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