1022         for (i = 0; i < vi->curr_queue_pairs; i++) {
1023                 struct receive_queue *rq = &vi->rq[i];
1024
1025                 napi_disable(&rq->napi);
1026                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1027                 virtnet_napi_enable(rq->vq, &rq->napi);
1028
1029                 /* In theory, this can happen: if we don't get any buffers in
1030                  * we will *never* try to fill again.
1031                  */
1032                 if (still_empty)
1033                         schedule_delayed_work(&vi->refill, HZ/2);
1034         }
1035 }
1036
1037 static int virtnet_receive(struct receive_queue *rq, int budget)
1038 {
1039         struct virtnet_info *vi = rq->vq->vdev->priv;
1040         unsigned int len, received = 0, bytes = 0;
1041         void *buf;
1042         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1043
1044         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1045                 void *ctx;
1046
1047                 while (received < budget &&
1048                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1049                         bytes += receive_buf(vi, rq, buf, len, ctx);
1050                         received++;
1051                 }
1052         } else {
1053                 while (received < budget &&
1054                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1055                         bytes += receive_buf(vi, rq, buf, len, NULL);
1056                         received++;
1057                 }
1058         }
1059
1060         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1061                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1062                         schedule_delayed_work(&vi->refill, 0);
1063         }
1064
1065         u64_stats_update_begin(&stats->rx_syncp);
1066         stats->rx_bytes += bytes;
1067         stats->rx_packets += received;
1068         u64_stats_update_end(&stats->rx_syncp);
1069
1070         return received;
1071 }
1072
1073 static void free_old_xmit_skbs(struct send_queue *sq)
1074 {
1075         struct sk_buff *skb;
1076         unsigned int len;
1077         struct virtnet_info *vi = sq->vq->vdev->priv;
1078         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1079         unsigned int packets = 0;
1080         unsigned int bytes = 0;
1081
1082         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1083                 pr_debug("Sent skb %p\n", skb);
1084
1085                 bytes += skb->len;
1086                 packets++;
1087
1088                 dev_kfree_skb_any(skb);
1089         }
1090
1091         /* Avoid overhead when no packets have been processed
1092          * happens when called speculatively from start_xmit.
1093          */
1094         if (!packets)
1095                 return;
1096
1097         u64_stats_update_begin(&stats->tx_syncp);
1098         stats->tx_bytes += bytes;
1099         stats->tx_packets += packets;
1100         u64_stats_update_end(&stats->tx_syncp);
1101 }
1102
1103 static void virtnet_poll_cleantx(struct receive_queue *rq)
1104 {
1105         struct virtnet_info *vi = rq->vq->vdev->priv;
1106         unsigned int index = vq2rxq(rq->vq);
1107         struct send_queue *sq = &vi->sq[index];
1108         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1109
1110         if (!sq->napi.weight)
1111                 return;
1112
1113         if (__netif_tx_trylock(txq)) {
1114                 free_old_xmit_skbs(sq);
1115                 __netif_tx_unlock(txq);
1116         }
1117
1118         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1119                 netif_tx_wake_queue(txq);
1120 }
1121
1122 static int virtnet_poll(struct napi_struct *napi, int budget)
1123 {
1124         struct receive_queue *rq =
1125                 container_of(napi, struct receive_queue, napi);
1126         unsigned int received;
1127
1128         virtnet_poll_cleantx(rq);
1129
1130         received = virtnet_receive(rq, budget);
1131
1132         /* Out of packets? */
1133         if (received < budget)
1134                 virtqueue_napi_complete(napi, rq->vq, received);
1135
1136         return received;
1137 }
1138
1139 static int virtnet_open(struct net_device *dev)
1140 {
1141         struct virtnet_info *vi = netdev_priv(dev);
1142         int i;
1143
1144         for (i = 0; i < vi->max_queue_pairs; i++) {
1145                 if (i < vi->curr_queue_pairs)
1146                         /* Make sure we have some buffers: if oom use wq. */
1147                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1148                                 schedule_delayed_work(&vi->refill, 0);
1149                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1150                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1151         }
1152
1153         return 0;
1154 }
1155
1156 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1157 {
1158         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1159         struct virtnet_info *vi = sq->vq->vdev->priv;
1160         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1161
1162         __netif_tx_lock(txq, raw_smp_processor_id());
1163         free_old_xmit_skbs(sq);
1164         __netif_tx_unlock(txq);
1165
1166         virtqueue_napi_complete(napi, sq->vq, 0);
1167
1168         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1169                 netif_tx_wake_queue(txq);
1170
1171         return 0;
1172 }
1173
1174 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1175 {
1176         struct virtio_net_hdr_mrg_rxbuf *hdr;
1177         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1178         struct virtnet_info *vi = sq->vq->vdev->priv;
1179         int num_sg;
1180         unsigned hdr_len = vi->hdr_len;
1181         bool can_push;
1182
1183         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1184
1185         can_push = vi->any_header_sg &&
1186                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1187                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1188         /* Even if we can, don't push here yet as this would skew
1189          * csum_start offset below. */
1190         if (can_push)
1191                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1192         else
1193                 hdr = skb_vnet_hdr(skb);
1194
1195         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1196                                     virtio_is_little_endian(vi->vdev), false))
1197                 BUG();
1198
1199         if (vi->mergeable_rx_bufs)
1200                 hdr->num_buffers = 0;
1201
1202         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1203         if (can_push) {
1204                 __skb_push(skb, hdr_len);
1205                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1206                 if (unlikely(num_sg < 0))
1207                         return num_sg;
1208                 /* Pull header back to avoid skew in tx bytes calculations. */
1209                 __skb_pull(skb, hdr_len);
1210         } else {
1211                 sg_set_buf(sq->sg, hdr, hdr_len);
1212                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1213                 if (unlikely(num_sg < 0))
1214                         return num_sg;
1215                 num_sg++;
1216         }
1217         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1218 }
1219
1220 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1221 {
1222         struct virtnet_info *vi = netdev_priv(dev);
1223         int qnum = skb_get_queue_mapping(skb);
1224         struct send_queue *sq = &vi->sq[qnum];
1225         int err;
1226         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1227         bool kick = !skb->xmit_more;
1228         bool use_napi = sq->napi.weight;
1229
1230         /* Free up any pending old buffers before queueing new ones. */
1231         free_old_xmit_skbs(sq);
1232
1233         if (use_napi && kick)
1234                 virtqueue_enable_cb_delayed(sq->vq);
1235
1236         /* timestamp packet in software */
1237         skb_tx_timestamp(skb);
1238
1239         /* Try to transmit */
1240         err = xmit_skb(sq, skb);
1241
1242         /* This should not happen! */
1243         if (unlikely(err)) {
1244                 dev->stats.tx_fifo_errors++;
1245                 if (net_ratelimit())
1246                         dev_warn(&dev->dev,
1247                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1248                 dev->stats.tx_dropped++;
1249                 dev_kfree_skb_any(skb);
1250                 return NETDEV_TX_OK;
1251         }
1252
1253         /* Don't wait up for transmitted skbs to be freed. */
1254         if (!use_napi) {
1255                 skb_orphan(skb);
1256                 nf_reset(skb);
1257         }
1258
1259         /* If running out of space, stop queue to avoid getting packets that we
1260          * are then unable to transmit.
1261          * An alternative would be to force queuing layer to requeue the skb by
1262          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1263          * returned in a normal path of operation: it means that driver is not
1264          * maintaining the TX queue stop/start state properly, and causes
1265          * the stack to do a non-trivial amount of useless work.
1266          * Since most packets only take 1 or 2 ring slots, stopping the queue
1267          * early means 16 slots are typically wasted.
1268          */
1269         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1270                 netif_stop_subqueue(dev, qnum);
1271                 if (!use_napi &&
1272                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1273                         /* More just got used, free them then recheck. */
1274                         free_old_xmit_skbs(sq);
1275                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1276                                 netif_start_subqueue(dev, qnum);
1277                                 virtqueue_disable_cb(sq->vq);
1278                         }
1279                 }
1280         }
1281
1282         if (kick || netif_xmit_stopped(txq))
1283                 virtqueue_kick(sq->vq);
1284
1285         return NETDEV_TX_OK;
1286 }
1287
1288 /*
1289  * Send command via the control virtqueue and check status.  Commands
1290  * supported by the hypervisor, as indicated by feature bits, should
1291  * never fail unless improperly formatted.
1292  */
1293 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1294                                  struct scatterlist *out)
1295 {
1296         struct scatterlist *sgs[4], hdr, stat;
1297         unsigned out_num = 0, tmp;
1298
1299         /* Caller should know better */
1300         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1301
1302         vi->ctrl_status = ~0;
1303         vi->ctrl_hdr.class = class;
1304         vi->ctrl_hdr.cmd = cmd;
1305         /* Add header */
1306         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1307         sgs[out_num++] = &hdr;
1308
1309         if (out)
1310                 sgs[out_num++] = out;
1311
1312         /* Add return status. */
1313         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1314         sgs[out_num] = &stat;
1315
1316         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1317         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1318
1319         if (unlikely(!virtqueue_kick(vi->cvq)))
1320                 return vi->ctrl_status == VIRTIO_NET_OK;
1321
1322         /* Spin for a response, the kick causes an ioport write, trapping
1323          * into the hypervisor, so the request should be handled immediately.
1324          */
1325         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1326                !virtqueue_is_broken(vi->cvq))
1327                 cpu_relax();
1328
1329         return vi->ctrl_status == VIRTIO_NET_OK;
1330 }
1331
1332 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1333 {
1334         struct virtnet_info *vi = netdev_priv(dev);
1335         struct virtio_device *vdev = vi->vdev;
1336         int ret;
1337         struct sockaddr *addr;
1338         struct scatterlist sg;
1339
1340         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1341         if (!addr)
1342                 return -ENOMEM;
1343
1344         ret = eth_prepare_mac_addr_change(dev, addr);
1345         if (ret)
1346                 goto out;
1347
1348         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1349                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1350                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1351                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1352                         dev_warn(&vdev->dev,
1353                                  "Failed to set mac address by vq command.\n");
1354                         ret = -EINVAL;
1355                         goto out;
1356                 }
1357         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1358                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1359                 unsigned int i;
1360
1361                 /* Naturally, this has an atomicity problem. */
1362                 for (i = 0; i < dev->addr_len; i++)
1363                         virtio_cwrite8(vdev,
1364                                        offsetof(struct virtio_net_config, mac) +
1365                                        i, addr->sa_data[i]);
1366         }
1367
1368         eth_commit_mac_addr_change(dev, p);
1369         ret = 0;
1370
1371 out:
1372         kfree(addr);
1373         return ret;
1374 }
1375
1376 static void virtnet_stats(struct net_device *dev,
1377                           struct rtnl_link_stats64 *tot)
1378 {
1379         struct virtnet_info *vi = netdev_priv(dev);
1380         int cpu;
1381         unsigned int start;
1382
1383         for_each_possible_cpu(cpu) {
1384                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1385                 u64 tpackets, tbytes, rpackets, rbytes;
1386
1387                 do {
1388                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1389                         tpackets = stats->tx_packets;
1390                         tbytes   = stats->tx_bytes;
1391                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1392
1393                 do {
1394                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1395                         rpackets = stats->rx_packets;
1396                         rbytes   = stats->rx_bytes;
1397                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1398
1399                 tot->rx_packets += rpackets;
1400                 tot->tx_packets += tpackets;
1401                 tot->rx_bytes   += rbytes;
1402                 tot->tx_bytes   += tbytes;
1403         }
1404
1405         tot->tx_dropped = dev->stats.tx_dropped;
1406         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1407         tot->rx_dropped = dev->stats.rx_dropped;
1408         tot->rx_length_errors = dev->stats.rx_length_errors;
1409         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1410 }
1411
1412 #ifdef CONFIG_NET_POLL_CONTROLLER
1413 static void virtnet_netpoll(struct net_device *dev)
1414 {
1415         struct virtnet_info *vi = netdev_priv(dev);
1416         int i;
1417
1418         for (i = 0; i < vi->curr_queue_pairs; i++)
1419                 napi_schedule(&vi->rq[i].napi);
1420 }
1421 #endif
1422
1423 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1424 {
1425         rtnl_lock();
1426         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1427                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1428                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1429         rtnl_unlock();
1430 }
1431
1432 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1433 {
1434         struct scatterlist sg;
1435         struct net_device *dev = vi->dev;
1436
1437         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1438                 return 0;
1439
1440         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1441         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1442
1443         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1444                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1445                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1446                          queue_pairs);
1447                 return -EINVAL;
1448         } else {
1449                 vi->curr_queue_pairs = queue_pairs;
1450                 /* virtnet_open() will refill when device is going to up. */
1451                 if (dev->flags & IFF_UP)
1452                         schedule_delayed_work(&vi->refill, 0);
1453         }
1454
1455         return 0;
1456 }
1457
1458 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1459 {
1460         int err;
1461
1462         rtnl_lock();
1463         err = _virtnet_set_queues(vi, queue_pairs);
1464         rtnl_unlock();
1465         return err;
1466 }
1467
1468 static int virtnet_close(struct net_device *dev)
1469 {
1470         struct virtnet_info *vi = netdev_priv(dev);
1471         int i;
1472
1473         /* Make sure refill_work doesn't re-enable napi! */
1474         cancel_delayed_work_sync(&vi->refill);
1475
1476         for (i = 0; i < vi->max_queue_pairs; i++) {
1477                 napi_disable(&vi->rq[i].napi);
1478                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1479         }
1480
1481         return 0;
1482 }
1483
1484 static void virtnet_set_rx_mode(struct net_device *dev)
1485 {
1486         struct virtnet_info *vi = netdev_priv(dev);
1487         struct scatterlist sg[2];
1488         struct virtio_net_ctrl_mac *mac_data;
1489         struct netdev_hw_addr *ha;
1490         int uc_count;
1491         int mc_count;
1492         void *buf;
1493         int i;
1494
1495         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1496         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1497                 return;
1498
1499         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1500         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1501
1502         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1503
1504         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1505                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1506                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1507                          vi->ctrl_promisc ? "en" : "dis");
1508
1509         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1510
1511         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1512                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1513                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1514                          vi->ctrl_allmulti ? "en" : "dis");
1515
1516         uc_count = netdev_uc_count(dev);
1517         mc_count = netdev_mc_count(dev);
1518         /* MAC filter - use one buffer for both lists */
1519         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1520                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1521         mac_data = buf;
1522         if (!buf)
1523                 return;
1524
1525         sg_init_table(sg, 2);
1526
1527         /* Store the unicast list and count in the front of the buffer */
1528         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1529         i = 0;
1530         netdev_for_each_uc_addr(ha, dev)
1531                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1532
1533         sg_set_buf(&sg[0], mac_data,
1534                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1535
1536         /* multicast list and count fill the end */
1537         mac_data = (void *)&mac_data->macs[uc_count][0];
1538
1539         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1540         i = 0;
1541         netdev_for_each_mc_addr(ha, dev)
1542                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1543
1544         sg_set_buf(&sg[1], mac_data,
1545                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1546
1547         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1548                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1549                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1550
1551         kfree(buf);
1552 }
1553
1554 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1555                                    __be16 proto, u16 vid)
1556 {
1557         struct virtnet_info *vi = netdev_priv(dev);
1558         struct scatterlist sg;
1559
1560         vi->ctrl_vid = vid;
1561         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1562
1563         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1564                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1565                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1566         return 0;
1567 }
1568
1569 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1570                                     __be16 proto, u16 vid)
1571 {
1572         struct virtnet_info *vi = netdev_priv(dev);
1573         struct scatterlist sg;
1574
1575         vi->ctrl_vid = vid;
1576         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1577
1578         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1579                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1580                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1581         return 0;
1582 }
1583
1584 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1585 {
1586         int i;
1587
1588         if (vi->affinity_hint_set) {
1589                 for (i = 0; i < vi->max_queue_pairs; i++) {
1590                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1591                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1592                 }
1593
1594                 vi->affinity_hint_set = false;
1595         }
1596 }
1597
1598 static void virtnet_set_affinity(struct virtnet_info *vi)
1599 {
1600         int i;
1601         int cpu;
1602
1603         /* In multiqueue mode, when the number of cpu is equal to the number of
1604          * queue pairs, we let the queue pairs to be private to one cpu by
1605          * setting the affinity hint to eliminate the contention.
1606          */
1607         if (vi->curr_queue_pairs == 1 ||
1608             vi->max_queue_pairs != num_online_cpus()) {
1609                 virtnet_clean_affinity(vi, -1);
1610                 return;
1611         }
1612
1613         i = 0;
1614         for_each_online_cpu(cpu) {
1615                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1616                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1617                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1618                 i++;
1619         }
1620
1621         vi->affinity_hint_set = true;
1622 }
1623
1624 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1625 {
1626         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1627                                                    node);
1628         virtnet_set_affinity(vi);
1629         return 0;
1630 }
1631
1632 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1633 {
1634         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1635                                                    node_dead);
1636         virtnet_set_affinity(vi);
1637         return 0;
1638 }
1639
1640 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1641 {
1642         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1643                                                    node);
1644
1645         virtnet_clean_affinity(vi, cpu);
1646         return 0;
1647 }
1648
1649 static enum cpuhp_state virtionet_online;
1650
1651 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1652 {
1653         int ret;
1654
1655         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1656         if (ret)
1657                 return ret;
1658         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1659                                                &vi->node_dead);
1660         if (!ret)
1661                 return ret;
1662         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1663         return ret;
1664 }
1665
1666 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1667 {
1668         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1669         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1670                                             &vi->node_dead);
1671 }
1672
1673 static void virtnet_get_ringparam(struct net_device *dev,
1674                                 struct ethtool_ringparam *ring)
1675 {
1676         struct virtnet_info *vi = netdev_priv(dev);
1677
1678         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1679         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1680         ring->rx_pending = ring->rx_max_pending;
1681         ring->tx_pending = ring->tx_max_pending;
1682 }
1683
1684
1685 static void virtnet_get_drvinfo(struct net_device *dev,
1686                                 struct ethtool_drvinfo *info)
1687 {
1688         struct virtnet_info *vi = netdev_priv(dev);
1689         struct virtio_device *vdev = vi->vdev;
1690
1691         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1692         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1693         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1694
1695 }
1696
1697 /* TODO: Eliminate OOO packets during switching */
1698 static int virtnet_set_channels(struct net_device *dev,
1699                                 struct ethtool_channels *channels)
1700 {
1701         struct virtnet_info *vi = netdev_priv(dev);
1702         u16 queue_pairs = channels->combined_count;
1703         int err;
1704
1705         /* We don't support separate rx/tx channels.
1706          * We don't allow setting 'other' channels.
1707          */
1708         if (channels->rx_count || channels->tx_count || channels->other_count)
1709                 return -EINVAL;
1710
1711         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1712                 return -EINVAL;
1713
1714         /* For now we don't support modifying channels while XDP is loaded
1715          * also when XDP is loaded all RX queues have XDP programs so we only
1716          * need to check a single RX queue.
1717          */
1718         if (vi->rq[0].xdp_prog)
1719                 return -EINVAL;
1720
1721         get_online_cpus();
1722         err = _virtnet_set_queues(vi, queue_pairs);
1723         if (!err) {
1724                 netif_set_real_num_tx_queues(dev, queue_pairs);
1725                 netif_set_real_num_rx_queues(dev, queue_pairs);
1726
1727                 virtnet_set_affinity(vi);
1728         }
1729         put_online_cpus();
1730
1731         return err;
1732 }
1733
1734 static void virtnet_get_channels(struct net_device *dev,
1735                                  struct ethtool_channels *channels)
1736 {
1737         struct virtnet_info *vi = netdev_priv(dev);
1738
1739         channels->combined_count = vi->curr_queue_pairs;
1740         channels->max_combined = vi->max_queue_pairs;
1741         channels->max_other = 0;
1742         channels->rx_count = 0;
1743         channels->tx_count = 0;
1744         channels->other_count = 0;
1745 }
1746
1747 /* Check if the user is trying to change anything besides speed/duplex */
1748 static bool
1749 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1750 {
1751         struct ethtool_link_ksettings diff1 = *cmd;
1752         struct ethtool_link_ksettings diff2 = {};
1753
1754         /* cmd is always set so we need to clear it, validate the port type
1755          * and also without autonegotiation we can ignore advertising
1756          */
1757         diff1.base.speed = 0;
1758         diff2.base.port = PORT_OTHER;
1759         ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
1760         diff1.base.duplex = 0;
1761         diff1.base.cmd = 0;
1762         diff1.base.link_mode_masks_nwords = 0;
1763
1764         return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
1765                 bitmap_empty(diff1.link_modes.supported,
1766                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1767                 bitmap_empty(diff1.link_modes.advertising,
1768                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1769                 bitmap_empty(diff1.link_modes.lp_advertising,
1770                              __ETHTOOL_LINK_MODE_MASK_NBITS);
1771 }
1772
1773 static int virtnet_set_link_ksettings(struct net_device *dev,
1774                                       const struct ethtool_link_ksettings *cmd)
1775 {
1776         struct virtnet_info *vi = netdev_priv(dev);
1777         u32 speed;
1778
1779         speed = cmd->base.speed;
1780         /* don't allow custom speed and duplex */
1781         if (!ethtool_validate_speed(speed) ||
1782             !ethtool_validate_duplex(cmd->base.duplex) ||
1783             !virtnet_validate_ethtool_cmd(cmd))
1784                 return -EINVAL;
1785         vi->speed = speed;
1786         vi->duplex = cmd->base.duplex;
1787
1788         return 0;
1789 }
1790
1791 static int virtnet_get_link_ksettings(struct net_device *dev,
1792                                       struct ethtool_link_ksettings *cmd)
1793 {
1794         struct virtnet_info *vi = netdev_priv(dev);
1795
1796         cmd->base.speed = vi->speed;
1797         cmd->base.duplex = vi->duplex;
1798         cmd->base.port = PORT_OTHER;
1799
1800         return 0;
1801 }
1802
1803 static void virtnet_init_settings(struct net_device *dev)
1804 {
1805         struct virtnet_info *vi = netdev_priv(dev);
1806
1807         vi->speed = SPEED_UNKNOWN;
1808         vi->duplex = DUPLEX_UNKNOWN;
1809 }
1810
1811 static const struct ethtool_ops virtnet_ethtool_ops = {
1812         .get_drvinfo = virtnet_get_drvinfo,
1813         .get_link = ethtool_op_get_link,
1814         .get_ringparam = virtnet_get_ringparam,
1815         .set_channels = virtnet_set_channels,
1816         .get_channels = virtnet_get_channels,
1817         .get_ts_info = ethtool_op_get_ts_info,
1818         .get_link_ksettings = virtnet_get_link_ksettings,
1819         .set_link_ksettings = virtnet_set_link_ksettings,
1820 };
1821
1822 static void virtnet_freeze_down(struct virtio_device *vdev)
1823 {
1824         struct virtnet_info *vi = vdev->priv;
1825         int i;
1826
1827         /* Make sure no work handler is accessing the device */
1828         flush_work(&vi->config_work);
1829
1830         netif_device_detach(vi->dev);
1831         netif_tx_disable(vi->dev);
1832         cancel_delayed_work_sync(&vi->refill);
1833
1834         if (netif_running(vi->dev)) {
1835                 for (i = 0; i < vi->max_queue_pairs; i++) {
1836                         napi_disable(&vi->rq[i].napi);
1837                         virtnet_napi_tx_disable(&vi->sq[i].napi);
1838                 }
1839         }
1840 }
1841
1842 static int init_vqs(struct virtnet_info *vi);
1843 static void _remove_vq_common(struct virtnet_info *vi);
1844
1845 static int virtnet_restore_up(struct virtio_device *vdev)
1846 {
1847         struct virtnet_info *vi = vdev->priv;
1848         int err, i;
1849
1850         err = init_vqs(vi);
1851         if (err)
1852                 return err;
1853
1854         virtio_device_ready(vdev);
1855
1856         if (netif_running(vi->dev)) {
1857                 for (i = 0; i < vi->curr_queue_pairs; i++)
1858                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1859                                 schedule_delayed_work(&vi->refill, 0);
1860
1861                 for (i = 0; i < vi->max_queue_pairs; i++) {
1862                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1863                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
1864                                                &vi->sq[i].napi);
1865                 }
1866         }
1867
1868         netif_device_attach(vi->dev);
1869         return err;
1870 }
1871
1872 static int virtnet_reset(struct virtnet_info *vi, int curr_qp, int xdp_qp)
1873 {
1874         struct virtio_device *dev = vi->vdev;
1875         int ret;
1876
1877         virtio_config_disable(dev);
1878         dev->failed = dev->config->get_status(dev) & VIRTIO_CONFIG_S_FAILED;
1879         virtnet_freeze_down(dev);
1880         _remove_vq_common(vi);
1881
1882         virtio_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE);
1883         virtio_add_status(dev, VIRTIO_CONFIG_S_DRIVER);
1884
1885         ret = virtio_finalize_features(dev);
1886         if (ret)
1887                 goto err;
1888
1889         vi->xdp_queue_pairs = xdp_qp;
1890         ret = virtnet_restore_up(dev);
1891         if (ret)
1892                 goto err;
1893         ret = _virtnet_set_queues(vi, curr_qp);
1894         if (ret)
1895                 goto err;
1896
1897         virtio_add_status(dev, VIRTIO_CONFIG_S_DRIVER_OK);
1898         virtio_config_enable(dev);
1899         return 0;
1900 err:
1901         virtio_add_status(dev, VIRTIO_CONFIG_S_FAILED);
1902         return ret;
1903 }
1904
1905 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1906                            struct netlink_ext_ack *extack)
1907 {
1908         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1909         struct virtnet_info *vi = netdev_priv(dev);
1910         struct bpf_prog *old_prog;
1911         u16 xdp_qp = 0, curr_qp;
1912         int i, err;
1913
1914         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1915             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1916             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
1917             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO)) {
1918                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
1919                 return -EOPNOTSUPP;
1920         }
1921
1922         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1923                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
1924                 return -EINVAL;
1925         }
1926
1927         if (dev->mtu > max_sz) {
1928                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
1929                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1930                 return -EINVAL;
1931         }
1932
1933         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1934         if (prog)
1935                 xdp_qp = nr_cpu_ids;
1936
1937         /* XDP requires extra queues for XDP_TX */
1938         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1939                 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
1940                 netdev_warn(dev, "request %i queues but max is %i\n",
1941                             curr_qp + xdp_qp, vi->max_queue_pairs);
1942                 return -ENOMEM;
1943         }
1944
1945         if (prog) {
1946                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1947                 if (IS_ERR(prog))
1948                         return PTR_ERR(prog);
1949         }
1950
1951         /* Changing the headroom in buffers is a disruptive operation because
1952          * existing buffers must be flushed and reallocated. This will happen
1953          * when a xdp program is initially added or xdp is disabled by removing
1954          * the xdp program resulting in number of XDP queues changing.
1955          */
1956         if (vi->xdp_queue_pairs != xdp_qp) {
1957                 err = virtnet_reset(vi, curr_qp + xdp_qp, xdp_qp);
1958                 if (err) {
1959                         dev_warn(&dev->dev, "XDP reset failure.\n");
1960                         goto virtio_reset_err;
1961                 }
1962         }
1963
1964         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
1965
1966         for (i = 0; i < vi->max_queue_pairs; i++) {
1967                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1968                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
1969                 if (old_prog)
1970                         bpf_prog_put(old_prog);
1971         }
1972
1973         return 0;
1974
1975 virtio_reset_err:
1976         /* On reset error do our best to unwind XDP changes inflight and return
1977          * error up to user space for resolution. The underlying reset hung on
1978          * us so not much we can do here.
1979          */
1980         if (prog)
1981                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
1982         return err;
1983 }
1984
1985 static u32 virtnet_xdp_query(struct net_device *dev)
1986 {
1987         struct virtnet_info *vi = netdev_priv(dev);
1988         const struct bpf_prog *xdp_prog;
1989         int i;
1990
1991         for (i = 0; i < vi->max_queue_pairs; i++) {
1992                 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1993                 if (xdp_prog)
1994                         return xdp_prog->aux->id;
1995         }
1996         return 0;
1997 }
1998
1999 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
2000 {
2001         switch (xdp->command) {
2002         case XDP_SETUP_PROG:
2003                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2004         case XDP_QUERY_PROG:
2005                 xdp->prog_id = virtnet_xdp_query(dev);
2006                 xdp->prog_attached = !!xdp->prog_id;
2007                 return 0;
2008         default:
2009                 return -EINVAL;
2010         }
2011 }
2012
2013 static const struct net_device_ops virtnet_netdev = {
2014         .ndo_open            = virtnet_open,
2015         .ndo_stop            = virtnet_close,
2016         .ndo_start_xmit      = start_xmit,
2017         .ndo_validate_addr   = eth_validate_addr,
2018         .ndo_set_mac_address = virtnet_set_mac_address,
2019         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2020         .ndo_get_stats64     = virtnet_stats,
2021         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2022         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2023 #ifdef CONFIG_NET_POLL_CONTROLLER
2024         .ndo_poll_controller = virtnet_netpoll,
2025 #endif
2026         .ndo_xdp                = virtnet_xdp,
2027         .ndo_features_check     = passthru_features_check,
2028 };
2029
2030 static void virtnet_config_changed_work(struct work_struct *work)
2031 {
2032         struct virtnet_info *vi =
2033                 container_of(work, struct virtnet_info, config_work);
2034         u16 v;
2035
2036         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2037                                  struct virtio_net_config, status, &v) < 0)
2038                 return;
2039
2040         if (v & VIRTIO_NET_S_ANNOUNCE) {
2041                 netdev_notify_peers(vi->dev);
2042                 virtnet_ack_link_announce(vi);
2043         }
2044
2045         /* Ignore unknown (future) status bits */
2046         v &= VIRTIO_NET_S_LINK_UP;
2047
2048         if (vi->status == v)
2049                 return;
2050
2051         vi->status = v;
2052
2053         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2054                 netif_carrier_on(vi->dev);
2055                 netif_tx_wake_all_queues(vi->dev);
2056         } else {
2057                 netif_carrier_off(vi->dev);
2058                 netif_tx_stop_all_queues(vi->dev);
2059         }
2060 }
2061
2062 static void virtnet_config_changed(struct virtio_device *vdev)
2063 {
2064         struct virtnet_info *vi = vdev->priv;
2065
2066         schedule_work(&vi->config_work);
2067 }
2068
2069 static void virtnet_free_queues(struct virtnet_info *vi)
2070 {
2071         int i;
2072
2073         for (i = 0; i < vi->max_queue_pairs; i++) {
2074                 napi_hash_del(&vi->rq[i].napi);
2075                 netif_napi_del(&vi->rq[i].napi);
2076                 netif_napi_del(&vi->sq[i].napi);
2077         }
2078
2079         /* We called napi_hash_del() before netif_napi_del(),
2080          * we need to respect an RCU grace period before freeing vi->rq
2081          */
2082         synchronize_net();
2083
2084         kfree(vi->rq);
2085         kfree(vi->sq);
2086 }
2087
2088 static void _free_receive_bufs(struct virtnet_info *vi)
2089 {
2090         struct bpf_prog *old_prog;
2091         int i;
2092
2093         for (i = 0; i < vi->max_queue_pairs; i++) {
2094                 while (vi->rq[i].pages)
2095                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2096
2097                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2098                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2099                 if (old_prog)
2100                         bpf_prog_put(old_prog);
2101         }
2102 }
2103
2104 static void free_receive_bufs(struct virtnet_info *vi)
2105 {
2106         rtnl_lock();
2107         _free_receive_bufs(vi);
2108         rtnl_unlock();
2109 }
2110
2111 static void free_receive_page_frags(struct virtnet_info *vi)
2112 {
2113         int i;
2114         for (i = 0; i < vi->max_queue_pairs; i++)
2115                 if (vi->rq[i].alloc_frag.page)
2116                         put_page(vi->rq[i].alloc_frag.page);
2117 }
2118
2119 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2120 {
2121         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2122                 return false;
2123         else if (q < vi->curr_queue_pairs)
2124                 return true;
2125         else
2126                 return false;
2127 }
2128
2129 static void free_unused_bufs(struct virtnet_info *vi)
2130 {
2131         void *buf;
2132         int i;
2133
2134         for (i = 0; i < vi->max_queue_pairs; i++) {
2135                 struct virtqueue *vq = vi->sq[i].vq;
2136                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2137                         if (!is_xdp_raw_buffer_queue(vi, i))
2138                                 dev_kfree_skb(buf);
2139                         else
2140                                 put_page(virt_to_head_page(buf));
2141                 }
2142         }
2143
2144         for (i = 0; i < vi->max_queue_pairs; i++) {
2145                 struct virtqueue *vq = vi->rq[i].vq;
2146
2147                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2148                         if (vi->mergeable_rx_bufs) {
2149                                 put_page(virt_to_head_page(buf));
2150                         } else if (vi->big_packets) {
2151                                 give_pages(&vi->rq[i], buf);
2152                         } else {
2153                                 put_page(virt_to_head_page(buf));
2154                         }
2155                 }
2156         }
2157 }
2158
2159 static void virtnet_del_vqs(struct virtnet_info *vi)
2160 {
2161         struct virtio_device *vdev = vi->vdev;
2162
2163         virtnet_clean_affinity(vi, -1);
2164
2165         vdev->config->del_vqs(vdev);
2166
2167         virtnet_free_queues(vi);
2168 }
2169
2170 /* How large should a single buffer be so a queue full of these can fit at
2171  * least one full packet?
2172  * Logic below assumes the mergeable buffer header is used.
2173  */
2174 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2175 {
2176         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2177         unsigned int rq_size = virtqueue_get_vring_size(vq);
2178         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2179         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2180         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2181
2182         return max(max(min_buf_len, hdr_len) - hdr_len,
2183                    (unsigned int)GOOD_PACKET_LEN);
2184 }
2185
2186 static int virtnet_find_vqs(struct virtnet_info *vi)
2187 {
2188         vq_callback_t **callbacks;
2189         struct virtqueue **vqs;
2190         int ret = -ENOMEM;
2191         int i, total_vqs;
2192         const char **names;
2193         bool *ctx;
2194
2195         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2196          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2197          * possible control vq.
2198          */
2199         total_vqs = vi->max_queue_pairs * 2 +
2200                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2201
2202         /* Allocate space for find_vqs parameters */
2203         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2204         if (!vqs)
2205                 goto err_vq;
2206         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2207         if (!callbacks)
2208                 goto err_callback;
2209         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2210         if (!names)
2211                 goto err_names;
2212         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2213                 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2214                 if (!ctx)
2215                         goto err_ctx;
2216         } else {
2217                 ctx = NULL;
2218         }
2219
2220         /* Parameters for control virtqueue, if any */
2221         if (vi->has_cvq) {
2222                 callbacks[total_vqs - 1] = NULL;
2223                 names[total_vqs - 1] = "control";
2224         }
2225
2226         /* Allocate/initialize parameters for send/receive virtqueues */
2227         for (i = 0; i < vi->max_queue_pairs; i++) {
2228                 callbacks[rxq2vq(i)] = skb_recv_done;
2229                 callbacks[txq2vq(i)] = skb_xmit_done;
2230                 sprintf(vi->rq[i].name, "input.%d", i);
2231                 sprintf(vi->sq[i].name, "output.%d", i);
2232                 names[rxq2vq(i)] = vi->rq[i].name;
2233                 names[txq2vq(i)] = vi->sq[i].name;
2234                 if (ctx)
2235                         ctx[rxq2vq(i)] = true;
2236         }
2237
2238         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2239                                          names, ctx, NULL);
2240         if (ret)
2241                 goto err_find;
2242
2243         if (vi->has_cvq) {
2244                 vi->cvq = vqs[total_vqs - 1];
2245                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2246                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2247         }
2248
2249         for (i = 0; i < vi->max_queue_pairs; i++) {
2250                 vi->rq[i].vq = vqs[rxq2vq(i)];
2251                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2252                 vi->sq[i].vq = vqs[txq2vq(i)];
2253         }
2254
2255         kfree(names);
2256         kfree(callbacks);
2257         kfree(vqs);
2258         kfree(ctx);
2259
2260         return 0;
2261
2262 err_find:
2263         kfree(ctx);
2264 err_ctx:
2265         kfree(names);
2266 err_names:
2267         kfree(callbacks);
2268 err_callback:
2269         kfree(vqs);
2270 err_vq:
2271         return ret;
2272 }
2273
2274 static int virtnet_alloc_queues(struct virtnet_info *vi)
2275 {
2276         int i;
2277
2278         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2279         if (!vi->sq)
2280                 goto err_sq;
2281         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2282         if (!vi->rq)
2283                 goto err_rq;
2284
2285         INIT_DELAYED_WORK(&vi->refill, refill_work);
2286         for (i = 0; i < vi->max_queue_pairs; i++) {
2287                 vi->rq[i].pages = NULL;
2288                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2289                                napi_weight);
2290                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2291                                   napi_tx ? napi_weight : 0);
2292
2293                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2294                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2295                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2296         }
2297
2298         return 0;
2299
2300 err_rq:
2301         kfree(vi->sq);
2302 err_sq:
2303         return -ENOMEM;
2304 }
2305
2306 static int init_vqs(struct virtnet_info *vi)
2307 {
2308         int ret;
2309
2310         /* Allocate send & receive queues */
2311         ret = virtnet_alloc_queues(vi);
2312         if (ret)
2313                 goto err;
2314
2315         ret = virtnet_find_vqs(vi);
2316         if (ret)
2317                 goto err_free;
2318
2319         get_online_cpus();
2320         virtnet_set_affinity(vi);
2321         put_online_cpus();
2322
2323         return 0;
2324
2325 err_free:
2326         virtnet_free_queues(vi);
2327 err:
2328         return ret;
2329 }
2330
2331 #ifdef CONFIG_SYSFS
2332 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2333                 struct rx_queue_attribute *attribute, char *buf)
2334 {
2335         struct virtnet_info *vi = netdev_priv(queue->dev);
2336         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2337         struct ewma_pkt_len *avg;
2338
2339         BUG_ON(queue_index >= vi->max_queue_pairs);
2340         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2341         return sprintf(buf, "%u\n",
2342                        get_mergeable_buf_len(&vi->rq[queue_index], avg));
2343 }
2344
2345 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2346         __ATTR_RO(mergeable_rx_buffer_size);
2347
2348 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2349         &mergeable_rx_buffer_size_attribute.attr,
2350         NULL
2351 };
2352
2353 static const struct attribute_group virtio_net_mrg_rx_group = {
2354         .name = "virtio_net",
2355         .attrs = virtio_net_mrg_rx_attrs
2356 };
2357 #endif
2358
2359 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2360                                     unsigned int fbit,
2361                                     const char *fname, const char *dname)
2362 {
2363         if (!virtio_has_feature(vdev, fbit))
2364                 return false;
2365
2366         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2367                 fname, dname);
2368
2369         return true;
2370 }
2371
2372 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2373         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2374
2375 static bool virtnet_validate_features(struct virtio_device *vdev)
2376 {
2377         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2378             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2379                              "VIRTIO_NET_F_CTRL_VQ") ||
2380              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2381                              "VIRTIO_NET_F_CTRL_VQ") ||
2382              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2383                              "VIRTIO_NET_F_CTRL_VQ") ||
2384              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2385              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2386                              "VIRTIO_NET_F_CTRL_VQ"))) {
2387                 return false;
2388         }
2389
2390         return true;
2391 }
2392
2393 #define MIN_MTU ETH_MIN_MTU
2394 #define MAX_MTU ETH_MAX_MTU
2395
2396 static int virtnet_validate(struct virtio_device *vdev)
2397 {
2398         if (!vdev->config->get) {
2399                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2400                         __func__);
2401                 return -EINVAL;
2402         }
2403
2404         if (!virtnet_validate_features(vdev))
2405                 return -EINVAL;
2406
2407         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2408                 int mtu = virtio_cread16(vdev,
2409                                          offsetof(struct virtio_net_config,
2410                                                   mtu));
2411                 if (mtu < MIN_MTU)
2412                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2413         }
2414
2415         return 0;
2416 }
2417
2418 static int virtnet_probe(struct virtio_device *vdev)
2419 {
2420         int i, err;
2421         struct net_device *dev;
2422         struct virtnet_info *vi;
2423         u16 max_queue_pairs;
2424         int mtu;
2425
2426         /* Find if host supports multiqueue virtio_net device */
2427         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2428                                    struct virtio_net_config,
2429                                    max_virtqueue_pairs, &max_queue_pairs);
2430
2431         /* We need at least 2 queue's */
2432         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2433             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2434             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2435                 max_queue_pairs = 1;
2436
2437         /* Allocate ourselves a network device with room for our info */
2438         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2439         if (!dev)
2440                 return -ENOMEM;
2441
2442         /* Set up network device as normal. */
2443         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2444         dev->netdev_ops = &virtnet_netdev;
2445         dev->features = NETIF_F_HIGHDMA;
2446
2447         dev->ethtool_ops = &virtnet_ethtool_ops;
2448         SET_NETDEV_DEV(dev, &vdev->dev);
2449
2450         /* Do we support "hardware" checksums? */
2451         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2452                 /* This opens up the world of extra features. */
2453                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2454                 if (csum)
2455                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2456
2457                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2458                         dev->hw_features |= NETIF_F_TSO
2459                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2460                 }
2461                 /* Individual feature bits: what can host handle? */
2462                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2463                         dev->hw_features |= NETIF_F_TSO;
2464                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2465                         dev->hw_features |= NETIF_F_TSO6;
2466                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2467                         dev->hw_features |= NETIF_F_TSO_ECN;
2468
2469                 dev->features |= NETIF_F_GSO_ROBUST;
2470
2471                 if (gso)
2472                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2473                 /* (!csum && gso) case will be fixed by register_netdev() */
2474         }
2475         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2476                 dev->features |= NETIF_F_RXCSUM;
2477
2478         dev->vlan_features = dev->features;
2479
2480         /* MTU range: 68 - 65535 */
2481         dev->min_mtu = MIN_MTU;
2482         dev->max_mtu = MAX_MTU;
2483
2484         /* Configuration may specify what MAC to use.  Otherwise random. */
2485         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2486                 virtio_cread_bytes(vdev,
2487                                    offsetof(struct virtio_net_config, mac),
2488                                    dev->dev_addr, dev->addr_len);
2489         else
2490                 eth_hw_addr_random(dev);
2491
2492         /* Set up our device-specific information */
2493         vi = netdev_priv(dev);
2494         vi->dev = dev;
2495         vi->vdev = vdev;
2496         vdev->priv = vi;
2497         vi->stats = alloc_percpu(struct virtnet_stats);
2498         err = -ENOMEM;
2499         if (vi->stats == NULL)
2500                 goto free;
2501
2502         for_each_possible_cpu(i) {
2503                 struct virtnet_stats *virtnet_stats;
2504                 virtnet_stats = per_cpu_ptr(vi->stats, i);
2505                 u64_stats_init(&virtnet_stats->tx_syncp);
2506                 u64_stats_init(&virtnet_stats->rx_syncp);
2507         }
2508
2509         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2510
2511         /* If we can receive ANY GSO packets, we must allocate large ones. */
2512         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2513             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2514             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2515             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2516                 vi->big_packets = true;
2517
2518         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2519                 vi->mergeable_rx_bufs = true;
2520
2521         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2522             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2523                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2524         else
2525                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2526
2527         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2528             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2529                 vi->any_header_sg = true;
2530
2531         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2532                 vi->has_cvq = true;
2533
2534         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2535                 mtu = virtio_cread16(vdev,
2536                                      offsetof(struct virtio_net_config,
2537                                               mtu));
2538                 if (mtu < dev->min_mtu) {
2539                         /* Should never trigger: MTU was previously validated
2540                          * in virtnet_validate.
2541                          */
2542                         dev_err(&vdev->dev, "device MTU appears to have changed "
2543                                 "it is now %d < %d", mtu, dev->min_mtu);
2544                         goto free_stats;
2545                 }
2546
2547                 dev->mtu = mtu;
2548                 dev->max_mtu = mtu;
2549
2550                 /* TODO: size buffers correctly in this case. */
2551                 if (dev->mtu > ETH_DATA_LEN)
2552                         vi->big_packets = true;
2553         }
2554
2555         if (vi->any_header_sg)
2556                 dev->needed_headroom = vi->hdr_len;
2557
2558         /* Enable multiqueue by default */
2559         if (num_online_cpus() >= max_queue_pairs)
2560                 vi->curr_queue_pairs = max_queue_pairs;
2561         else
2562                 vi->curr_queue_pairs = num_online_cpus();
2563         vi->max_queue_pairs = max_queue_pairs;
2564
2565         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2566         err = init_vqs(vi);
2567         if (err)
2568                 goto free_stats;
2569
2570 #ifdef CONFIG_SYSFS
2571         if (vi->mergeable_rx_bufs)
2572                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2573 #endif
2574         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2575         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2576
2577         virtnet_init_settings(dev);
2578
2579         err = register_netdev(dev);
2580         if (err) {
2581                 pr_debug("virtio_net: registering device failed\n");
2582                 goto free_vqs;
2583         }
2584
2585         virtio_device_ready(vdev);
2586
2587         err = virtnet_cpu_notif_add(vi);
2588         if (err) {
2589                 pr_debug("virtio_net: registering cpu notifier failed\n");
2590                 goto free_unregister_netdev;
2591         }
2592
2593         virtnet_set_queues(vi, vi->curr_queue_pairs);
2594
2595         /* Assume link up if device can't report link status,
2596            otherwise get link status from config. */
2597         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2598                 netif_carrier_off(dev);
2599                 schedule_work(&vi->config_work);
2600         } else {
2601                 vi->status = VIRTIO_NET_S_LINK_UP;
2602                 netif_carrier_on(dev);
2603         }
2604
2605         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2606                  dev->name, max_queue_pairs);
2607
2608         return 0;
2609
2610 free_unregister_netdev:
2611         vi->vdev->config->reset(vdev);
2612
2613         unregister_netdev(dev);
2614 free_vqs:
2615         cancel_delayed_work_sync(&vi->refill);
2616         free_receive_page_frags(vi);
2617         virtnet_del_vqs(vi);
2618 free_stats:
2619         free_percpu(vi->stats);
2620 free:
2621         free_netdev(dev);
2622         return err;
2623 }
2624
2625 static void _remove_vq_common(struct virtnet_info *vi)
2626 {
2627         vi->vdev->config->reset(vi->vdev);
2628         free_unused_bufs(vi);
2629         _free_receive_bufs(vi);
2630         free_receive_page_frags(vi);
2631         virtnet_del_vqs(vi);
2632 }
2633
2634 static void remove_vq_common(struct virtnet_info *vi)
2635 {
2636         vi->vdev->config->reset(vi->vdev);
2637
2638         /* Free unused buffers in both send and recv, if any. */
2639         free_unused_bufs(vi);
2640
2641         free_receive_bufs(vi);
2642
2643         free_receive_page_frags(vi);
2644
2645         virtnet_del_vqs(vi);
2646 }
2647
2648 static void virtnet_remove(struct virtio_device *vdev)
2649 {
2650         struct virtnet_info *vi = vdev->priv;
2651
2652         virtnet_cpu_notif_remove(vi);
2653
2654         /* Make sure no work handler is accessing the device. */
2655         flush_work(&vi->config_work);
2656
2657         unregister_netdev(vi->dev);
2658
2659         remove_vq_common(vi);
2660
2661         free_percpu(vi->stats);
2662         free_netdev(vi->dev);
2663 }
2664
2665 #ifdef CONFIG_PM_SLEEP
2666 static int virtnet_freeze(struct virtio_device *vdev)
2667 {
2668         struct virtnet_info *vi = vdev->priv;
2669
2670         virtnet_cpu_notif_remove(vi);
2671         virtnet_freeze_down(vdev);
2672         remove_vq_common(vi);
2673
2674         return 0;
2675 }
2676
2677 static int virtnet_restore(struct virtio_device *vdev)
2678 {
2679         struct virtnet_info *vi = vdev->priv;
2680         int err;
2681
2682         err = virtnet_restore_up(vdev);
2683         if (err)
2684                 return err;
2685         virtnet_set_queues(vi, vi->curr_queue_pairs);
2686
2687         err = virtnet_cpu_notif_add(vi);
2688         if (err)
2689                 return err;
2690
2691         return 0;
2692 }
2693 #endif
2694
2695 static struct virtio_device_id id_table[] = {
2696         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2697         { 0 },
2698 };
2699
2700 #define VIRTNET_FEATURES \
2701         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2702         VIRTIO_NET_F_MAC, \
2703         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2704         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2705         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2706         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2707         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2708         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2709         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2710         VIRTIO_NET_F_MTU
2711
2712 static unsigned int features[] = {
2713         VIRTNET_FEATURES,
2714 };
2715
2716 static unsigned int features_legacy[] = {
2717         VIRTNET_FEATURES,
2718         VIRTIO_NET_F_GSO,
2719         VIRTIO_F_ANY_LAYOUT,
2720 };
2721
2722 static struct virtio_driver virtio_net_driver = {
2723         .feature_table = features,
2724         .feature_table_size = ARRAY_SIZE(features),
2725         .feature_table_legacy = features_legacy,
2726         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2727         .driver.name =  KBUILD_MODNAME,
2728         .driver.owner = THIS_MODULE,
2729         .id_table =     id_table,
2730         .validate =     virtnet_validate,
2731         .probe =        virtnet_probe,
2732         .remove =       virtnet_remove,
2733         .config_changed = virtnet_config_changed,
2734 #ifdef CONFIG_PM_SLEEP
2735         .freeze =       virtnet_freeze,
2736         .restore =      virtnet_restore,
2737 #endif
2738 };
2739
2740 static __init int virtio_net_driver_init(void)
2741 {
2742         int ret;
2743
2744         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2745                                       virtnet_cpu_online,
2746                                       virtnet_cpu_down_prep);
2747         if (ret < 0)
2748                 goto out;
2749         virtionet_online = ret;
2750         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2751                                       NULL, virtnet_cpu_dead);
2752         if (ret)
2753                 goto err_dead;
2754
2755         ret = register_virtio_driver(&virtio_net_driver);
2756         if (ret)
2757                 goto err_virtio;
2758         return 0;
2759 err_virtio:
2760         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2761 err_dead:
2762         cpuhp_remove_multi_state(virtionet_online);
2763 out:
2764         return ret;
2765 }
2766 module_init(virtio_net_driver_init);
2767
2768 static __exit void virtio_net_driver_exit(void)
2769 {
2770         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2771         cpuhp_remove_multi_state(virtionet_online);
2772         unregister_virtio_driver(&virtio_net_driver);
2773 }
2774 module_exit(virtio_net_driver_exit);
2775
2776 MODULE_DEVICE_TABLE(virtio, id_table);
2777 MODULE_DESCRIPTION("Virtio network driver");
2778 MODULE_LICENSE("GPL");