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 = NAPI_POLL_WEIGHT;
  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         /* Host can handle any s/g split between our header and packet data */
 110         bool any_header_sg;
 111
 112         /* enable config space updates */
 113         bool config_enable;
 114
 115         /* Active statistics */
 116         struct virtnet_stats __percpu *stats;
 117
 118         /* Work struct for refilling if we run low on memory. */
 119         struct delayed_work refill;
 120
 121         /* Work struct for config space updates */
 122         struct work_struct config_work;
 123
 124         /* Lock for config space updates */
 125         struct mutex config_lock;
 126
 127         /* Does the affinity hint is set for virtqueues? */
 128         bool affinity_hint_set;
 129
 130         /* Per-cpu variable to show the mapping from CPU to virtqueue */
 131         int __percpu *vq_index;
 132
 133         /* CPU hot plug notifier */
 134         struct notifier_block nb;
 135 };
 136
 137 struct skb_vnet_hdr {
 138         union {
 139                 struct virtio_net_hdr hdr;
 140                 struct virtio_net_hdr_mrg_rxbuf mhdr;
 141         };
 142 };
 143
 144 struct padded_vnet_hdr {
 145         struct virtio_net_hdr hdr;
 146         /*
 147          * virtio_net_hdr should be in a separated sg buffer because of a
 148          * QEMU bug, and data sg buffer shares same page with this header sg.
 149          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
 150          */
 151         char padding[6];
 152 };
 153
 154 /* Converting between virtqueue no. and kernel tx/rx queue no.
 155  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 156  */
 157 static int vq2txq(struct virtqueue *vq)
 158 {
 159         return (vq->index - 1) / 2;
 160 }
 161
 162 static int txq2vq(int txq)
 163 {
 164         return txq * 2 + 1;
 165 }
 166
 167 static int vq2rxq(struct virtqueue *vq)
 168 {
 169         return vq->index / 2;
 170 }
 171
 172 static int rxq2vq(int rxq)
 173 {
 174         return rxq * 2;
 175 }
 176
 177 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 178 {
 179         return (struct skb_vnet_hdr *)skb->cb;
 180 }
 181
 182 /*
 183  * private is used to chain pages for big packets, put the whole
 184  * most recent used list in the beginning for reuse
 185  */
 186 static void give_pages(struct receive_queue *rq, struct page *page)
 187 {
 188         struct page *end;
 189
 190         /* Find end of list, sew whole thing into vi->rq.pages. */
 191         for (end = page; end->private; end = (struct page *)end->private);
 192         end->private = (unsigned long)rq->pages;
 193         rq->pages = page;
 194 }
 195
 196 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 197 {
 198         struct page *p = rq->pages;
 199
 200         if (p) {
 201                 rq->pages = (struct page *)p->private;
 202                 /* clear private here, it is used to chain pages */
 203                 p->private = 0;
 204         } else
 205                 p = alloc_page(gfp_mask);
 206         return p;
 207 }
 208
 209 static void skb_xmit_done(struct virtqueue *vq)
 210 {
 211         struct virtnet_info *vi = vq->vdev->priv;
 212
 213         /* Suppress further interrupts. */
 214         virtqueue_disable_cb(vq);
 215
 216         /* We were probably waiting for more output buffers. */
 217         netif_wake_subqueue(vi->dev, vq2txq(vq));
 218 }
 219
 220 static void set_skb_frag(struct sk_buff *skb, struct page *page,
 221                          unsigned int offset, unsigned int *len)
 222 {
 223         int size = min((unsigned)PAGE_SIZE - offset, *len);
 224         int i = skb_shinfo(skb)->nr_frags;
 225
 226         __skb_fill_page_desc(skb, i, page, offset, size);
 227
 228         skb->data_len += size;
 229         skb->len += size;
 230         skb->truesize += PAGE_SIZE;
 231         skb_shinfo(skb)->nr_frags++;
 232         skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
 233         *len -= size;
 234 }
 235
 236 /* Called from bottom half context */
 237 static struct sk_buff *page_to_skb(struct receive_queue *rq,
 238                                    struct page *page, unsigned int len)
 239 {
 240         struct virtnet_info *vi = rq->vq->vdev->priv;
 241         struct sk_buff *skb;
 242         struct skb_vnet_hdr *hdr;
 243         unsigned int copy, hdr_len, offset;
 244         char *p;
 245
 246         p = page_address(page);
 247
 248         /* copy small packet so we can reuse these pages for small data */
 249         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 250         if (unlikely(!skb))
 251                 return NULL;
 252
 253         hdr = skb_vnet_hdr(skb);
 254
 255         if (vi->mergeable_rx_bufs) {
 256                 hdr_len = sizeof hdr->mhdr;
 257                 offset = hdr_len;
 258         } else {
 259                 hdr_len = sizeof hdr->hdr;
 260                 offset = sizeof(struct padded_vnet_hdr);
 261         }
 262
 263         memcpy(hdr, p, hdr_len);
 264
 265         len -= hdr_len;
 266         p += offset;
 267
 268         copy = len;
 269         if (copy > skb_tailroom(skb))
 270                 copy = skb_tailroom(skb);
 271         memcpy(skb_put(skb, copy), p, copy);
 272
 273         len -= copy;
 274         offset += copy;
 275
 276         /*
 277          * Verify that we can indeed put this data into a skb.
 278          * This is here to handle cases when the device erroneously
 279          * tries to receive more than is possible. This is usually
 280          * the case of a broken device.
 281          */
 282         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 283                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 284                 dev_kfree_skb(skb);
 285                 return NULL;
 286         }
 287
 288         while (len) {
 289                 set_skb_frag(skb, page, offset, &len);
 290                 page = (struct page *)page->private;
 291                 offset = 0;
 292         }
 293
 294         if (page)
 295                 give_pages(rq, page);
 296
 297         return skb;
 298 }
 299
 300 static int receive_mergeable(struct receive_queue *rq, struct sk_buff *skb)
 301 {
 302         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 303         struct page *page;
 304         int num_buf, i, len;
 305
 306         num_buf = hdr->mhdr.num_buffers;
 307         while (--num_buf) {
 308                 i = skb_shinfo(skb)->nr_frags;
 309                 if (i >= MAX_SKB_FRAGS) {
 310                         pr_debug("%s: packet too long\n", skb->dev->name);
 311                         skb->dev->stats.rx_length_errors++;
 312                         return -EINVAL;
 313                 }
 314                 page = virtqueue_get_buf(rq->vq, &len);
 315                 if (!page) {
 316                         pr_debug("%s: rx error: %d buffers missing\n",
 317                                  skb->dev->name, hdr->mhdr.num_buffers);
 318                         skb->dev->stats.rx_length_errors++;
 319                         return -EINVAL;
 320                 }
 321
 322                 if (len > PAGE_SIZE)
 323                         len = PAGE_SIZE;
 324
 325                 set_skb_frag(skb, page, 0, &len);
 326
 327                 --rq->num;
 328         }
 329         return 0;
 330 }
 331
 332 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
 333 {
 334         struct virtnet_info *vi = rq->vq->vdev->priv;
 335         struct net_device *dev = vi->dev;
 336         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 337         struct sk_buff *skb;
 338         struct page *page;
 339         struct skb_vnet_hdr *hdr;
 340
 341         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 342                 pr_debug("%s: short packet %i\n", dev->name, len);
 343                 dev->stats.rx_length_errors++;
 344                 if (vi->mergeable_rx_bufs || vi->big_packets)
 345                         give_pages(rq, buf);
 346                 else
 347                         dev_kfree_skb(buf);
 348                 return;
 349         }
 350
 351         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
 352                 skb = buf;
 353                 len -= sizeof(struct virtio_net_hdr);
 354                 skb_trim(skb, len);
 355         } else {
 356                 page = buf;
 357                 skb = page_to_skb(rq, page, len);
 358                 if (unlikely(!skb)) {
 359                         dev->stats.rx_dropped++;
 360                         give_pages(rq, page);
 361                         return;
 362                 }
 363                 if (vi->mergeable_rx_bufs)
 364                         if (receive_mergeable(rq, skb)) {
 365                                 dev_kfree_skb(skb);
 366                                 return;
 367                         }
 368         }
 369
 370         hdr = skb_vnet_hdr(skb);
 371
 372         u64_stats_update_begin(&stats->rx_syncp);
 373         stats->rx_bytes += skb->len;
 374         stats->rx_packets++;
 375         u64_stats_update_end(&stats->rx_syncp);
 376
 377         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 378                 pr_debug("Needs csum!\n");
 379                 if (!skb_partial_csum_set(skb,
 380                                           hdr->hdr.csum_start,
 381                                           hdr->hdr.csum_offset))
 382                         goto frame_err;
 383         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 384                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 385         }
 386
 387         skb->protocol = eth_type_trans(skb, dev);
 388         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 389                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 390
 391         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 392                 pr_debug("GSO!\n");
 393                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 394                 case VIRTIO_NET_HDR_GSO_TCPV4:
 395                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 396                         break;
 397                 case VIRTIO_NET_HDR_GSO_UDP:
 398                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 399                         break;
 400                 case VIRTIO_NET_HDR_GSO_TCPV6:
 401                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 402                         break;
 403                 default:
 404                         net_warn_ratelimited("%s: bad gso type %u.\n",
 405                                              dev->name, hdr->hdr.gso_type);
 406                         goto frame_err;
 407                 }
 408
 409                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 410                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 411
 412                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 413                 if (skb_shinfo(skb)->gso_size == 0) {
 414                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 415                         goto frame_err;
 416                 }
 417
 418                 /* Header must be checked, and gso_segs computed. */
 419                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 420                 skb_shinfo(skb)->gso_segs = 0;
 421         }
 422
 423         netif_receive_skb(skb);
 424         return;
 425
 426 frame_err:
 427         dev->stats.rx_frame_errors++;
 428         dev_kfree_skb(skb);
 429 }
 430
 431 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
 432 {
 433         struct virtnet_info *vi = rq->vq->vdev->priv;
 434         struct sk_buff *skb;
 435         struct skb_vnet_hdr *hdr;
 436         int err;
 437
 438         skb = __netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN, gfp);
 439         if (unlikely(!skb))
 440                 return -ENOMEM;
 441
 442         skb_put(skb, MAX_PACKET_LEN);
 443
 444         hdr = skb_vnet_hdr(skb);
 445         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
 446
 447         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 448
 449         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 450         if (err < 0)
 451                 dev_kfree_skb(skb);
 452
 453         return err;
 454 }
 455
 456 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
 457 {
 458         struct page *first, *list = NULL;
 459         char *p;
 460         int i, err, offset;
 461
 462         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 463         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 464                 first = get_a_page(rq, gfp);
 465                 if (!first) {
 466                         if (list)
 467                                 give_pages(rq, list);
 468                         return -ENOMEM;
 469                 }
 470                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 471
 472                 /* chain new page in list head to match sg */
 473                 first->private = (unsigned long)list;
 474                 list = first;
 475         }
 476
 477         first = get_a_page(rq, gfp);
 478         if (!first) {
 479                 give_pages(rq, list);
 480                 return -ENOMEM;
 481         }
 482         p = page_address(first);
 483
 484         /* rq->sg[0], rq->sg[1] share the same page */
 485         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
 486         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
 487
 488         /* rq->sg[1] for data packet, from offset */
 489         offset = sizeof(struct padded_vnet_hdr);
 490         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 491
 492         /* chain first in list head */
 493         first->private = (unsigned long)list;
 494         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 495                                   first, gfp);
 496         if (err < 0)
 497                 give_pages(rq, first);
 498
 499         return err;
 500 }
 501
 502 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 503 {
 504         struct page *page;
 505         int err;
 506
 507         page = get_a_page(rq, gfp);
 508         if (!page)
 509                 return -ENOMEM;
 510
 511         sg_init_one(rq->sg, page_address(page), PAGE_SIZE);
 512
 513         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, page, gfp);
 514         if (err < 0)
 515                 give_pages(rq, page);
 516
 517         return err;
 518 }
 519
 520 /*
 521  * Returns false if we couldn't fill entirely (OOM).
 522  *
 523  * Normally run in the receive path, but can also be run from ndo_open
 524  * before we're receiving packets, or from refill_work which is
 525  * careful to disable receiving (using napi_disable).
 526  */
 527 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
 528 {
 529         struct virtnet_info *vi = rq->vq->vdev->priv;
 530         int err;
 531         bool oom;
 532
 533         do {
 534                 if (vi->mergeable_rx_bufs)
 535                         err = add_recvbuf_mergeable(rq, gfp);
 536                 else if (vi->big_packets)
 537                         err = add_recvbuf_big(rq, gfp);
 538                 else
 539                         err = add_recvbuf_small(rq, gfp);
 540
 541                 oom = err == -ENOMEM;
 542                 if (err)
 543                         break;
 544                 ++rq->num;
 545         } while (rq->vq->num_free);
 546         if (unlikely(rq->num > rq->max))
 547                 rq->max = rq->num;
 548         virtqueue_kick(rq->vq);
 549         return !oom;
 550 }
 551
 552 static void skb_recv_done(struct virtqueue *rvq)
 553 {
 554         struct virtnet_info *vi = rvq->vdev->priv;
 555         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 556
 557         /* Schedule NAPI, Suppress further interrupts if successful. */
 558         if (napi_schedule_prep(&rq->napi)) {
 559                 virtqueue_disable_cb(rvq);
 560                 __napi_schedule(&rq->napi);
 561         }
 562 }
 563
 564 static void virtnet_napi_enable(struct receive_queue *rq)
 565 {
 566         napi_enable(&rq->napi);
 567
 568         /* If all buffers were filled by other side before we napi_enabled, we
 569          * won't get another interrupt, so process any outstanding packets
 570          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 571          * We synchronize against interrupts via NAPI_STATE_SCHED */
 572         if (napi_schedule_prep(&rq->napi)) {
 573                 virtqueue_disable_cb(rq->vq);
 574                 local_bh_disable();
 575                 __napi_schedule(&rq->napi);
 576                 local_bh_enable();
 577         }
 578 }
 579
 580 static void refill_work(struct work_struct *work)
 581 {
 582         struct virtnet_info *vi =
 583                 container_of(work, struct virtnet_info, refill.work);
 584         bool still_empty;
 585         int i;
 586
 587         for (i = 0; i < vi->curr_queue_pairs; i++) {
 588                 struct receive_queue *rq = &vi->rq[i];
 589
 590                 napi_disable(&rq->napi);
 591                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 592                 virtnet_napi_enable(rq);
 593
 594                 /* In theory, this can happen: if we don't get any buffers in
 595                  * we will *never* try to fill again.
 596                  */
 597                 if (still_empty)
 598                         schedule_delayed_work(&vi->refill, HZ/2);
 599         }
 600 }
 601
 602 static int virtnet_poll(struct napi_struct *napi, int budget)
 603 {
 604         struct receive_queue *rq =
 605                 container_of(napi, struct receive_queue, napi);
 606         struct virtnet_info *vi = rq->vq->vdev->priv;
 607         void *buf;
 608         unsigned int r, len, received = 0;
 609
 610 again:
 611         while (received < budget &&
 612                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 613                 receive_buf(rq, buf, len);
 614                 --rq->num;
 615                 received++;
 616         }
 617
 618         if (rq->num < rq->max / 2) {
 619                 if (!try_fill_recv(rq, GFP_ATOMIC))
 620                         schedule_delayed_work(&vi->refill, 0);
 621         }
 622
 623         /* Out of packets? */
 624         if (received < budget) {
 625                 r = virtqueue_enable_cb_prepare(rq->vq);
 626                 napi_complete(napi);
 627                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 628                     napi_schedule_prep(napi)) {
 629                         virtqueue_disable_cb(rq->vq);
 630                         __napi_schedule(napi);
 631                         goto again;
 632                 }
 633         }
 634
 635         return received;
 636 }
 637
 638 static int virtnet_open(struct net_device *dev)
 639 {
 640         struct virtnet_info *vi = netdev_priv(dev);
 641         int i;
 642
 643         for (i = 0; i < vi->max_queue_pairs; i++) {
 644                 if (i < vi->curr_queue_pairs)
 645                         /* Make sure we have some buffers: if oom use wq. */
 646                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 647                                 schedule_delayed_work(&vi->refill, 0);
 648                 virtnet_napi_enable(&vi->rq[i]);
 649         }
 650
 651         return 0;
 652 }
 653
 654 static void free_old_xmit_skbs(struct send_queue *sq)
 655 {
 656         struct sk_buff *skb;
 657         unsigned int len;
 658         struct virtnet_info *vi = sq->vq->vdev->priv;
 659         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 660
 661         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 662                 pr_debug("Sent skb %p\n", skb);
 663
 664                 u64_stats_update_begin(&stats->tx_syncp);
 665                 stats->tx_bytes += skb->len;
 666                 stats->tx_packets++;
 667                 u64_stats_update_end(&stats->tx_syncp);
 668
 669                 dev_kfree_skb_any(skb);
 670         }
 671 }
 672
 673 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 674 {
 675         struct skb_vnet_hdr *hdr;
 676         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 677         struct virtnet_info *vi = sq->vq->vdev->priv;
 678         unsigned num_sg;
 679         unsigned hdr_len;
 680         bool can_push;
 681
 682         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 683         if (vi->mergeable_rx_bufs)
 684                 hdr_len = sizeof hdr->mhdr;
 685         else
 686                 hdr_len = sizeof hdr->hdr;
 687
 688         can_push = vi->any_header_sg &&
 689                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 690                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 691         /* Even if we can, don't push here yet as this would skew
 692          * csum_start offset below. */
 693         if (can_push)
 694                 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
 695         else
 696                 hdr = skb_vnet_hdr(skb);
 697
 698         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 699                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 700                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 701                 hdr->hdr.csum_offset = skb->csum_offset;
 702         } else {
 703                 hdr->hdr.flags = 0;
 704                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 705         }
 706
 707         if (skb_is_gso(skb)) {
 708                 hdr->hdr.hdr_len = skb_headlen(skb);
 709                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 710                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 711                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 712                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 713                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 714                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 715                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 716                 else
 717                         BUG();
 718                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 719                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 720         } else {
 721                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 722                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 723         }
 724
 725         if (vi->mergeable_rx_bufs)
 726                 hdr->mhdr.num_buffers = 0;
 727
 728         if (can_push) {
 729                 __skb_push(skb, hdr_len);
 730                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 731                 /* Pull header back to avoid skew in tx bytes calculations. */
 732                 __skb_pull(skb, hdr_len);
 733         } else {
 734                 sg_set_buf(sq->sg, hdr, hdr_len);
 735                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 736         }
 737         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 738 }
 739
 740 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 741 {
 742         struct virtnet_info *vi = netdev_priv(dev);
 743         int qnum = skb_get_queue_mapping(skb);
 744         struct send_queue *sq = &vi->sq[qnum];
 745         int err;
 746
 747         /* Free up any pending old buffers before queueing new ones. */
 748         free_old_xmit_skbs(sq);
 749
 750         /* Try to transmit */
 751         err = xmit_skb(sq, skb);
 752
 753         /* This should not happen! */
 754         if (unlikely(err)) {
 755                 dev->stats.tx_fifo_errors++;
 756                 if (net_ratelimit())
 757                         dev_warn(&dev->dev,
 758                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 759                 dev->stats.tx_dropped++;
 760                 kfree_skb(skb);
 761                 return NETDEV_TX_OK;
 762         }
 763         virtqueue_kick(sq->vq);
 764
 765         /* Don't wait up for transmitted skbs to be freed. */
 766         skb_orphan(skb);
 767         nf_reset(skb);
 768
 769         /* Apparently nice girls don't return TX_BUSY; stop the queue
 770          * before it gets out of hand.  Naturally, this wastes entries. */
 771         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 772                 netif_stop_subqueue(dev, qnum);
 773                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 774                         /* More just got used, free them then recheck. */
 775                         free_old_xmit_skbs(sq);
 776                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 777                                 netif_start_subqueue(dev, qnum);
 778                                 virtqueue_disable_cb(sq->vq);
 779                         }
 780                 }
 781         }
 782
 783         return NETDEV_TX_OK;
 784 }
 785
 786 /*
 787  * Send command via the control virtqueue and check status.  Commands
 788  * supported by the hypervisor, as indicated by feature bits, should
 789  * never fail unless improperly formated.
 790  */
 791 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 792                                  struct scatterlist *out,
 793                                  struct scatterlist *in)
 794 {
 795         struct scatterlist *sgs[4], hdr, stat;
 796         struct virtio_net_ctrl_hdr ctrl;
 797         virtio_net_ctrl_ack status = ~0;
 798         unsigned out_num = 0, in_num = 0, tmp;
 799
 800         /* Caller should know better */
 801         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 802
 803         ctrl.class = class;
 804         ctrl.cmd = cmd;
 805         /* Add header */
 806         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 807         sgs[out_num++] = &hdr;
 808
 809         if (out)
 810                 sgs[out_num++] = out;
 811         if (in)
 812                 sgs[out_num + in_num++] = in;
 813
 814         /* Add return status. */
 815         sg_init_one(&stat, &status, sizeof(status));
 816         sgs[out_num + in_num++] = &stat;
 817
 818         BUG_ON(out_num + in_num > ARRAY_SIZE(sgs));
 819         BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, in_num, vi, GFP_ATOMIC)
 820                < 0);
 821
 822         virtqueue_kick(vi->cvq);
 823
 824         /* Spin for a response, the kick causes an ioport write, trapping
 825          * into the hypervisor, so the request should be handled immediately.
 826          */
 827         while (!virtqueue_get_buf(vi->cvq, &tmp))
 828                 cpu_relax();
 829
 830         return status == VIRTIO_NET_OK;
 831 }
 832
 833 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 834 {
 835         struct virtnet_info *vi = netdev_priv(dev);
 836         struct virtio_device *vdev = vi->vdev;
 837         int ret;
 838         struct sockaddr *addr = p;
 839         struct scatterlist sg;
 840
 841         ret = eth_prepare_mac_addr_change(dev, p);
 842         if (ret)
 843                 return ret;
 844
 845         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 846                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 847                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 848                                           VIRTIO_NET_CTRL_MAC_ADDR_SET,
 849                                           &sg, NULL)) {
 850                         dev_warn(&vdev->dev,
 851                                  "Failed to set mac address by vq command.\n");
 852                         return -EINVAL;
 853                 }
 854         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 855                 unsigned int i;
 856
 857                 /* Naturally, this has an atomicity problem. */
 858                 for (i = 0; i < dev->addr_len; i++)
 859                         virtio_cwrite8(vdev,
 860                                        offsetof(struct virtio_net_config, mac) +
 861                                        i, addr->sa_data[i]);
 862         }
 863
 864         eth_commit_mac_addr_change(dev, p);
 865
 866         return 0;
 867 }
 868
 869 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 870                                                struct rtnl_link_stats64 *tot)
 871 {
 872         struct virtnet_info *vi = netdev_priv(dev);
 873         int cpu;
 874         unsigned int start;
 875
 876         for_each_possible_cpu(cpu) {
 877                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 878                 u64 tpackets, tbytes, rpackets, rbytes;
 879
 880                 do {
 881                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 882                         tpackets = stats->tx_packets;
 883                         tbytes   = stats->tx_bytes;
 884                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 885
 886                 do {
 887                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 888                         rpackets = stats->rx_packets;
 889                         rbytes   = stats->rx_bytes;
 890                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 891
 892                 tot->rx_packets += rpackets;
 893                 tot->tx_packets += tpackets;
 894                 tot->rx_bytes   += rbytes;
 895                 tot->tx_bytes   += tbytes;
 896         }
 897
 898         tot->tx_dropped = dev->stats.tx_dropped;
 899         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 900         tot->rx_dropped = dev->stats.rx_dropped;
 901         tot->rx_length_errors = dev->stats.rx_length_errors;
 902         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 903
 904         return tot;
 905 }
 906
 907 #ifdef CONFIG_NET_POLL_CONTROLLER
 908 static void virtnet_netpoll(struct net_device *dev)
 909 {
 910         struct virtnet_info *vi = netdev_priv(dev);
 911         int i;
 912
 913         for (i = 0; i < vi->curr_queue_pairs; i++)
 914                 napi_schedule(&vi->rq[i].napi);
 915 }
 916 #endif
 917
 918 static void virtnet_ack_link_announce(struct virtnet_info *vi)
 919 {
 920         rtnl_lock();
 921         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
 922                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
 923                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
 924         rtnl_unlock();
 925 }
 926
 927 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
 928 {
 929         struct scatterlist sg;
 930         struct virtio_net_ctrl_mq s;
 931         struct net_device *dev = vi->dev;
 932
 933         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
 934                 return 0;
 935
 936         s.virtqueue_pairs = queue_pairs;
 937         sg_init_one(&sg, &s, sizeof(s));
 938
 939         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
 940                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
 941                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
 942                          queue_pairs);
 943                 return -EINVAL;
 944         } else {
 945                 vi->curr_queue_pairs = queue_pairs;
 946                 /* virtnet_open() will refill when device is going to up. */
 947                 if (dev->flags & IFF_UP)
 948                         schedule_delayed_work(&vi->refill, 0);
 949         }
 950
 951         return 0;
 952 }
 953
 954 static int virtnet_close(struct net_device *dev)
 955 {
 956         struct virtnet_info *vi = netdev_priv(dev);
 957         int i;
 958
 959         /* Make sure refill_work doesn't re-enable napi! */
 960         cancel_delayed_work_sync(&vi->refill);
 961
 962         for (i = 0; i < vi->max_queue_pairs; i++)
 963                 napi_disable(&vi->rq[i].napi);
 964
 965         return 0;
 966 }
 967
 968 static void virtnet_set_rx_mode(struct net_device *dev)
 969 {
 970         struct virtnet_info *vi = netdev_priv(dev);
 971         struct scatterlist sg[2];
 972         u8 promisc, allmulti;
 973         struct virtio_net_ctrl_mac *mac_data;
 974         struct netdev_hw_addr *ha;
 975         int uc_count;
 976         int mc_count;
 977         void *buf;
 978         int i;
 979
 980         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 981         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 982                 return;
 983
 984         promisc = ((dev->flags & IFF_PROMISC) != 0);
 985         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 986
 987         sg_init_one(sg, &promisc, sizeof(promisc));
 988
 989         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 990                                   VIRTIO_NET_CTRL_RX_PROMISC,
 991                                   sg, NULL))
 992                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 993                          promisc ? "en" : "dis");
 994
 995         sg_init_one(sg, &allmulti, sizeof(allmulti));
 996
 997         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 998                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 999                                   sg, NULL))
1000                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1001                          allmulti ? "en" : "dis");
1002
1003         uc_count = netdev_uc_count(dev);
1004         mc_count = netdev_mc_count(dev);
1005         /* MAC filter - use one buffer for both lists */
1006         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1007                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1008         mac_data = buf;
1009         if (!buf)
1010                 return;
1011
1012         sg_init_table(sg, 2);
1013
1014         /* Store the unicast list and count in the front of the buffer */
1015         mac_data->entries = uc_count;
1016         i = 0;
1017         netdev_for_each_uc_addr(ha, dev)
1018                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1019
1020         sg_set_buf(&sg[0], mac_data,
1021                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1022
1023         /* multicast list and count fill the end */
1024         mac_data = (void *)&mac_data->macs[uc_count][0];
1025
1026         mac_data->entries = mc_count;
1027         i = 0;
1028         netdev_for_each_mc_addr(ha, dev)
1029                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1030
1031         sg_set_buf(&sg[1], mac_data,
1032                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1033
1034         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1035                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
1036                                   sg, NULL))
1037                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1038
1039         kfree(buf);
1040 }
1041
1042 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1043                                    __be16 proto, u16 vid)
1044 {
1045         struct virtnet_info *vi = netdev_priv(dev);
1046         struct scatterlist sg;
1047
1048         sg_init_one(&sg, &vid, sizeof(vid));
1049
1050         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1051                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1052                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1053         return 0;
1054 }
1055
1056 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1057                                     __be16 proto, u16 vid)
1058 {
1059         struct virtnet_info *vi = netdev_priv(dev);
1060         struct scatterlist sg;
1061
1062         sg_init_one(&sg, &vid, sizeof(vid));
1063
1064         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1065                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1066                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1067         return 0;
1068 }
1069
1070 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1071 {
1072         int i;
1073         int cpu;
1074
1075         if (vi->affinity_hint_set) {
1076                 for (i = 0; i < vi->max_queue_pairs; i++) {
1077                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1078                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1079                 }
1080
1081                 vi->affinity_hint_set = false;
1082         }
1083
1084         i = 0;
1085         for_each_online_cpu(cpu) {
1086                 if (cpu == hcpu) {
1087                         *per_cpu_ptr(vi->vq_index, cpu) = -1;
1088                 } else {
1089                         *per_cpu_ptr(vi->vq_index, cpu) =
1090                                 ++i % vi->curr_queue_pairs;
1091                 }
1092         }
1093 }
1094
1095 static void virtnet_set_affinity(struct virtnet_info *vi)
1096 {
1097         int i;
1098         int cpu;
1099
1100         /* In multiqueue mode, when the number of cpu is equal to the number of
1101          * queue pairs, we let the queue pairs to be private to one cpu by
1102          * setting the affinity hint to eliminate the contention.
1103          */
1104         if (vi->curr_queue_pairs == 1 ||
1105             vi->max_queue_pairs != num_online_cpus()) {
1106                 virtnet_clean_affinity(vi, -1);
1107                 return;
1108         }
1109
1110         i = 0;
1111         for_each_online_cpu(cpu) {
1112                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1113                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1114                 *per_cpu_ptr(vi->vq_index, cpu) = i;
1115                 i++;
1116         }
1117
1118         vi->affinity_hint_set = true;
1119 }
1120
1121 static int virtnet_cpu_callback(struct notifier_block *nfb,
1122                                 unsigned long action, void *hcpu)
1123 {
1124         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1125
1126         mutex_lock(&vi->config_lock);
1127
1128         if (!vi->config_enable)
1129                 goto done;
1130
1131         switch(action & ~CPU_TASKS_FROZEN) {
1132         case CPU_ONLINE:
1133         case CPU_DOWN_FAILED:
1134         case CPU_DEAD:
1135                 virtnet_set_affinity(vi);
1136                 break;
1137         case CPU_DOWN_PREPARE:
1138                 virtnet_clean_affinity(vi, (long)hcpu);
1139                 break;
1140         default:
1141                 break;
1142         }
1143
1144 done:
1145         mutex_unlock(&vi->config_lock);
1146         return NOTIFY_OK;
1147 }
1148
1149 static void virtnet_get_ringparam(struct net_device *dev,
1150                                 struct ethtool_ringparam *ring)
1151 {
1152         struct virtnet_info *vi = netdev_priv(dev);
1153
1154         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1155         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1156         ring->rx_pending = ring->rx_max_pending;
1157         ring->tx_pending = ring->tx_max_pending;
1158 }
1159
1160
1161 static void virtnet_get_drvinfo(struct net_device *dev,
1162                                 struct ethtool_drvinfo *info)
1163 {
1164         struct virtnet_info *vi = netdev_priv(dev);
1165         struct virtio_device *vdev = vi->vdev;
1166
1167         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1168         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1169         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1170
1171 }
1172
1173 /* TODO: Eliminate OOO packets during switching */
1174 static int virtnet_set_channels(struct net_device *dev,
1175                                 struct ethtool_channels *channels)
1176 {
1177         struct virtnet_info *vi = netdev_priv(dev);
1178         u16 queue_pairs = channels->combined_count;
1179         int err;
1180
1181         /* We don't support separate rx/tx channels.
1182          * We don't allow setting 'other' channels.
1183          */
1184         if (channels->rx_count || channels->tx_count || channels->other_count)
1185                 return -EINVAL;
1186
1187         if (queue_pairs > vi->max_queue_pairs)
1188                 return -EINVAL;
1189
1190         get_online_cpus();
1191         err = virtnet_set_queues(vi, queue_pairs);
1192         if (!err) {
1193                 netif_set_real_num_tx_queues(dev, queue_pairs);
1194                 netif_set_real_num_rx_queues(dev, queue_pairs);
1195
1196                 virtnet_set_affinity(vi);
1197         }
1198         put_online_cpus();
1199
1200         return err;
1201 }
1202
1203 static void virtnet_get_channels(struct net_device *dev,
1204                                  struct ethtool_channels *channels)
1205 {
1206         struct virtnet_info *vi = netdev_priv(dev);
1207
1208         channels->combined_count = vi->curr_queue_pairs;
1209         channels->max_combined = vi->max_queue_pairs;
1210         channels->max_other = 0;
1211         channels->rx_count = 0;
1212         channels->tx_count = 0;
1213         channels->other_count = 0;
1214 }
1215
1216 static const struct ethtool_ops virtnet_ethtool_ops = {
1217         .get_drvinfo = virtnet_get_drvinfo,
1218         .get_link = ethtool_op_get_link,
1219         .get_ringparam = virtnet_get_ringparam,
1220         .set_channels = virtnet_set_channels,
1221         .get_channels = virtnet_get_channels,
1222 };
1223
1224 #define MIN_MTU 68
1225 #define MAX_MTU 65535
1226
1227 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1228 {
1229         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1230                 return -EINVAL;
1231         dev->mtu = new_mtu;
1232         return 0;
1233 }
1234
1235 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
1236  * txq based on the processor id.
1237  */
1238 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
1239 {
1240         int txq;
1241         struct virtnet_info *vi = netdev_priv(dev);
1242
1243         if (skb_rx_queue_recorded(skb)) {
1244                 txq = skb_get_rx_queue(skb);
1245         } else {
1246                 txq = *__this_cpu_ptr(vi->vq_index);
1247                 if (txq == -1)
1248                         txq = 0;
1249         }
1250
1251         while (unlikely(txq >= dev->real_num_tx_queues))
1252                 txq -= dev->real_num_tx_queues;
1253
1254         return txq;
1255 }
1256
1257 static const struct net_device_ops virtnet_netdev = {
1258         .ndo_open            = virtnet_open,
1259         .ndo_stop            = virtnet_close,
1260         .ndo_start_xmit      = start_xmit,
1261         .ndo_validate_addr   = eth_validate_addr,
1262         .ndo_set_mac_address = virtnet_set_mac_address,
1263         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1264         .ndo_change_mtu      = virtnet_change_mtu,
1265         .ndo_get_stats64     = virtnet_stats,
1266         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1267         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1268         .ndo_select_queue     = virtnet_select_queue,
1269 #ifdef CONFIG_NET_POLL_CONTROLLER
1270         .ndo_poll_controller = virtnet_netpoll,
1271 #endif
1272 };
1273
1274 static void virtnet_config_changed_work(struct work_struct *work)
1275 {
1276         struct virtnet_info *vi =
1277                 container_of(work, struct virtnet_info, config_work);
1278         u16 v;
1279
1280         mutex_lock(&vi->config_lock);
1281         if (!vi->config_enable)
1282                 goto done;
1283
1284         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1285                                  struct virtio_net_config, status, &v) < 0)
1286                 goto done;
1287
1288         if (v & VIRTIO_NET_S_ANNOUNCE) {
1289                 netdev_notify_peers(vi->dev);
1290                 virtnet_ack_link_announce(vi);
1291         }
1292
1293         /* Ignore unknown (future) status bits */
1294         v &= VIRTIO_NET_S_LINK_UP;
1295
1296         if (vi->status == v)
1297                 goto done;
1298
1299         vi->status = v;
1300
1301         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1302                 netif_carrier_on(vi->dev);
1303                 netif_tx_wake_all_queues(vi->dev);
1304         } else {
1305                 netif_carrier_off(vi->dev);
1306                 netif_tx_stop_all_queues(vi->dev);
1307         }
1308 done:
1309         mutex_unlock(&vi->config_lock);
1310 }
1311
1312 static void virtnet_config_changed(struct virtio_device *vdev)
1313 {
1314         struct virtnet_info *vi = vdev->priv;
1315
1316         schedule_work(&vi->config_work);
1317 }
1318
1319 static void virtnet_free_queues(struct virtnet_info *vi)
1320 {
1321         kfree(vi->rq);
1322         kfree(vi->sq);
1323 }
1324
1325 static void free_receive_bufs(struct virtnet_info *vi)
1326 {
1327         int i;
1328
1329         for (i = 0; i < vi->max_queue_pairs; i++) {
1330                 while (vi->rq[i].pages)
1331                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1332         }
1333 }
1334
1335 static void free_unused_bufs(struct virtnet_info *vi)
1336 {
1337         void *buf;
1338         int i;
1339
1340         for (i = 0; i < vi->max_queue_pairs; i++) {
1341                 struct virtqueue *vq = vi->sq[i].vq;
1342                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1343                         dev_kfree_skb(buf);
1344         }
1345
1346         for (i = 0; i < vi->max_queue_pairs; i++) {
1347                 struct virtqueue *vq = vi->rq[i].vq;
1348
1349                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1350                         if (vi->mergeable_rx_bufs || vi->big_packets)
1351                                 give_pages(&vi->rq[i], buf);
1352                         else
1353                                 dev_kfree_skb(buf);
1354                         --vi->rq[i].num;
1355                 }
1356                 BUG_ON(vi->rq[i].num != 0);
1357         }
1358 }
1359
1360 static void virtnet_del_vqs(struct virtnet_info *vi)
1361 {
1362         struct virtio_device *vdev = vi->vdev;
1363
1364         virtnet_clean_affinity(vi, -1);
1365
1366         vdev->config->del_vqs(vdev);
1367
1368         virtnet_free_queues(vi);
1369 }
1370
1371 static int virtnet_find_vqs(struct virtnet_info *vi)
1372 {
1373         vq_callback_t **callbacks;
1374         struct virtqueue **vqs;
1375         int ret = -ENOMEM;
1376         int i, total_vqs;
1377         const char **names;
1378
1379         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1380          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1381          * possible control vq.
1382          */
1383         total_vqs = vi->max_queue_pairs * 2 +
1384                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1385
1386         /* Allocate space for find_vqs parameters */
1387         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1388         if (!vqs)
1389                 goto err_vq;
1390         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1391         if (!callbacks)
1392                 goto err_callback;
1393         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1394         if (!names)
1395                 goto err_names;
1396
1397         /* Parameters for control virtqueue, if any */
1398         if (vi->has_cvq) {
1399                 callbacks[total_vqs - 1] = NULL;
1400                 names[total_vqs - 1] = "control";
1401         }
1402
1403         /* Allocate/initialize parameters for send/receive virtqueues */
1404         for (i = 0; i < vi->max_queue_pairs; i++) {
1405                 callbacks[rxq2vq(i)] = skb_recv_done;
1406                 callbacks[txq2vq(i)] = skb_xmit_done;
1407                 sprintf(vi->rq[i].name, "input.%d", i);
1408                 sprintf(vi->sq[i].name, "output.%d", i);
1409                 names[rxq2vq(i)] = vi->rq[i].name;
1410                 names[txq2vq(i)] = vi->sq[i].name;
1411         }
1412
1413         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1414                                          names);
1415         if (ret)
1416                 goto err_find;
1417
1418         if (vi->has_cvq) {
1419                 vi->cvq = vqs[total_vqs - 1];
1420                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1421                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1422         }
1423
1424         for (i = 0; i < vi->max_queue_pairs; i++) {
1425                 vi->rq[i].vq = vqs[rxq2vq(i)];
1426                 vi->sq[i].vq = vqs[txq2vq(i)];
1427         }
1428
1429         kfree(names);
1430         kfree(callbacks);
1431         kfree(vqs);
1432
1433         return 0;
1434
1435 err_find:
1436         kfree(names);
1437 err_names:
1438         kfree(callbacks);
1439 err_callback:
1440         kfree(vqs);
1441 err_vq:
1442         return ret;
1443 }
1444
1445 static int virtnet_alloc_queues(struct virtnet_info *vi)
1446 {
1447         int i;
1448
1449         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1450         if (!vi->sq)
1451                 goto err_sq;
1452         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1453         if (!vi->rq)
1454                 goto err_rq;
1455
1456         INIT_DELAYED_WORK(&vi->refill, refill_work);
1457         for (i = 0; i < vi->max_queue_pairs; i++) {
1458                 vi->rq[i].pages = NULL;
1459                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1460                                napi_weight);
1461
1462                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1463                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1464         }
1465
1466         return 0;
1467
1468 err_rq:
1469         kfree(vi->sq);
1470 err_sq:
1471         return -ENOMEM;
1472 }
1473
1474 static int init_vqs(struct virtnet_info *vi)
1475 {
1476         int ret;
1477
1478         /* Allocate send & receive queues */
1479         ret = virtnet_alloc_queues(vi);
1480         if (ret)
1481                 goto err;
1482
1483         ret = virtnet_find_vqs(vi);
1484         if (ret)
1485                 goto err_free;
1486
1487         get_online_cpus();
1488         virtnet_set_affinity(vi);
1489         put_online_cpus();
1490
1491         return 0;
1492
1493 err_free:
1494         virtnet_free_queues(vi);
1495 err:
1496         return ret;
1497 }
1498
1499 static int virtnet_probe(struct virtio_device *vdev)
1500 {
1501         int i, err;
1502         struct net_device *dev;
1503         struct virtnet_info *vi;
1504         u16 max_queue_pairs;
1505
1506         /* Find if host supports multiqueue virtio_net device */
1507         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1508                                    struct virtio_net_config,
1509                                    max_virtqueue_pairs, &max_queue_pairs);
1510
1511         /* We need at least 2 queue's */
1512         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1513             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1514             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1515                 max_queue_pairs = 1;
1516
1517         /* Allocate ourselves a network device with room for our info */
1518         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1519         if (!dev)
1520                 return -ENOMEM;
1521
1522         /* Set up network device as normal. */
1523         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1524         dev->netdev_ops = &virtnet_netdev;
1525         dev->features = NETIF_F_HIGHDMA;
1526
1527         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1528         SET_NETDEV_DEV(dev, &vdev->dev);
1529
1530         /* Do we support "hardware" checksums? */
1531         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1532                 /* This opens up the world of extra features. */
1533                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1534                 if (csum)
1535                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1536
1537                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1538                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1539                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1540                 }
1541                 /* Individual feature bits: what can host handle? */
1542                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1543                         dev->hw_features |= NETIF_F_TSO;
1544                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1545                         dev->hw_features |= NETIF_F_TSO6;
1546                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1547                         dev->hw_features |= NETIF_F_TSO_ECN;
1548                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1549                         dev->hw_features |= NETIF_F_UFO;
1550
1551                 if (gso)
1552                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1553                 /* (!csum && gso) case will be fixed by register_netdev() */
1554         }
1555         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1556                 dev->features |= NETIF_F_RXCSUM;
1557
1558         dev->vlan_features = dev->features;
1559
1560         /* Configuration may specify what MAC to use.  Otherwise random. */
1561         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1562                 virtio_cread_bytes(vdev,
1563                                    offsetof(struct virtio_net_config, mac),
1564                                    dev->dev_addr, dev->addr_len);
1565         else
1566                 eth_hw_addr_random(dev);
1567
1568         /* Set up our device-specific information */
1569         vi = netdev_priv(dev);
1570         vi->dev = dev;
1571         vi->vdev = vdev;
1572         vdev->priv = vi;
1573         vi->stats = alloc_percpu(struct virtnet_stats);
1574         err = -ENOMEM;
1575         if (vi->stats == NULL)
1576                 goto free;
1577
1578         vi->vq_index = alloc_percpu(int);
1579         if (vi->vq_index == NULL)
1580                 goto free_stats;
1581
1582         mutex_init(&vi->config_lock);
1583         vi->config_enable = true;
1584         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1585
1586         /* If we can receive ANY GSO packets, we must allocate large ones. */
1587         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1588             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1589             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1590                 vi->big_packets = true;
1591
1592         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1593                 vi->mergeable_rx_bufs = true;
1594
1595         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1596                 vi->any_header_sg = true;
1597
1598         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1599                 vi->has_cvq = true;
1600
1601         /* Use single tx/rx queue pair as default */
1602         vi->curr_queue_pairs = 1;
1603         vi->max_queue_pairs = max_queue_pairs;
1604
1605         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1606         err = init_vqs(vi);
1607         if (err)
1608                 goto free_index;
1609
1610         netif_set_real_num_tx_queues(dev, 1);
1611         netif_set_real_num_rx_queues(dev, 1);
1612
1613         err = register_netdev(dev);
1614         if (err) {
1615                 pr_debug("virtio_net: registering device failed\n");
1616                 goto free_vqs;
1617         }
1618
1619         /* Last of all, set up some receive buffers. */
1620         for (i = 0; i < vi->curr_queue_pairs; i++) {
1621                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1622
1623                 /* If we didn't even get one input buffer, we're useless. */
1624                 if (vi->rq[i].num == 0) {
1625                         free_unused_bufs(vi);
1626                         err = -ENOMEM;
1627                         goto free_recv_bufs;
1628                 }
1629         }
1630
1631         vi->nb.notifier_call = &virtnet_cpu_callback;
1632         err = register_hotcpu_notifier(&vi->nb);
1633         if (err) {
1634                 pr_debug("virtio_net: registering cpu notifier failed\n");
1635                 goto free_recv_bufs;
1636         }
1637
1638         /* Assume link up if device can't report link status,
1639            otherwise get link status from config. */
1640         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1641                 netif_carrier_off(dev);
1642                 schedule_work(&vi->config_work);
1643         } else {
1644                 vi->status = VIRTIO_NET_S_LINK_UP;
1645                 netif_carrier_on(dev);
1646         }
1647
1648         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1649                  dev->name, max_queue_pairs);
1650
1651         return 0;
1652
1653 free_recv_bufs:
1654         free_receive_bufs(vi);
1655         unregister_netdev(dev);
1656 free_vqs:
1657         cancel_delayed_work_sync(&vi->refill);
1658         virtnet_del_vqs(vi);
1659 free_index:
1660         free_percpu(vi->vq_index);
1661 free_stats:
1662         free_percpu(vi->stats);
1663 free:
1664         free_netdev(dev);
1665         return err;
1666 }
1667
1668 static void remove_vq_common(struct virtnet_info *vi)
1669 {
1670         vi->vdev->config->reset(vi->vdev);
1671
1672         /* Free unused buffers in both send and recv, if any. */
1673         free_unused_bufs(vi);
1674
1675         free_receive_bufs(vi);
1676
1677         virtnet_del_vqs(vi);
1678 }
1679
1680 static void virtnet_remove(struct virtio_device *vdev)
1681 {
1682         struct virtnet_info *vi = vdev->priv;
1683
1684         unregister_hotcpu_notifier(&vi->nb);
1685
1686         /* Prevent config work handler from accessing the device. */
1687         mutex_lock(&vi->config_lock);
1688         vi->config_enable = false;
1689         mutex_unlock(&vi->config_lock);
1690
1691         unregister_netdev(vi->dev);
1692
1693         remove_vq_common(vi);
1694
1695         flush_work(&vi->config_work);
1696
1697         free_percpu(vi->vq_index);
1698         free_percpu(vi->stats);
1699         free_netdev(vi->dev);
1700 }
1701
1702 #ifdef CONFIG_PM_SLEEP
1703 static int virtnet_freeze(struct virtio_device *vdev)
1704 {
1705         struct virtnet_info *vi = vdev->priv;
1706         int i;
1707
1708         /* Prevent config work handler from accessing the device */
1709         mutex_lock(&vi->config_lock);
1710         vi->config_enable = false;
1711         mutex_unlock(&vi->config_lock);
1712
1713         netif_device_detach(vi->dev);
1714         cancel_delayed_work_sync(&vi->refill);
1715
1716         if (netif_running(vi->dev))
1717                 for (i = 0; i < vi->max_queue_pairs; i++) {
1718                         napi_disable(&vi->rq[i].napi);
1719                         netif_napi_del(&vi->rq[i].napi);
1720                 }
1721
1722         remove_vq_common(vi);
1723
1724         flush_work(&vi->config_work);
1725
1726         return 0;
1727 }
1728
1729 static int virtnet_restore(struct virtio_device *vdev)
1730 {
1731         struct virtnet_info *vi = vdev->priv;
1732         int err, i;
1733
1734         err = init_vqs(vi);
1735         if (err)
1736                 return err;
1737
1738         if (netif_running(vi->dev))
1739                 for (i = 0; i < vi->max_queue_pairs; i++)
1740                         virtnet_napi_enable(&vi->rq[i]);
1741
1742         netif_device_attach(vi->dev);
1743
1744         for (i = 0; i < vi->curr_queue_pairs; i++)
1745                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1746                         schedule_delayed_work(&vi->refill, 0);
1747
1748         mutex_lock(&vi->config_lock);
1749         vi->config_enable = true;
1750         mutex_unlock(&vi->config_lock);
1751
1752         rtnl_lock();
1753         virtnet_set_queues(vi, vi->curr_queue_pairs);
1754         rtnl_unlock();
1755
1756         return 0;
1757 }
1758 #endif
1759
1760 static struct virtio_device_id id_table[] = {
1761         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1762         { 0 },
1763 };
1764
1765 static unsigned int features[] = {
1766         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1767         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1768         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1769         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1770         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1771         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1772         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1773         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1774         VIRTIO_NET_F_CTRL_MAC_ADDR,
1775         VIRTIO_F_ANY_LAYOUT,
1776 };
1777
1778 static struct virtio_driver virtio_net_driver = {
1779         .feature_table = features,
1780         .feature_table_size = ARRAY_SIZE(features),
1781         .driver.name =  KBUILD_MODNAME,
1782         .driver.owner = THIS_MODULE,
1783         .id_table =     id_table,
1784         .probe =        virtnet_probe,
1785         .remove =       virtnet_remove,
1786         .config_changed = virtnet_config_changed,
1787 #ifdef CONFIG_PM_SLEEP
1788         .freeze =       virtnet_freeze,
1789         .restore =      virtnet_restore,
1790 #endif
1791 };
1792
1793 module_virtio_driver(virtio_net_driver);
1794
1795 MODULE_DEVICE_TABLE(virtio, id_table);
1796 MODULE_DESCRIPTION("Virtio network driver");
1797 MODULE_LICENSE("GPL");