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, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  18  */
  19 //#define DEBUG
  20 #include <linux/netdevice.h>
  21 #include <linux/etherdevice.h>
  22 #include <linux/ethtool.h>
  23 #include <linux/module.h>
  24 #include <linux/virtio.h>
  25 #include <linux/virtio_net.h>
  26 #include <linux/scatterlist.h>
  27 #include <linux/if_vlan.h>
  28 #include <linux/slab.h>
  29 #include <linux/cpu.h>
  30
  31 static int napi_weight = 128;
  32 module_param(napi_weight, int, 0444);
  33
  34 static bool csum = true, gso = true;
  35 module_param(csum, bool, 0444);
  36 module_param(gso, bool, 0444);
  37
  38 /* FIXME: MTU in config. */
  39 #define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  40 #define GOOD_COPY_LEN   128
  41
  42 #define VIRTNET_DRIVER_VERSION "1.0.0"
  43
  44 struct virtnet_stats {
  45         struct u64_stats_sync tx_syncp;
  46         struct u64_stats_sync rx_syncp;
  47         u64 tx_bytes;
  48         u64 tx_packets;
  49
  50         u64 rx_bytes;
  51         u64 rx_packets;
  52 };
  53
  54 /* Internal representation of a send virtqueue */
  55 struct send_queue {
  56         /* Virtqueue associated with this send _queue */
  57         struct virtqueue *vq;
  58
  59         /* TX: fragments + linear part + virtio header */
  60         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  61
  62         /* Name of the send queue: output.$index */
  63         char name[40];
  64 };
  65
  66 /* Internal representation of a receive virtqueue */
  67 struct receive_queue {
  68         /* Virtqueue associated with this receive_queue */
  69         struct virtqueue *vq;
  70
  71         struct napi_struct napi;
  72
  73         /* Number of input buffers, and max we've ever had. */
  74         unsigned int num, max;
  75
  76         /* Chain pages by the private ptr. */
  77         struct page *pages;
  78
  79         /* RX: fragments + linear part + virtio header */
  80         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  81
  82         /* Name of this receive queue: input.$index */
  83         char name[40];
  84 };
  85
  86 struct virtnet_info {
  87         struct virtio_device *vdev;
  88         struct virtqueue *cvq;
  89         struct net_device *dev;
  90         struct send_queue *sq;
  91         struct receive_queue *rq;
  92         unsigned int status;
  93
  94         /* Max # of queue pairs supported by the device */
  95         u16 max_queue_pairs;
  96
  97         /* # of queue pairs currently used by the driver */
  98         u16 curr_queue_pairs;
  99
 100         /* I like... big packets and I cannot lie! */
 101         bool big_packets;
 102
 103         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 104         bool mergeable_rx_bufs;
 105
 106         /* Has control virtqueue */
 107         bool has_cvq;
 108
 109         /* enable config space updates */
 110         bool config_enable;
 111
 112         /* Active statistics */
 113         struct virtnet_stats __percpu *stats;
 114
 115         /* Work struct for refilling if we run low on memory. */
 116         struct delayed_work refill;
 117
 118         /* Work struct for config space updates */
 119         struct work_struct config_work;
 120
 121         /* Lock for config space updates */
 122         struct mutex config_lock;
 123
 124         /* Does the affinity hint is set for virtqueues? */
 125         bool affinity_hint_set;
 126
 127         /* Per-cpu variable to show the mapping from CPU to virtqueue */
 128         int __percpu *vq_index;
 129
 130         /* CPU hot plug notifier */
 131         struct notifier_block nb;
 132 };
 133
 134 struct skb_vnet_hdr {
 135         union {
 136                 struct virtio_net_hdr hdr;
 137                 struct virtio_net_hdr_mrg_rxbuf mhdr;
 138         };
 139 };
 140
 141 struct padded_vnet_hdr {
 142         struct virtio_net_hdr hdr;
 143         /*
 144          * virtio_net_hdr should be in a separated sg buffer because of a
 145          * QEMU bug, and data sg buffer shares same page with this header sg.
 146          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
 147          */
 148         char padding[6];
 149 };
 150
 151 /* Converting between virtqueue no. and kernel tx/rx queue no.
 152  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 153  */
 154 static int vq2txq(struct virtqueue *vq)
 155 {
 156         return (virtqueue_get_queue_index(vq) - 1) / 2;
 157 }
 158
 159 static int txq2vq(int txq)
 160 {
 161         return txq * 2 + 1;
 162 }
 163
 164 static int vq2rxq(struct virtqueue *vq)
 165 {
 166         return virtqueue_get_queue_index(vq) / 2;
 167 }
 168
 169 static int rxq2vq(int rxq)
 170 {
 171         return rxq * 2;
 172 }
 173
 174 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 175 {
 176         return (struct skb_vnet_hdr *)skb->cb;
 177 }
 178
 179 /*
 180  * private is used to chain pages for big packets, put the whole
 181  * most recent used list in the beginning for reuse
 182  */
 183 static void give_pages(struct receive_queue *rq, struct page *page)
 184 {
 185         struct page *end;
 186
 187         /* Find end of list, sew whole thing into vi->rq.pages. */
 188         for (end = page; end->private; end = (struct page *)end->private);
 189         end->private = (unsigned long)rq->pages;
 190         rq->pages = page;
 191 }
 192
 193 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 194 {
 195         struct page *p = rq->pages;
 196
 197         if (p) {
 198                 rq->pages = (struct page *)p->private;
 199                 /* clear private here, it is used to chain pages */
 200                 p->private = 0;
 201         } else
 202                 p = alloc_page(gfp_mask);
 203         return p;
 204 }
 205
 206 static void skb_xmit_done(struct virtqueue *vq)
 207 {
 208         struct virtnet_info *vi = vq->vdev->priv;
 209
 210         /* Suppress further interrupts. */
 211         virtqueue_disable_cb(vq);
 212
 213         /* We were probably waiting for more output buffers. */
 214         netif_wake_subqueue(vi->dev, vq2txq(vq));
 215 }
 216
 217 static void set_skb_frag(struct sk_buff *skb, struct page *page,
 218                          unsigned int offset, unsigned int *len)
 219 {
 220         int size = min((unsigned)PAGE_SIZE - offset, *len);
 221         int i = skb_shinfo(skb)->nr_frags;
 222
 223         __skb_fill_page_desc(skb, i, page, offset, size);
 224
 225         skb->data_len += size;
 226         skb->len += size;
 227         skb->truesize += PAGE_SIZE;
 228         skb_shinfo(skb)->nr_frags++;
 229         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 230         *len -= size;
 231 }
 232
 233 /* Called from bottom half context */
 234 static struct sk_buff *page_to_skb(struct receive_queue *rq,
 235                                    struct page *page, unsigned int len)
 236 {
 237         struct virtnet_info *vi = rq->vq->vdev->priv;
 238         struct sk_buff *skb;
 239         struct skb_vnet_hdr *hdr;
 240         unsigned int copy, hdr_len, offset;
 241         char *p;
 242
 243         p = page_address(page);
 244
 245         /* copy small packet so we can reuse these pages for small data */
 246         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 247         if (unlikely(!skb))
 248                 return NULL;
 249
 250         hdr = skb_vnet_hdr(skb);
 251
 252         if (vi->mergeable_rx_bufs) {
 253                 hdr_len = sizeof hdr->mhdr;
 254                 offset = hdr_len;
 255         } else {
 256                 hdr_len = sizeof hdr->hdr;
 257                 offset = sizeof(struct padded_vnet_hdr);
 258         }
 259
 260         memcpy(hdr, p, hdr_len);
 261
 262         len -= hdr_len;
 263         p += offset;
 264
 265         copy = len;
 266         if (copy > skb_tailroom(skb))
 267                 copy = skb_tailroom(skb);
 268         memcpy(skb_put(skb, copy), p, copy);
 269
 270         len -= copy;
 271         offset += copy;
 272
 273         /*
 274          * Verify that we can indeed put this data into a skb.
 275          * This is here to handle cases when the device erroneously
 276          * tries to receive more than is possible. This is usually
 277          * the case of a broken device.
 278          */
 279         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 280                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 281                 dev_kfree_skb(skb);
 282                 return NULL;
 283         }
 284
 285         while (len) {
 286                 set_skb_frag(skb, page, offset, &len);
 287                 page = (struct page *)page->private;
 288                 offset = 0;
 289         }
 290
 291         if (page)
 292                 give_pages(rq, page);
 293
 294         return skb;
 295 }
 296
 297 static int receive_mergeable(struct receive_queue *rq, struct sk_buff *skb)
 298 {
 299         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 300         struct page *page;
 301         int num_buf, i, len;
 302
 303         num_buf = hdr->mhdr.num_buffers;
 304         while (--num_buf) {
 305                 i = skb_shinfo(skb)->nr_frags;
 306                 if (i >= MAX_SKB_FRAGS) {
 307                         pr_debug("%s: packet too long\n", skb->dev->name);
 308                         skb->dev->stats.rx_length_errors++;
 309                         return -EINVAL;
 310                 }
 311                 page = virtqueue_get_buf(rq->vq, &len);
 312                 if (!page) {
 313                         pr_debug("%s: rx error: %d buffers missing\n",
 314                                  skb->dev->name, hdr->mhdr.num_buffers);
 315                         skb->dev->stats.rx_length_errors++;
 316                         return -EINVAL;
 317                 }
 318
 319                 if (len > PAGE_SIZE)
 320                         len = PAGE_SIZE;
 321
 322                 set_skb_frag(skb, page, 0, &len);
 323
 324                 --rq->num;
 325         }
 326         return 0;
 327 }
 328
 329 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
 330 {
 331         struct virtnet_info *vi = rq->vq->vdev->priv;
 332         struct net_device *dev = vi->dev;
 333         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 334         struct sk_buff *skb;
 335         struct page *page;
 336         struct skb_vnet_hdr *hdr;
 337
 338         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 339                 pr_debug("%s: short packet %i\n", dev->name, len);
 340                 dev->stats.rx_length_errors++;
 341                 if (vi->mergeable_rx_bufs || vi->big_packets)
 342                         give_pages(rq, buf);
 343                 else
 344                         dev_kfree_skb(buf);
 345                 return;
 346         }
 347
 348         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
 349                 skb = buf;
 350                 len -= sizeof(struct virtio_net_hdr);
 351                 skb_trim(skb, len);
 352         } else {
 353                 page = buf;
 354                 skb = page_to_skb(rq, page, len);
 355                 if (unlikely(!skb)) {
 356                         dev->stats.rx_dropped++;
 357                         give_pages(rq, page);
 358                         return;
 359                 }
 360                 if (vi->mergeable_rx_bufs)
 361                         if (receive_mergeable(rq, skb)) {
 362                                 dev_kfree_skb(skb);
 363                                 return;
 364                         }
 365         }
 366
 367         hdr = skb_vnet_hdr(skb);
 368
 369         u64_stats_update_begin(&stats->rx_syncp);
 370         stats->rx_bytes += skb->len;
 371         stats->rx_packets++;
 372         u64_stats_update_end(&stats->rx_syncp);
 373
 374         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 375                 pr_debug("Needs csum!\n");
 376                 if (!skb_partial_csum_set(skb,
 377                                           hdr->hdr.csum_start,
 378                                           hdr->hdr.csum_offset))
 379                         goto frame_err;
 380         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 381                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 382         }
 383
 384         skb->protocol = eth_type_trans(skb, dev);
 385         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 386                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 387
 388         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 389                 pr_debug("GSO!\n");
 390                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 391                 case VIRTIO_NET_HDR_GSO_TCPV4:
 392                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 393                         break;
 394                 case VIRTIO_NET_HDR_GSO_UDP:
 395                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 396                         break;
 397                 case VIRTIO_NET_HDR_GSO_TCPV6:
 398                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 399                         break;
 400                 default:
 401                         net_warn_ratelimited("%s: bad gso type %u.\n",
 402                                              dev->name, hdr->hdr.gso_type);
 403                         goto frame_err;
 404                 }
 405
 406                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 407                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 408
 409                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 410                 if (skb_shinfo(skb)->gso_size == 0) {
 411                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 412                         goto frame_err;
 413                 }
 414
 415                 /* Header must be checked, and gso_segs computed. */
 416                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 417                 skb_shinfo(skb)->gso_segs = 0;
 418         }
 419
 420         netif_receive_skb(skb);
 421         return;
 422
 423 frame_err:
 424         dev->stats.rx_frame_errors++;
 425         dev_kfree_skb(skb);
 426 }
 427
 428 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
 429 {
 430         struct virtnet_info *vi = rq->vq->vdev->priv;
 431         struct sk_buff *skb;
 432         struct skb_vnet_hdr *hdr;
 433         int err;
 434
 435         skb = __netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN, gfp);
 436         if (unlikely(!skb))
 437                 return -ENOMEM;
 438
 439         skb_put(skb, MAX_PACKET_LEN);
 440
 441         hdr = skb_vnet_hdr(skb);
 442         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
 443
 444         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 445
 446         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 447         if (err < 0)
 448                 dev_kfree_skb(skb);
 449
 450         return err;
 451 }
 452
 453 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
 454 {
 455         struct page *first, *list = NULL;
 456         char *p;
 457         int i, err, offset;
 458
 459         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 460         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 461                 first = get_a_page(rq, gfp);
 462                 if (!first) {
 463                         if (list)
 464                                 give_pages(rq, list);
 465                         return -ENOMEM;
 466                 }
 467                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 468
 469                 /* chain new page in list head to match sg */
 470                 first->private = (unsigned long)list;
 471                 list = first;
 472         }
 473
 474         first = get_a_page(rq, gfp);
 475         if (!first) {
 476                 give_pages(rq, list);
 477                 return -ENOMEM;
 478         }
 479         p = page_address(first);
 480
 481         /* rq->sg[0], rq->sg[1] share the same page */
 482         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
 483         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
 484
 485         /* rq->sg[1] for data packet, from offset */
 486         offset = sizeof(struct padded_vnet_hdr);
 487         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 488
 489         /* chain first in list head */
 490         first->private = (unsigned long)list;
 491         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 492                                   first, gfp);
 493         if (err < 0)
 494                 give_pages(rq, first);
 495
 496         return err;
 497 }
 498
 499 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 500 {
 501         struct page *page;
 502         int err;
 503
 504         page = get_a_page(rq, gfp);
 505         if (!page)
 506                 return -ENOMEM;
 507
 508         sg_init_one(rq->sg, page_address(page), PAGE_SIZE);
 509
 510         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, page, gfp);
 511         if (err < 0)
 512                 give_pages(rq, page);
 513
 514         return err;
 515 }
 516
 517 /*
 518  * Returns false if we couldn't fill entirely (OOM).
 519  *
 520  * Normally run in the receive path, but can also be run from ndo_open
 521  * before we're receiving packets, or from refill_work which is
 522  * careful to disable receiving (using napi_disable).
 523  */
 524 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
 525 {
 526         struct virtnet_info *vi = rq->vq->vdev->priv;
 527         int err;
 528         bool oom;
 529
 530         do {
 531                 if (vi->mergeable_rx_bufs)
 532                         err = add_recvbuf_mergeable(rq, gfp);
 533                 else if (vi->big_packets)
 534                         err = add_recvbuf_big(rq, gfp);
 535                 else
 536                         err = add_recvbuf_small(rq, gfp);
 537
 538                 oom = err == -ENOMEM;
 539                 if (err)
 540                         break;
 541                 ++rq->num;
 542         } while (rq->vq->num_free);
 543         if (unlikely(rq->num > rq->max))
 544                 rq->max = rq->num;
 545         virtqueue_kick(rq->vq);
 546         return !oom;
 547 }
 548
 549 static void skb_recv_done(struct virtqueue *rvq)
 550 {
 551         struct virtnet_info *vi = rvq->vdev->priv;
 552         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 553
 554         /* Schedule NAPI, Suppress further interrupts if successful. */
 555         if (napi_schedule_prep(&rq->napi)) {
 556                 virtqueue_disable_cb(rvq);
 557                 __napi_schedule(&rq->napi);
 558         }
 559 }
 560
 561 static void virtnet_napi_enable(struct receive_queue *rq)
 562 {
 563         napi_enable(&rq->napi);
 564
 565         /* If all buffers were filled by other side before we napi_enabled, we
 566          * won't get another interrupt, so process any outstanding packets
 567          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 568          * We synchronize against interrupts via NAPI_STATE_SCHED */
 569         if (napi_schedule_prep(&rq->napi)) {
 570                 virtqueue_disable_cb(rq->vq);
 571                 local_bh_disable();
 572                 __napi_schedule(&rq->napi);
 573                 local_bh_enable();
 574         }
 575 }
 576
 577 static void refill_work(struct work_struct *work)
 578 {
 579         struct virtnet_info *vi =
 580                 container_of(work, struct virtnet_info, refill.work);
 581         bool still_empty;
 582         int i;
 583
 584         for (i = 0; i < vi->max_queue_pairs; i++) {
 585                 struct receive_queue *rq = &vi->rq[i];
 586
 587                 napi_disable(&rq->napi);
 588                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 589                 virtnet_napi_enable(rq);
 590
 591                 /* In theory, this can happen: if we don't get any buffers in
 592                  * we will *never* try to fill again.
 593                  */
 594                 if (still_empty)
 595                         schedule_delayed_work(&vi->refill, HZ/2);
 596         }
 597 }
 598
 599 static int virtnet_poll(struct napi_struct *napi, int budget)
 600 {
 601         struct receive_queue *rq =
 602                 container_of(napi, struct receive_queue, napi);
 603         struct virtnet_info *vi = rq->vq->vdev->priv;
 604         void *buf;
 605         unsigned int len, received = 0;
 606
 607 again:
 608         while (received < budget &&
 609                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 610                 receive_buf(rq, buf, len);
 611                 --rq->num;
 612                 received++;
 613         }
 614
 615         if (rq->num < rq->max / 2) {
 616                 if (!try_fill_recv(rq, GFP_ATOMIC))
 617                         schedule_delayed_work(&vi->refill, 0);
 618         }
 619
 620         /* Out of packets? */
 621         if (received < budget) {
 622                 napi_complete(napi);
 623                 if (unlikely(!virtqueue_enable_cb(rq->vq)) &&
 624                     napi_schedule_prep(napi)) {
 625                         virtqueue_disable_cb(rq->vq);
 626                         __napi_schedule(napi);
 627                         goto again;
 628                 }
 629         }
 630
 631         return received;
 632 }
 633
 634 static int virtnet_open(struct net_device *dev)
 635 {
 636         struct virtnet_info *vi = netdev_priv(dev);
 637         int i;
 638
 639         for (i = 0; i < vi->max_queue_pairs; i++) {
 640                 /* Make sure we have some buffers: if oom use wq. */
 641                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 642                         schedule_delayed_work(&vi->refill, 0);
 643                 virtnet_napi_enable(&vi->rq[i]);
 644         }
 645
 646         return 0;
 647 }
 648
 649 static void free_old_xmit_skbs(struct send_queue *sq)
 650 {
 651         struct sk_buff *skb;
 652         unsigned int len;
 653         struct virtnet_info *vi = sq->vq->vdev->priv;
 654         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 655
 656         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 657                 pr_debug("Sent skb %p\n", skb);
 658
 659                 u64_stats_update_begin(&stats->tx_syncp);
 660                 stats->tx_bytes += skb->len;
 661                 stats->tx_packets++;
 662                 u64_stats_update_end(&stats->tx_syncp);
 663
 664                 dev_kfree_skb_any(skb);
 665         }
 666 }
 667
 668 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 669 {
 670         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 671         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 672         struct virtnet_info *vi = sq->vq->vdev->priv;
 673         unsigned num_sg;
 674
 675         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 676
 677         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 678                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 679                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 680                 hdr->hdr.csum_offset = skb->csum_offset;
 681         } else {
 682                 hdr->hdr.flags = 0;
 683                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 684         }
 685
 686         if (skb_is_gso(skb)) {
 687                 hdr->hdr.hdr_len = skb_headlen(skb);
 688                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 689                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 690                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 691                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 692                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 693                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 694                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 695                 else
 696                         BUG();
 697                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 698                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 699         } else {
 700                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 701                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 702         }
 703
 704         hdr->mhdr.num_buffers = 0;
 705
 706         /* Encode metadata header at front. */
 707         if (vi->mergeable_rx_bufs)
 708                 sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
 709         else
 710                 sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
 711
 712         num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 713         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 714 }
 715
 716 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 717 {
 718         struct virtnet_info *vi = netdev_priv(dev);
 719         int qnum = skb_get_queue_mapping(skb);
 720         struct send_queue *sq = &vi->sq[qnum];
 721         int err;
 722
 723         /* Free up any pending old buffers before queueing new ones. */
 724         free_old_xmit_skbs(sq);
 725
 726         /* Try to transmit */
 727         err = xmit_skb(sq, skb);
 728
 729         /* This should not happen! */
 730         if (unlikely(err)) {
 731                 dev->stats.tx_fifo_errors++;
 732                 if (net_ratelimit())
 733                         dev_warn(&dev->dev,
 734                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 735                 dev->stats.tx_dropped++;
 736                 kfree_skb(skb);
 737                 return NETDEV_TX_OK;
 738         }
 739         virtqueue_kick(sq->vq);
 740
 741         /* Don't wait up for transmitted skbs to be freed. */
 742         skb_orphan(skb);
 743         nf_reset(skb);
 744
 745         /* Apparently nice girls don't return TX_BUSY; stop the queue
 746          * before it gets out of hand.  Naturally, this wastes entries. */
 747         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 748                 netif_stop_subqueue(dev, qnum);
 749                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 750                         /* More just got used, free them then recheck. */
 751                         free_old_xmit_skbs(sq);
 752                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 753                                 netif_start_subqueue(dev, qnum);
 754                                 virtqueue_disable_cb(sq->vq);
 755                         }
 756                 }
 757         }
 758
 759         return NETDEV_TX_OK;
 760 }
 761
 762 /*
 763  * Send command via the control virtqueue and check status.  Commands
 764  * supported by the hypervisor, as indicated by feature bits, should
 765  * never fail unless improperly formated.
 766  */
 767 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 768                                  struct scatterlist *out,
 769                                  struct scatterlist *in)
 770 {
 771         struct scatterlist *sgs[4], hdr, stat;
 772         struct virtio_net_ctrl_hdr ctrl;
 773         virtio_net_ctrl_ack status = ~0;
 774         unsigned out_num = 0, in_num = 0, tmp;
 775
 776         /* Caller should know better */
 777         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 778
 779         ctrl.class = class;
 780         ctrl.cmd = cmd;
 781         /* Add header */
 782         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 783         sgs[out_num++] = &hdr;
 784
 785         if (out)
 786                 sgs[out_num++] = out;
 787         if (in)
 788                 sgs[out_num + in_num++] = in;
 789
 790         /* Add return status. */
 791         sg_init_one(&stat, &status, sizeof(status));
 792         sgs[out_num + in_num++] = &stat;
 793
 794         BUG_ON(out_num + in_num > ARRAY_SIZE(sgs));
 795         BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, in_num, vi, GFP_ATOMIC)
 796                < 0);
 797
 798         virtqueue_kick(vi->cvq);
 799
 800         /* Spin for a response, the kick causes an ioport write, trapping
 801          * into the hypervisor, so the request should be handled immediately.
 802          */
 803         while (!virtqueue_get_buf(vi->cvq, &tmp))
 804                 cpu_relax();
 805
 806         return status == VIRTIO_NET_OK;
 807 }
 808
 809 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 810 {
 811         struct virtnet_info *vi = netdev_priv(dev);
 812         struct virtio_device *vdev = vi->vdev;
 813         int ret;
 814         struct sockaddr *addr = p;
 815         struct scatterlist sg;
 816
 817         ret = eth_prepare_mac_addr_change(dev, p);
 818         if (ret)
 819                 return ret;
 820
 821         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 822                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 823                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 824                                           VIRTIO_NET_CTRL_MAC_ADDR_SET,
 825                                           &sg, NULL)) {
 826                         dev_warn(&vdev->dev,
 827                                  "Failed to set mac address by vq command.\n");
 828                         return -EINVAL;
 829                 }
 830         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 831                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 832                                   addr->sa_data, dev->addr_len);
 833         }
 834
 835         eth_commit_mac_addr_change(dev, p);
 836
 837         return 0;
 838 }
 839
 840 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 841                                                struct rtnl_link_stats64 *tot)
 842 {
 843         struct virtnet_info *vi = netdev_priv(dev);
 844         int cpu;
 845         unsigned int start;
 846
 847         for_each_possible_cpu(cpu) {
 848                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 849                 u64 tpackets, tbytes, rpackets, rbytes;
 850
 851                 do {
 852                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 853                         tpackets = stats->tx_packets;
 854                         tbytes   = stats->tx_bytes;
 855                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 856
 857                 do {
 858                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 859                         rpackets = stats->rx_packets;
 860                         rbytes   = stats->rx_bytes;
 861                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 862
 863                 tot->rx_packets += rpackets;
 864                 tot->tx_packets += tpackets;
 865                 tot->rx_bytes   += rbytes;
 866                 tot->tx_bytes   += tbytes;
 867         }
 868
 869         tot->tx_dropped = dev->stats.tx_dropped;
 870         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 871         tot->rx_dropped = dev->stats.rx_dropped;
 872         tot->rx_length_errors = dev->stats.rx_length_errors;
 873         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 874
 875         return tot;
 876 }
 877
 878 #ifdef CONFIG_NET_POLL_CONTROLLER
 879 static void virtnet_netpoll(struct net_device *dev)
 880 {
 881         struct virtnet_info *vi = netdev_priv(dev);
 882         int i;
 883
 884         for (i = 0; i < vi->curr_queue_pairs; i++)
 885                 napi_schedule(&vi->rq[i].napi);
 886 }
 887 #endif
 888
 889 static void virtnet_ack_link_announce(struct virtnet_info *vi)
 890 {
 891         rtnl_lock();
 892         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
 893                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
 894                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
 895         rtnl_unlock();
 896 }
 897
 898 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
 899 {
 900         struct scatterlist sg;
 901         struct virtio_net_ctrl_mq s;
 902         struct net_device *dev = vi->dev;
 903
 904         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
 905                 return 0;
 906
 907         s.virtqueue_pairs = queue_pairs;
 908         sg_init_one(&sg, &s, sizeof(s));
 909
 910         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
 911                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
 912                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
 913                          queue_pairs);
 914                 return -EINVAL;
 915         } else
 916                 vi->curr_queue_pairs = queue_pairs;
 917
 918         return 0;
 919 }
 920
 921 static int virtnet_close(struct net_device *dev)
 922 {
 923         struct virtnet_info *vi = netdev_priv(dev);
 924         int i;
 925
 926         /* Make sure refill_work doesn't re-enable napi! */
 927         cancel_delayed_work_sync(&vi->refill);
 928
 929         for (i = 0; i < vi->max_queue_pairs; i++)
 930                 napi_disable(&vi->rq[i].napi);
 931
 932         return 0;
 933 }
 934
 935 static void virtnet_set_rx_mode(struct net_device *dev)
 936 {
 937         struct virtnet_info *vi = netdev_priv(dev);
 938         struct scatterlist sg[2];
 939         u8 promisc, allmulti;
 940         struct virtio_net_ctrl_mac *mac_data;
 941         struct netdev_hw_addr *ha;
 942         int uc_count;
 943         int mc_count;
 944         void *buf;
 945         int i;
 946
 947         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 948         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 949                 return;
 950
 951         promisc = ((dev->flags & IFF_PROMISC) != 0);
 952         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 953
 954         sg_init_one(sg, &promisc, sizeof(promisc));
 955
 956         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 957                                   VIRTIO_NET_CTRL_RX_PROMISC,
 958                                   sg, NULL))
 959                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 960                          promisc ? "en" : "dis");
 961
 962         sg_init_one(sg, &allmulti, sizeof(allmulti));
 963
 964         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 965                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 966                                   sg, NULL))
 967                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
 968                          allmulti ? "en" : "dis");
 969
 970         uc_count = netdev_uc_count(dev);
 971         mc_count = netdev_mc_count(dev);
 972         /* MAC filter - use one buffer for both lists */
 973         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
 974                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
 975         mac_data = buf;
 976         if (!buf)
 977                 return;
 978
 979         sg_init_table(sg, 2);
 980
 981         /* Store the unicast list and count in the front of the buffer */
 982         mac_data->entries = uc_count;
 983         i = 0;
 984         netdev_for_each_uc_addr(ha, dev)
 985                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 986
 987         sg_set_buf(&sg[0], mac_data,
 988                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
 989
 990         /* multicast list and count fill the end */
 991         mac_data = (void *)&mac_data->macs[uc_count][0];
 992
 993         mac_data->entries = mc_count;
 994         i = 0;
 995         netdev_for_each_mc_addr(ha, dev)
 996                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 997
 998         sg_set_buf(&sg[1], mac_data,
 999                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1000
1001         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1002                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
1003                                   sg, NULL))
1004                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1005
1006         kfree(buf);
1007 }
1008
1009 static int virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
1010 {
1011         struct virtnet_info *vi = netdev_priv(dev);
1012         struct scatterlist sg;
1013
1014         sg_init_one(&sg, &vid, sizeof(vid));
1015
1016         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1017                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1018                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1019         return 0;
1020 }
1021
1022 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
1023 {
1024         struct virtnet_info *vi = netdev_priv(dev);
1025         struct scatterlist sg;
1026
1027         sg_init_one(&sg, &vid, sizeof(vid));
1028
1029         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1030                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1031                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1032         return 0;
1033 }
1034
1035 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1036 {
1037         int i;
1038         int cpu;
1039
1040         if (vi->affinity_hint_set) {
1041                 for (i = 0; i < vi->max_queue_pairs; i++) {
1042                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1043                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1044                 }
1045
1046                 vi->affinity_hint_set = false;
1047         }
1048
1049         i = 0;
1050         for_each_online_cpu(cpu) {
1051                 if (cpu == hcpu) {
1052                         *per_cpu_ptr(vi->vq_index, cpu) = -1;
1053                 } else {
1054                         *per_cpu_ptr(vi->vq_index, cpu) =
1055                                 ++i % vi->curr_queue_pairs;
1056                 }
1057         }
1058 }
1059
1060 static void virtnet_set_affinity(struct virtnet_info *vi)
1061 {
1062         int i;
1063         int cpu;
1064
1065         /* In multiqueue mode, when the number of cpu is equal to the number of
1066          * queue pairs, we let the queue pairs to be private to one cpu by
1067          * setting the affinity hint to eliminate the contention.
1068          */
1069         if (vi->curr_queue_pairs == 1 ||
1070             vi->max_queue_pairs != num_online_cpus()) {
1071                 virtnet_clean_affinity(vi, -1);
1072                 return;
1073         }
1074
1075         i = 0;
1076         for_each_online_cpu(cpu) {
1077                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1078                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1079                 *per_cpu_ptr(vi->vq_index, cpu) = i;
1080                 i++;
1081         }
1082
1083         vi->affinity_hint_set = true;
1084 }
1085
1086 static int virtnet_cpu_callback(struct notifier_block *nfb,
1087                                 unsigned long action, void *hcpu)
1088 {
1089         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1090
1091         switch(action & ~CPU_TASKS_FROZEN) {
1092         case CPU_ONLINE:
1093         case CPU_DOWN_FAILED:
1094         case CPU_DEAD:
1095                 virtnet_set_affinity(vi);
1096                 break;
1097         case CPU_DOWN_PREPARE:
1098                 virtnet_clean_affinity(vi, (long)hcpu);
1099                 break;
1100         default:
1101                 break;
1102         }
1103         return NOTIFY_OK;
1104 }
1105
1106 static void virtnet_get_ringparam(struct net_device *dev,
1107                                 struct ethtool_ringparam *ring)
1108 {
1109         struct virtnet_info *vi = netdev_priv(dev);
1110
1111         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1112         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1113         ring->rx_pending = ring->rx_max_pending;
1114         ring->tx_pending = ring->tx_max_pending;
1115 }
1116
1117
1118 static void virtnet_get_drvinfo(struct net_device *dev,
1119                                 struct ethtool_drvinfo *info)
1120 {
1121         struct virtnet_info *vi = netdev_priv(dev);
1122         struct virtio_device *vdev = vi->vdev;
1123
1124         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1125         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1126         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1127
1128 }
1129
1130 /* TODO: Eliminate OOO packets during switching */
1131 static int virtnet_set_channels(struct net_device *dev,
1132                                 struct ethtool_channels *channels)
1133 {
1134         struct virtnet_info *vi = netdev_priv(dev);
1135         u16 queue_pairs = channels->combined_count;
1136         int err;
1137
1138         /* We don't support separate rx/tx channels.
1139          * We don't allow setting 'other' channels.
1140          */
1141         if (channels->rx_count || channels->tx_count || channels->other_count)
1142                 return -EINVAL;
1143
1144         if (queue_pairs > vi->max_queue_pairs)
1145                 return -EINVAL;
1146
1147         get_online_cpus();
1148         err = virtnet_set_queues(vi, queue_pairs);
1149         if (!err) {
1150                 netif_set_real_num_tx_queues(dev, queue_pairs);
1151                 netif_set_real_num_rx_queues(dev, queue_pairs);
1152
1153                 virtnet_set_affinity(vi);
1154         }
1155         put_online_cpus();
1156
1157         return err;
1158 }
1159
1160 static void virtnet_get_channels(struct net_device *dev,
1161                                  struct ethtool_channels *channels)
1162 {
1163         struct virtnet_info *vi = netdev_priv(dev);
1164
1165         channels->combined_count = vi->curr_queue_pairs;
1166         channels->max_combined = vi->max_queue_pairs;
1167         channels->max_other = 0;
1168         channels->rx_count = 0;
1169         channels->tx_count = 0;
1170         channels->other_count = 0;
1171 }
1172
1173 static const struct ethtool_ops virtnet_ethtool_ops = {
1174         .get_drvinfo = virtnet_get_drvinfo,
1175         .get_link = ethtool_op_get_link,
1176         .get_ringparam = virtnet_get_ringparam,
1177         .set_channels = virtnet_set_channels,
1178         .get_channels = virtnet_get_channels,
1179 };
1180
1181 #define MIN_MTU 68
1182 #define MAX_MTU 65535
1183
1184 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1185 {
1186         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1187                 return -EINVAL;
1188         dev->mtu = new_mtu;
1189         return 0;
1190 }
1191
1192 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
1193  * txq based on the processor id.
1194  */
1195 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
1196 {
1197         int txq;
1198         struct virtnet_info *vi = netdev_priv(dev);
1199
1200         if (skb_rx_queue_recorded(skb)) {
1201                 txq = skb_get_rx_queue(skb);
1202         } else {
1203                 txq = *__this_cpu_ptr(vi->vq_index);
1204                 if (txq == -1)
1205                         txq = 0;
1206         }
1207
1208         while (unlikely(txq >= dev->real_num_tx_queues))
1209                 txq -= dev->real_num_tx_queues;
1210
1211         return txq;
1212 }
1213
1214 static const struct net_device_ops virtnet_netdev = {
1215         .ndo_open            = virtnet_open,
1216         .ndo_stop            = virtnet_close,
1217         .ndo_start_xmit      = start_xmit,
1218         .ndo_validate_addr   = eth_validate_addr,
1219         .ndo_set_mac_address = virtnet_set_mac_address,
1220         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1221         .ndo_change_mtu      = virtnet_change_mtu,
1222         .ndo_get_stats64     = virtnet_stats,
1223         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1224         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1225         .ndo_select_queue     = virtnet_select_queue,
1226 #ifdef CONFIG_NET_POLL_CONTROLLER
1227         .ndo_poll_controller = virtnet_netpoll,
1228 #endif
1229 };
1230
1231 static void virtnet_config_changed_work(struct work_struct *work)
1232 {
1233         struct virtnet_info *vi =
1234                 container_of(work, struct virtnet_info, config_work);
1235         u16 v;
1236
1237         mutex_lock(&vi->config_lock);
1238         if (!vi->config_enable)
1239                 goto done;
1240
1241         if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1242                               offsetof(struct virtio_net_config, status),
1243                               &v) < 0)
1244                 goto done;
1245
1246         if (v & VIRTIO_NET_S_ANNOUNCE) {
1247                 netdev_notify_peers(vi->dev);
1248                 virtnet_ack_link_announce(vi);
1249         }
1250
1251         /* Ignore unknown (future) status bits */
1252         v &= VIRTIO_NET_S_LINK_UP;
1253
1254         if (vi->status == v)
1255                 goto done;
1256
1257         vi->status = v;
1258
1259         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1260                 netif_carrier_on(vi->dev);
1261                 netif_tx_wake_all_queues(vi->dev);
1262         } else {
1263                 netif_carrier_off(vi->dev);
1264                 netif_tx_stop_all_queues(vi->dev);
1265         }
1266 done:
1267         mutex_unlock(&vi->config_lock);
1268 }
1269
1270 static void virtnet_config_changed(struct virtio_device *vdev)
1271 {
1272         struct virtnet_info *vi = vdev->priv;
1273
1274         schedule_work(&vi->config_work);
1275 }
1276
1277 static void virtnet_free_queues(struct virtnet_info *vi)
1278 {
1279         kfree(vi->rq);
1280         kfree(vi->sq);
1281 }
1282
1283 static void free_receive_bufs(struct virtnet_info *vi)
1284 {
1285         int i;
1286
1287         for (i = 0; i < vi->max_queue_pairs; i++) {
1288                 while (vi->rq[i].pages)
1289                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1290         }
1291 }
1292
1293 static void free_unused_bufs(struct virtnet_info *vi)
1294 {
1295         void *buf;
1296         int i;
1297
1298         for (i = 0; i < vi->max_queue_pairs; i++) {
1299                 struct virtqueue *vq = vi->sq[i].vq;
1300                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1301                         dev_kfree_skb(buf);
1302         }
1303
1304         for (i = 0; i < vi->max_queue_pairs; i++) {
1305                 struct virtqueue *vq = vi->rq[i].vq;
1306
1307                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1308                         if (vi->mergeable_rx_bufs || vi->big_packets)
1309                                 give_pages(&vi->rq[i], buf);
1310                         else
1311                                 dev_kfree_skb(buf);
1312                         --vi->rq[i].num;
1313                 }
1314                 BUG_ON(vi->rq[i].num != 0);
1315         }
1316 }
1317
1318 static void virtnet_del_vqs(struct virtnet_info *vi)
1319 {
1320         struct virtio_device *vdev = vi->vdev;
1321
1322         virtnet_clean_affinity(vi, -1);
1323
1324         vdev->config->del_vqs(vdev);
1325
1326         virtnet_free_queues(vi);
1327 }
1328
1329 static int virtnet_find_vqs(struct virtnet_info *vi)
1330 {
1331         vq_callback_t **callbacks;
1332         struct virtqueue **vqs;
1333         int ret = -ENOMEM;
1334         int i, total_vqs;
1335         const char **names;
1336
1337         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1338          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1339          * possible control vq.
1340          */
1341         total_vqs = vi->max_queue_pairs * 2 +
1342                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1343
1344         /* Allocate space for find_vqs parameters */
1345         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1346         if (!vqs)
1347                 goto err_vq;
1348         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1349         if (!callbacks)
1350                 goto err_callback;
1351         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1352         if (!names)
1353                 goto err_names;
1354
1355         /* Parameters for control virtqueue, if any */
1356         if (vi->has_cvq) {
1357                 callbacks[total_vqs - 1] = NULL;
1358                 names[total_vqs - 1] = "control";
1359         }
1360
1361         /* Allocate/initialize parameters for send/receive virtqueues */
1362         for (i = 0; i < vi->max_queue_pairs; i++) {
1363                 callbacks[rxq2vq(i)] = skb_recv_done;
1364                 callbacks[txq2vq(i)] = skb_xmit_done;
1365                 sprintf(vi->rq[i].name, "input.%d", i);
1366                 sprintf(vi->sq[i].name, "output.%d", i);
1367                 names[rxq2vq(i)] = vi->rq[i].name;
1368                 names[txq2vq(i)] = vi->sq[i].name;
1369         }
1370
1371         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1372                                          names);
1373         if (ret)
1374                 goto err_find;
1375
1376         if (vi->has_cvq) {
1377                 vi->cvq = vqs[total_vqs - 1];
1378                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1379                         vi->dev->features |= NETIF_F_HW_VLAN_FILTER;
1380         }
1381
1382         for (i = 0; i < vi->max_queue_pairs; i++) {
1383                 vi->rq[i].vq = vqs[rxq2vq(i)];
1384                 vi->sq[i].vq = vqs[txq2vq(i)];
1385         }
1386
1387         kfree(names);
1388         kfree(callbacks);
1389         kfree(vqs);
1390
1391         return 0;
1392
1393 err_find:
1394         kfree(names);
1395 err_names:
1396         kfree(callbacks);
1397 err_callback:
1398         kfree(vqs);
1399 err_vq:
1400         return ret;
1401 }
1402
1403 static int virtnet_alloc_queues(struct virtnet_info *vi)
1404 {
1405         int i;
1406
1407         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1408         if (!vi->sq)
1409                 goto err_sq;
1410         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1411         if (!vi->rq)
1412                 goto err_rq;
1413
1414         INIT_DELAYED_WORK(&vi->refill, refill_work);
1415         for (i = 0; i < vi->max_queue_pairs; i++) {
1416                 vi->rq[i].pages = NULL;
1417                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1418                                napi_weight);
1419
1420                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1421                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1422         }
1423
1424         return 0;
1425
1426 err_rq:
1427         kfree(vi->sq);
1428 err_sq:
1429         return -ENOMEM;
1430 }
1431
1432 static int init_vqs(struct virtnet_info *vi)
1433 {
1434         int ret;
1435
1436         /* Allocate send & receive queues */
1437         ret = virtnet_alloc_queues(vi);
1438         if (ret)
1439                 goto err;
1440
1441         ret = virtnet_find_vqs(vi);
1442         if (ret)
1443                 goto err_free;
1444
1445         get_online_cpus();
1446         virtnet_set_affinity(vi);
1447         put_online_cpus();
1448
1449         return 0;
1450
1451 err_free:
1452         virtnet_free_queues(vi);
1453 err:
1454         return ret;
1455 }
1456
1457 static int virtnet_probe(struct virtio_device *vdev)
1458 {
1459         int i, err;
1460         struct net_device *dev;
1461         struct virtnet_info *vi;
1462         u16 max_queue_pairs;
1463
1464         /* Find if host supports multiqueue virtio_net device */
1465         err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1466                                 offsetof(struct virtio_net_config,
1467                                 max_virtqueue_pairs), &max_queue_pairs);
1468
1469         /* We need at least 2 queue's */
1470         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1471             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1472             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1473                 max_queue_pairs = 1;
1474
1475         /* Allocate ourselves a network device with room for our info */
1476         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1477         if (!dev)
1478                 return -ENOMEM;
1479
1480         /* Set up network device as normal. */
1481         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1482         dev->netdev_ops = &virtnet_netdev;
1483         dev->features = NETIF_F_HIGHDMA;
1484
1485         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1486         SET_NETDEV_DEV(dev, &vdev->dev);
1487
1488         /* Do we support "hardware" checksums? */
1489         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1490                 /* This opens up the world of extra features. */
1491                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1492                 if (csum)
1493                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1494
1495                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1496                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1497                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1498                 }
1499                 /* Individual feature bits: what can host handle? */
1500                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1501                         dev->hw_features |= NETIF_F_TSO;
1502                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1503                         dev->hw_features |= NETIF_F_TSO6;
1504                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1505                         dev->hw_features |= NETIF_F_TSO_ECN;
1506                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1507                         dev->hw_features |= NETIF_F_UFO;
1508
1509                 if (gso)
1510                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1511                 /* (!csum && gso) case will be fixed by register_netdev() */
1512         }
1513
1514         /* Configuration may specify what MAC to use.  Otherwise random. */
1515         if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1516                                   offsetof(struct virtio_net_config, mac),
1517                                   dev->dev_addr, dev->addr_len) < 0)
1518                 eth_hw_addr_random(dev);
1519
1520         /* Set up our device-specific information */
1521         vi = netdev_priv(dev);
1522         vi->dev = dev;
1523         vi->vdev = vdev;
1524         vdev->priv = vi;
1525         vi->stats = alloc_percpu(struct virtnet_stats);
1526         err = -ENOMEM;
1527         if (vi->stats == NULL)
1528                 goto free;
1529
1530         vi->vq_index = alloc_percpu(int);
1531         if (vi->vq_index == NULL)
1532                 goto free_stats;
1533
1534         mutex_init(&vi->config_lock);
1535         vi->config_enable = true;
1536         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1537
1538         /* If we can receive ANY GSO packets, we must allocate large ones. */
1539         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1540             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1541             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1542                 vi->big_packets = true;
1543
1544         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1545                 vi->mergeable_rx_bufs = true;
1546
1547         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1548                 vi->has_cvq = true;
1549
1550         /* Use single tx/rx queue pair as default */
1551         vi->curr_queue_pairs = 1;
1552         vi->max_queue_pairs = max_queue_pairs;
1553
1554         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1555         err = init_vqs(vi);
1556         if (err)
1557                 goto free_index;
1558
1559         netif_set_real_num_tx_queues(dev, 1);
1560         netif_set_real_num_rx_queues(dev, 1);
1561
1562         err = register_netdev(dev);
1563         if (err) {
1564                 pr_debug("virtio_net: registering device failed\n");
1565                 goto free_vqs;
1566         }
1567
1568         /* Last of all, set up some receive buffers. */
1569         for (i = 0; i < vi->max_queue_pairs; i++) {
1570                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1571
1572                 /* If we didn't even get one input buffer, we're useless. */
1573                 if (vi->rq[i].num == 0) {
1574                         free_unused_bufs(vi);
1575                         err = -ENOMEM;
1576                         goto free_recv_bufs;
1577                 }
1578         }
1579
1580         vi->nb.notifier_call = &virtnet_cpu_callback;
1581         err = register_hotcpu_notifier(&vi->nb);
1582         if (err) {
1583                 pr_debug("virtio_net: registering cpu notifier failed\n");
1584                 goto free_recv_bufs;
1585         }
1586
1587         /* Assume link up if device can't report link status,
1588            otherwise get link status from config. */
1589         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1590                 netif_carrier_off(dev);
1591                 schedule_work(&vi->config_work);
1592         } else {
1593                 vi->status = VIRTIO_NET_S_LINK_UP;
1594                 netif_carrier_on(dev);
1595         }
1596
1597         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1598                  dev->name, max_queue_pairs);
1599
1600         return 0;
1601
1602 free_recv_bufs:
1603         free_receive_bufs(vi);
1604         unregister_netdev(dev);
1605 free_vqs:
1606         cancel_delayed_work_sync(&vi->refill);
1607         virtnet_del_vqs(vi);
1608 free_index:
1609         free_percpu(vi->vq_index);
1610 free_stats:
1611         free_percpu(vi->stats);
1612 free:
1613         free_netdev(dev);
1614         return err;
1615 }
1616
1617 static void remove_vq_common(struct virtnet_info *vi)
1618 {
1619         vi->vdev->config->reset(vi->vdev);
1620
1621         /* Free unused buffers in both send and recv, if any. */
1622         free_unused_bufs(vi);
1623
1624         free_receive_bufs(vi);
1625
1626         virtnet_del_vqs(vi);
1627 }
1628
1629 static void virtnet_remove(struct virtio_device *vdev)
1630 {
1631         struct virtnet_info *vi = vdev->priv;
1632
1633         unregister_hotcpu_notifier(&vi->nb);
1634
1635         /* Prevent config work handler from accessing the device. */
1636         mutex_lock(&vi->config_lock);
1637         vi->config_enable = false;
1638         mutex_unlock(&vi->config_lock);
1639
1640         unregister_netdev(vi->dev);
1641
1642         remove_vq_common(vi);
1643
1644         flush_work(&vi->config_work);
1645
1646         free_percpu(vi->vq_index);
1647         free_percpu(vi->stats);
1648         free_netdev(vi->dev);
1649 }
1650
1651 #ifdef CONFIG_PM
1652 static int virtnet_freeze(struct virtio_device *vdev)
1653 {
1654         struct virtnet_info *vi = vdev->priv;
1655         int i;
1656
1657         /* Prevent config work handler from accessing the device */
1658         mutex_lock(&vi->config_lock);
1659         vi->config_enable = false;
1660         mutex_unlock(&vi->config_lock);
1661
1662         netif_device_detach(vi->dev);
1663         cancel_delayed_work_sync(&vi->refill);
1664
1665         if (netif_running(vi->dev))
1666                 for (i = 0; i < vi->max_queue_pairs; i++) {
1667                         napi_disable(&vi->rq[i].napi);
1668                         netif_napi_del(&vi->rq[i].napi);
1669                 }
1670
1671         remove_vq_common(vi);
1672
1673         flush_work(&vi->config_work);
1674
1675         return 0;
1676 }
1677
1678 static int virtnet_restore(struct virtio_device *vdev)
1679 {
1680         struct virtnet_info *vi = vdev->priv;
1681         int err, i;
1682
1683         err = init_vqs(vi);
1684         if (err)
1685                 return err;
1686
1687         if (netif_running(vi->dev))
1688                 for (i = 0; i < vi->max_queue_pairs; i++)
1689                         virtnet_napi_enable(&vi->rq[i]);
1690
1691         netif_device_attach(vi->dev);
1692
1693         for (i = 0; i < vi->max_queue_pairs; i++)
1694                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1695                         schedule_delayed_work(&vi->refill, 0);
1696
1697         mutex_lock(&vi->config_lock);
1698         vi->config_enable = true;
1699         mutex_unlock(&vi->config_lock);
1700
1701         virtnet_set_queues(vi, vi->curr_queue_pairs);
1702
1703         return 0;
1704 }
1705 #endif
1706
1707 static struct virtio_device_id id_table[] = {
1708         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1709         { 0 },
1710 };
1711
1712 static unsigned int features[] = {
1713         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1714         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1715         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1716         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1717         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1718         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1719         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1720         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1721         VIRTIO_NET_F_CTRL_MAC_ADDR,
1722 };
1723
1724 static struct virtio_driver virtio_net_driver = {
1725         .feature_table = features,
1726         .feature_table_size = ARRAY_SIZE(features),
1727         .driver.name =  KBUILD_MODNAME,
1728         .driver.owner = THIS_MODULE,
1729         .id_table =     id_table,
1730         .probe =        virtnet_probe,
1731         .remove =       virtnet_remove,
1732         .config_changed = virtnet_config_changed,
1733 #ifdef CONFIG_PM
1734         .freeze =       virtnet_freeze,
1735         .restore =      virtnet_restore,
1736 #endif
1737 };
1738
1739 module_virtio_driver(virtio_net_driver);
1740
1741 MODULE_DEVICE_TABLE(virtio, id_table);
1742 MODULE_DESCRIPTION("Virtio network driver");
1743 MODULE_LICENSE("GPL");