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         /* Page_frag for GFP_KERNEL packet buffer allocation when we run
 128          * low on memory.
 129          */
 130         struct page_frag alloc_frag;
 131
 132         /* Does the affinity hint is set for virtqueues? */
 133         bool affinity_hint_set;
 134
 135         /* CPU hot plug notifier */
 136         struct notifier_block nb;
 137 };
 138
 139 struct skb_vnet_hdr {
 140         union {
 141                 struct virtio_net_hdr hdr;
 142                 struct virtio_net_hdr_mrg_rxbuf mhdr;
 143         };
 144 };
 145
 146 struct padded_vnet_hdr {
 147         struct virtio_net_hdr hdr;
 148         /*
 149          * virtio_net_hdr should be in a separated sg buffer because of a
 150          * QEMU bug, and data sg buffer shares same page with this header sg.
 151          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
 152          */
 153         char padding[6];
 154 };
 155
 156 /* Converting between virtqueue no. and kernel tx/rx queue no.
 157  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 158  */
 159 static int vq2txq(struct virtqueue *vq)
 160 {
 161         return (vq->index - 1) / 2;
 162 }
 163
 164 static int txq2vq(int txq)
 165 {
 166         return txq * 2 + 1;
 167 }
 168
 169 static int vq2rxq(struct virtqueue *vq)
 170 {
 171         return vq->index / 2;
 172 }
 173
 174 static int rxq2vq(int rxq)
 175 {
 176         return rxq * 2;
 177 }
 178
 179 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 180 {
 181         return (struct skb_vnet_hdr *)skb->cb;
 182 }
 183
 184 /*
 185  * private is used to chain pages for big packets, put the whole
 186  * most recent used list in the beginning for reuse
 187  */
 188 static void give_pages(struct receive_queue *rq, struct page *page)
 189 {
 190         struct page *end;
 191
 192         /* Find end of list, sew whole thing into vi->rq.pages. */
 193         for (end = page; end->private; end = (struct page *)end->private);
 194         end->private = (unsigned long)rq->pages;
 195         rq->pages = page;
 196 }
 197
 198 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 199 {
 200         struct page *p = rq->pages;
 201
 202         if (p) {
 203                 rq->pages = (struct page *)p->private;
 204                 /* clear private here, it is used to chain pages */
 205                 p->private = 0;
 206         } else
 207                 p = alloc_page(gfp_mask);
 208         return p;
 209 }
 210
 211 static void skb_xmit_done(struct virtqueue *vq)
 212 {
 213         struct virtnet_info *vi = vq->vdev->priv;
 214
 215         /* Suppress further interrupts. */
 216         virtqueue_disable_cb(vq);
 217
 218         /* We were probably waiting for more output buffers. */
 219         netif_wake_subqueue(vi->dev, vq2txq(vq));
 220 }
 221
 222 /* Called from bottom half context */
 223 static struct sk_buff *page_to_skb(struct receive_queue *rq,
 224                                    struct page *page, unsigned int offset,
 225                                    unsigned int len, unsigned int truesize)
 226 {
 227         struct virtnet_info *vi = rq->vq->vdev->priv;
 228         struct sk_buff *skb;
 229         struct skb_vnet_hdr *hdr;
 230         unsigned int copy, hdr_len, hdr_padded_len;
 231         char *p;
 232
 233         p = page_address(page) + offset;
 234
 235         /* copy small packet so we can reuse these pages for small data */
 236         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 237         if (unlikely(!skb))
 238                 return NULL;
 239
 240         hdr = skb_vnet_hdr(skb);
 241
 242         if (vi->mergeable_rx_bufs) {
 243                 hdr_len = sizeof hdr->mhdr;
 244                 hdr_padded_len = sizeof hdr->mhdr;
 245         } else {
 246                 hdr_len = sizeof hdr->hdr;
 247                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 248         }
 249
 250         memcpy(hdr, p, hdr_len);
 251
 252         len -= hdr_len;
 253         offset += hdr_padded_len;
 254         p += hdr_padded_len;
 255
 256         copy = len;
 257         if (copy > skb_tailroom(skb))
 258                 copy = skb_tailroom(skb);
 259         memcpy(skb_put(skb, copy), p, copy);
 260
 261         len -= copy;
 262         offset += copy;
 263
 264         if (vi->mergeable_rx_bufs) {
 265                 if (len)
 266                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 267                 else
 268                         put_page(page);
 269                 return skb;
 270         }
 271
 272         /*
 273          * Verify that we can indeed put this data into a skb.
 274          * This is here to handle cases when the device erroneously
 275          * tries to receive more than is possible. This is usually
 276          * the case of a broken device.
 277          */
 278         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 279                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 280                 dev_kfree_skb(skb);
 281                 return NULL;
 282         }
 283         BUG_ON(offset >= PAGE_SIZE);
 284         while (len) {
 285                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 286                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 287                                 frag_size, truesize);
 288                 len -= frag_size;
 289                 page = (struct page *)page->private;
 290                 offset = 0;
 291         }
 292
 293         if (page)
 294                 give_pages(rq, page);
 295
 296         return skb;
 297 }
 298
 299 static int receive_mergeable(struct receive_queue *rq, struct sk_buff *head_skb)
 300 {
 301         struct skb_vnet_hdr *hdr = skb_vnet_hdr(head_skb);
 302         struct sk_buff *curr_skb = head_skb;
 303         char *buf;
 304         struct page *page;
 305         int num_buf, len, offset;
 306
 307         num_buf = hdr->mhdr.num_buffers;
 308         while (--num_buf) {
 309                 int num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 310                 buf = virtqueue_get_buf(rq->vq, &len);
 311                 if (unlikely(!buf)) {
 312                         pr_debug("%s: rx error: %d buffers missing\n",
 313                                  head_skb->dev->name, hdr->mhdr.num_buffers);
 314                         head_skb->dev->stats.rx_length_errors++;
 315                         return -EINVAL;
 316                 }
 317                 if (unlikely(len > MAX_PACKET_LEN)) {
 318                         pr_debug("%s: rx error: merge buffer too long\n",
 319                                  head_skb->dev->name);
 320                         len = MAX_PACKET_LEN;
 321                 }
 322                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 323                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 324                         if (unlikely(!nskb)) {
 325                                 head_skb->dev->stats.rx_dropped++;
 326                                 return -ENOMEM;
 327                         }
 328                         if (curr_skb == head_skb)
 329                                 skb_shinfo(curr_skb)->frag_list = nskb;
 330                         else
 331                                 curr_skb->next = nskb;
 332                         curr_skb = nskb;
 333                         head_skb->truesize += nskb->truesize;
 334                         num_skb_frags = 0;
 335                 }
 336                 if (curr_skb != head_skb) {
 337                         head_skb->data_len += len;
 338                         head_skb->len += len;
 339                         head_skb->truesize += MAX_PACKET_LEN;
 340                 }
 341                 page = virt_to_head_page(buf);
 342                 offset = buf - (char *)page_address(page);
 343                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 344                         put_page(page);
 345                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 346                                              len, MAX_PACKET_LEN);
 347                 } else {
 348                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 349                                         offset, len,
 350                                         MAX_PACKET_LEN);
 351                 }
 352                 --rq->num;
 353         }
 354         return 0;
 355 }
 356
 357 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
 358 {
 359         struct virtnet_info *vi = rq->vq->vdev->priv;
 360         struct net_device *dev = vi->dev;
 361         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 362         struct sk_buff *skb;
 363         struct page *page;
 364         struct skb_vnet_hdr *hdr;
 365
 366         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 367                 pr_debug("%s: short packet %i\n", dev->name, len);
 368                 dev->stats.rx_length_errors++;
 369                 if (vi->big_packets)
 370                         give_pages(rq, buf);
 371                 else if (vi->mergeable_rx_bufs)
 372                         put_page(virt_to_head_page(buf));
 373                 else
 374                         dev_kfree_skb(buf);
 375                 return;
 376         }
 377
 378         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
 379                 skb = buf;
 380                 len -= sizeof(struct virtio_net_hdr);
 381                 skb_trim(skb, len);
 382         } else if (vi->mergeable_rx_bufs) {
 383                 struct page *page = virt_to_head_page(buf);
 384                 skb = page_to_skb(rq, page,
 385                                   (char *)buf - (char *)page_address(page),
 386                                   len, MAX_PACKET_LEN);
 387                 if (unlikely(!skb)) {
 388                         dev->stats.rx_dropped++;
 389                         put_page(page);
 390                         return;
 391                 }
 392                 if (receive_mergeable(rq, skb)) {
 393                         dev_kfree_skb(skb);
 394                         return;
 395                 }
 396         } else {
 397                 page = buf;
 398                 skb = page_to_skb(rq, page, 0, len, PAGE_SIZE);
 399                 if (unlikely(!skb)) {
 400                         dev->stats.rx_dropped++;
 401                         give_pages(rq, page);
 402                         return;
 403                 }
 404         }
 405
 406         hdr = skb_vnet_hdr(skb);
 407
 408         u64_stats_update_begin(&stats->rx_syncp);
 409         stats->rx_bytes += skb->len;
 410         stats->rx_packets++;
 411         u64_stats_update_end(&stats->rx_syncp);
 412
 413         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 414                 pr_debug("Needs csum!\n");
 415                 if (!skb_partial_csum_set(skb,
 416                                           hdr->hdr.csum_start,
 417                                           hdr->hdr.csum_offset))
 418                         goto frame_err;
 419         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 420                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 421         }
 422
 423         skb->protocol = eth_type_trans(skb, dev);
 424         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 425                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 426
 427         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 428                 pr_debug("GSO!\n");
 429                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 430                 case VIRTIO_NET_HDR_GSO_TCPV4:
 431                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 432                         break;
 433                 case VIRTIO_NET_HDR_GSO_UDP:
 434                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 435                         break;
 436                 case VIRTIO_NET_HDR_GSO_TCPV6:
 437                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 438                         break;
 439                 default:
 440                         net_warn_ratelimited("%s: bad gso type %u.\n",
 441                                              dev->name, hdr->hdr.gso_type);
 442                         goto frame_err;
 443                 }
 444
 445                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 446                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 447
 448                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 449                 if (skb_shinfo(skb)->gso_size == 0) {
 450                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 451                         goto frame_err;
 452                 }
 453
 454                 /* Header must be checked, and gso_segs computed. */
 455                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 456                 skb_shinfo(skb)->gso_segs = 0;
 457         }
 458
 459         netif_receive_skb(skb);
 460         return;
 461
 462 frame_err:
 463         dev->stats.rx_frame_errors++;
 464         dev_kfree_skb(skb);
 465 }
 466
 467 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
 468 {
 469         struct virtnet_info *vi = rq->vq->vdev->priv;
 470         struct sk_buff *skb;
 471         struct skb_vnet_hdr *hdr;
 472         int err;
 473
 474         skb = __netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN, gfp);
 475         if (unlikely(!skb))
 476                 return -ENOMEM;
 477
 478         skb_put(skb, MAX_PACKET_LEN);
 479
 480         hdr = skb_vnet_hdr(skb);
 481         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
 482
 483         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 484
 485         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 486         if (err < 0)
 487                 dev_kfree_skb(skb);
 488
 489         return err;
 490 }
 491
 492 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
 493 {
 494         struct page *first, *list = NULL;
 495         char *p;
 496         int i, err, offset;
 497
 498         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 499         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 500                 first = get_a_page(rq, gfp);
 501                 if (!first) {
 502                         if (list)
 503                                 give_pages(rq, list);
 504                         return -ENOMEM;
 505                 }
 506                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 507
 508                 /* chain new page in list head to match sg */
 509                 first->private = (unsigned long)list;
 510                 list = first;
 511         }
 512
 513         first = get_a_page(rq, gfp);
 514         if (!first) {
 515                 give_pages(rq, list);
 516                 return -ENOMEM;
 517         }
 518         p = page_address(first);
 519
 520         /* rq->sg[0], rq->sg[1] share the same page */
 521         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
 522         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
 523
 524         /* rq->sg[1] for data packet, from offset */
 525         offset = sizeof(struct padded_vnet_hdr);
 526         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 527
 528         /* chain first in list head */
 529         first->private = (unsigned long)list;
 530         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 531                                   first, gfp);
 532         if (err < 0)
 533                 give_pages(rq, first);
 534
 535         return err;
 536 }
 537
 538 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 539 {
 540         struct virtnet_info *vi = rq->vq->vdev->priv;
 541         char *buf = NULL;
 542         int err;
 543
 544         if (gfp & __GFP_WAIT) {
 545                 if (skb_page_frag_refill(MAX_PACKET_LEN, &vi->alloc_frag,
 546                                          gfp)) {
 547                         buf = (char *)page_address(vi->alloc_frag.page) +
 548                               vi->alloc_frag.offset;
 549                         get_page(vi->alloc_frag.page);
 550                         vi->alloc_frag.offset += MAX_PACKET_LEN;
 551                 }
 552         } else {
 553                 buf = netdev_alloc_frag(MAX_PACKET_LEN);
 554         }
 555         if (!buf)
 556                 return -ENOMEM;
 557
 558         sg_init_one(rq->sg, buf, MAX_PACKET_LEN);
 559         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
 560         if (err < 0)
 561                 put_page(virt_to_head_page(buf));
 562
 563         return err;
 564 }
 565
 566 /*
 567  * Returns false if we couldn't fill entirely (OOM).
 568  *
 569  * Normally run in the receive path, but can also be run from ndo_open
 570  * before we're receiving packets, or from refill_work which is
 571  * careful to disable receiving (using napi_disable).
 572  */
 573 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
 574 {
 575         struct virtnet_info *vi = rq->vq->vdev->priv;
 576         int err;
 577         bool oom;
 578
 579         do {
 580                 if (vi->mergeable_rx_bufs)
 581                         err = add_recvbuf_mergeable(rq, gfp);
 582                 else if (vi->big_packets)
 583                         err = add_recvbuf_big(rq, gfp);
 584                 else
 585                         err = add_recvbuf_small(rq, gfp);
 586
 587                 oom = err == -ENOMEM;
 588                 if (err)
 589                         break;
 590                 ++rq->num;
 591         } while (rq->vq->num_free);
 592         if (unlikely(rq->num > rq->max))
 593                 rq->max = rq->num;
 594         if (unlikely(!virtqueue_kick(rq->vq)))
 595                 return false;
 596         return !oom;
 597 }
 598
 599 static void skb_recv_done(struct virtqueue *rvq)
 600 {
 601         struct virtnet_info *vi = rvq->vdev->priv;
 602         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 603
 604         /* Schedule NAPI, Suppress further interrupts if successful. */
 605         if (napi_schedule_prep(&rq->napi)) {
 606                 virtqueue_disable_cb(rvq);
 607                 __napi_schedule(&rq->napi);
 608         }
 609 }
 610
 611 static void virtnet_napi_enable(struct receive_queue *rq)
 612 {
 613         napi_enable(&rq->napi);
 614
 615         /* If all buffers were filled by other side before we napi_enabled, we
 616          * won't get another interrupt, so process any outstanding packets
 617          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 618          * We synchronize against interrupts via NAPI_STATE_SCHED */
 619         if (napi_schedule_prep(&rq->napi)) {
 620                 virtqueue_disable_cb(rq->vq);
 621                 local_bh_disable();
 622                 __napi_schedule(&rq->napi);
 623                 local_bh_enable();
 624         }
 625 }
 626
 627 static void refill_work(struct work_struct *work)
 628 {
 629         struct virtnet_info *vi =
 630                 container_of(work, struct virtnet_info, refill.work);
 631         bool still_empty;
 632         int i;
 633
 634         for (i = 0; i < vi->curr_queue_pairs; i++) {
 635                 struct receive_queue *rq = &vi->rq[i];
 636
 637                 napi_disable(&rq->napi);
 638                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 639                 virtnet_napi_enable(rq);
 640
 641                 /* In theory, this can happen: if we don't get any buffers in
 642                  * we will *never* try to fill again.
 643                  */
 644                 if (still_empty)
 645                         schedule_delayed_work(&vi->refill, HZ/2);
 646         }
 647 }
 648
 649 static int virtnet_poll(struct napi_struct *napi, int budget)
 650 {
 651         struct receive_queue *rq =
 652                 container_of(napi, struct receive_queue, napi);
 653         struct virtnet_info *vi = rq->vq->vdev->priv;
 654         void *buf;
 655         unsigned int r, len, received = 0;
 656
 657 again:
 658         while (received < budget &&
 659                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 660                 receive_buf(rq, buf, len);
 661                 --rq->num;
 662                 received++;
 663         }
 664
 665         if (rq->num < rq->max / 2) {
 666                 if (!try_fill_recv(rq, GFP_ATOMIC))
 667                         schedule_delayed_work(&vi->refill, 0);
 668         }
 669
 670         /* Out of packets? */
 671         if (received < budget) {
 672                 r = virtqueue_enable_cb_prepare(rq->vq);
 673                 napi_complete(napi);
 674                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 675                     napi_schedule_prep(napi)) {
 676                         virtqueue_disable_cb(rq->vq);
 677                         __napi_schedule(napi);
 678                         goto again;
 679                 }
 680         }
 681
 682         return received;
 683 }
 684
 685 static int virtnet_open(struct net_device *dev)
 686 {
 687         struct virtnet_info *vi = netdev_priv(dev);
 688         int i;
 689
 690         for (i = 0; i < vi->max_queue_pairs; i++) {
 691                 if (i < vi->curr_queue_pairs)
 692                         /* Make sure we have some buffers: if oom use wq. */
 693                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 694                                 schedule_delayed_work(&vi->refill, 0);
 695                 virtnet_napi_enable(&vi->rq[i]);
 696         }
 697
 698         return 0;
 699 }
 700
 701 static void free_old_xmit_skbs(struct send_queue *sq)
 702 {
 703         struct sk_buff *skb;
 704         unsigned int len;
 705         struct virtnet_info *vi = sq->vq->vdev->priv;
 706         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 707
 708         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 709                 pr_debug("Sent skb %p\n", skb);
 710
 711                 u64_stats_update_begin(&stats->tx_syncp);
 712                 stats->tx_bytes += skb->len;
 713                 stats->tx_packets++;
 714                 u64_stats_update_end(&stats->tx_syncp);
 715
 716                 dev_kfree_skb_any(skb);
 717         }
 718 }
 719
 720 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 721 {
 722         struct skb_vnet_hdr *hdr;
 723         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 724         struct virtnet_info *vi = sq->vq->vdev->priv;
 725         unsigned num_sg;
 726         unsigned hdr_len;
 727         bool can_push;
 728
 729         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 730         if (vi->mergeable_rx_bufs)
 731                 hdr_len = sizeof hdr->mhdr;
 732         else
 733                 hdr_len = sizeof hdr->hdr;
 734
 735         can_push = vi->any_header_sg &&
 736                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 737                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 738         /* Even if we can, don't push here yet as this would skew
 739          * csum_start offset below. */
 740         if (can_push)
 741                 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
 742         else
 743                 hdr = skb_vnet_hdr(skb);
 744
 745         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 746                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 747                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 748                 hdr->hdr.csum_offset = skb->csum_offset;
 749         } else {
 750                 hdr->hdr.flags = 0;
 751                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 752         }
 753
 754         if (skb_is_gso(skb)) {
 755                 hdr->hdr.hdr_len = skb_headlen(skb);
 756                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 757                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 758                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 759                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 760                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 761                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 762                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 763                 else
 764                         BUG();
 765                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 766                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 767         } else {
 768                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 769                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 770         }
 771
 772         if (vi->mergeable_rx_bufs)
 773                 hdr->mhdr.num_buffers = 0;
 774
 775         if (can_push) {
 776                 __skb_push(skb, hdr_len);
 777                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 778                 /* Pull header back to avoid skew in tx bytes calculations. */
 779                 __skb_pull(skb, hdr_len);
 780         } else {
 781                 sg_set_buf(sq->sg, hdr, hdr_len);
 782                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 783         }
 784         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 785 }
 786
 787 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 788 {
 789         struct virtnet_info *vi = netdev_priv(dev);
 790         int qnum = skb_get_queue_mapping(skb);
 791         struct send_queue *sq = &vi->sq[qnum];
 792         int err;
 793
 794         /* Free up any pending old buffers before queueing new ones. */
 795         free_old_xmit_skbs(sq);
 796
 797         /* Try to transmit */
 798         err = xmit_skb(sq, skb);
 799
 800         /* This should not happen! */
 801         if (unlikely(err) || unlikely(!virtqueue_kick(sq->vq))) {
 802                 dev->stats.tx_fifo_errors++;
 803                 if (net_ratelimit())
 804                         dev_warn(&dev->dev,
 805                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 806                 dev->stats.tx_dropped++;
 807                 kfree_skb(skb);
 808                 return NETDEV_TX_OK;
 809         }
 810
 811         /* Don't wait up for transmitted skbs to be freed. */
 812         skb_orphan(skb);
 813         nf_reset(skb);
 814
 815         /* Apparently nice girls don't return TX_BUSY; stop the queue
 816          * before it gets out of hand.  Naturally, this wastes entries. */
 817         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 818                 netif_stop_subqueue(dev, qnum);
 819                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 820                         /* More just got used, free them then recheck. */
 821                         free_old_xmit_skbs(sq);
 822                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 823                                 netif_start_subqueue(dev, qnum);
 824                                 virtqueue_disable_cb(sq->vq);
 825                         }
 826                 }
 827         }
 828
 829         return NETDEV_TX_OK;
 830 }
 831
 832 /*
 833  * Send command via the control virtqueue and check status.  Commands
 834  * supported by the hypervisor, as indicated by feature bits, should
 835  * never fail unless improperly formated.
 836  */
 837 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 838                                  struct scatterlist *out,
 839                                  struct scatterlist *in)
 840 {
 841         struct scatterlist *sgs[4], hdr, stat;
 842         struct virtio_net_ctrl_hdr ctrl;
 843         virtio_net_ctrl_ack status = ~0;
 844         unsigned out_num = 0, in_num = 0, tmp;
 845
 846         /* Caller should know better */
 847         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 848
 849         ctrl.class = class;
 850         ctrl.cmd = cmd;
 851         /* Add header */
 852         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 853         sgs[out_num++] = &hdr;
 854
 855         if (out)
 856                 sgs[out_num++] = out;
 857         if (in)
 858                 sgs[out_num + in_num++] = in;
 859
 860         /* Add return status. */
 861         sg_init_one(&stat, &status, sizeof(status));
 862         sgs[out_num + in_num++] = &stat;
 863
 864         BUG_ON(out_num + in_num > ARRAY_SIZE(sgs));
 865         BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, in_num, vi, GFP_ATOMIC)
 866                < 0);
 867
 868         if (unlikely(!virtqueue_kick(vi->cvq)))
 869                 return status == VIRTIO_NET_OK;
 870
 871         /* Spin for a response, the kick causes an ioport write, trapping
 872          * into the hypervisor, so the request should be handled immediately.
 873          */
 874         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
 875                !virtqueue_is_broken(vi->cvq))
 876                 cpu_relax();
 877
 878         return status == VIRTIO_NET_OK;
 879 }
 880
 881 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 882 {
 883         struct virtnet_info *vi = netdev_priv(dev);
 884         struct virtio_device *vdev = vi->vdev;
 885         int ret;
 886         struct sockaddr *addr = p;
 887         struct scatterlist sg;
 888
 889         ret = eth_prepare_mac_addr_change(dev, p);
 890         if (ret)
 891                 return ret;
 892
 893         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 894                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 895                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 896                                           VIRTIO_NET_CTRL_MAC_ADDR_SET,
 897                                           &sg, NULL)) {
 898                         dev_warn(&vdev->dev,
 899                                  "Failed to set mac address by vq command.\n");
 900                         return -EINVAL;
 901                 }
 902         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 903                 unsigned int i;
 904
 905                 /* Naturally, this has an atomicity problem. */
 906                 for (i = 0; i < dev->addr_len; i++)
 907                         virtio_cwrite8(vdev,
 908                                        offsetof(struct virtio_net_config, mac) +
 909                                        i, addr->sa_data[i]);
 910         }
 911
 912         eth_commit_mac_addr_change(dev, p);
 913
 914         return 0;
 915 }
 916
 917 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 918                                                struct rtnl_link_stats64 *tot)
 919 {
 920         struct virtnet_info *vi = netdev_priv(dev);
 921         int cpu;
 922         unsigned int start;
 923
 924         for_each_possible_cpu(cpu) {
 925                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 926                 u64 tpackets, tbytes, rpackets, rbytes;
 927
 928                 do {
 929                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 930                         tpackets = stats->tx_packets;
 931                         tbytes   = stats->tx_bytes;
 932                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 933
 934                 do {
 935                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 936                         rpackets = stats->rx_packets;
 937                         rbytes   = stats->rx_bytes;
 938                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 939
 940                 tot->rx_packets += rpackets;
 941                 tot->tx_packets += tpackets;
 942                 tot->rx_bytes   += rbytes;
 943                 tot->tx_bytes   += tbytes;
 944         }
 945
 946         tot->tx_dropped = dev->stats.tx_dropped;
 947         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 948         tot->rx_dropped = dev->stats.rx_dropped;
 949         tot->rx_length_errors = dev->stats.rx_length_errors;
 950         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 951
 952         return tot;
 953 }
 954
 955 #ifdef CONFIG_NET_POLL_CONTROLLER
 956 static void virtnet_netpoll(struct net_device *dev)
 957 {
 958         struct virtnet_info *vi = netdev_priv(dev);
 959         int i;
 960
 961         for (i = 0; i < vi->curr_queue_pairs; i++)
 962                 napi_schedule(&vi->rq[i].napi);
 963 }
 964 #endif
 965
 966 static void virtnet_ack_link_announce(struct virtnet_info *vi)
 967 {
 968         rtnl_lock();
 969         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
 970                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
 971                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
 972         rtnl_unlock();
 973 }
 974
 975 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
 976 {
 977         struct scatterlist sg;
 978         struct virtio_net_ctrl_mq s;
 979         struct net_device *dev = vi->dev;
 980
 981         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
 982                 return 0;
 983
 984         s.virtqueue_pairs = queue_pairs;
 985         sg_init_one(&sg, &s, sizeof(s));
 986
 987         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
 988                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
 989                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
 990                          queue_pairs);
 991                 return -EINVAL;
 992         } else {
 993                 vi->curr_queue_pairs = queue_pairs;
 994                 /* virtnet_open() will refill when device is going to up. */
 995                 if (dev->flags & IFF_UP)
 996                         schedule_delayed_work(&vi->refill, 0);
 997         }
 998
 999         return 0;
1000 }
1001
1002 static int virtnet_close(struct net_device *dev)
1003 {
1004         struct virtnet_info *vi = netdev_priv(dev);
1005         int i;
1006
1007         /* Make sure refill_work doesn't re-enable napi! */
1008         cancel_delayed_work_sync(&vi->refill);
1009
1010         for (i = 0; i < vi->max_queue_pairs; i++)
1011                 napi_disable(&vi->rq[i].napi);
1012
1013         return 0;
1014 }
1015
1016 static void virtnet_set_rx_mode(struct net_device *dev)
1017 {
1018         struct virtnet_info *vi = netdev_priv(dev);
1019         struct scatterlist sg[2];
1020         u8 promisc, allmulti;
1021         struct virtio_net_ctrl_mac *mac_data;
1022         struct netdev_hw_addr *ha;
1023         int uc_count;
1024         int mc_count;
1025         void *buf;
1026         int i;
1027
1028         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
1029         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1030                 return;
1031
1032         promisc = ((dev->flags & IFF_PROMISC) != 0);
1033         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1034
1035         sg_init_one(sg, &promisc, sizeof(promisc));
1036
1037         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1038                                   VIRTIO_NET_CTRL_RX_PROMISC,
1039                                   sg, NULL))
1040                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1041                          promisc ? "en" : "dis");
1042
1043         sg_init_one(sg, &allmulti, sizeof(allmulti));
1044
1045         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1046                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
1047                                   sg, NULL))
1048                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1049                          allmulti ? "en" : "dis");
1050
1051         uc_count = netdev_uc_count(dev);
1052         mc_count = netdev_mc_count(dev);
1053         /* MAC filter - use one buffer for both lists */
1054         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1055                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1056         mac_data = buf;
1057         if (!buf)
1058                 return;
1059
1060         sg_init_table(sg, 2);
1061
1062         /* Store the unicast list and count in the front of the buffer */
1063         mac_data->entries = uc_count;
1064         i = 0;
1065         netdev_for_each_uc_addr(ha, dev)
1066                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1067
1068         sg_set_buf(&sg[0], mac_data,
1069                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1070
1071         /* multicast list and count fill the end */
1072         mac_data = (void *)&mac_data->macs[uc_count][0];
1073
1074         mac_data->entries = mc_count;
1075         i = 0;
1076         netdev_for_each_mc_addr(ha, dev)
1077                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1078
1079         sg_set_buf(&sg[1], mac_data,
1080                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1081
1082         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1083                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
1084                                   sg, NULL))
1085                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1086
1087         kfree(buf);
1088 }
1089
1090 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1091                                    __be16 proto, u16 vid)
1092 {
1093         struct virtnet_info *vi = netdev_priv(dev);
1094         struct scatterlist sg;
1095
1096         sg_init_one(&sg, &vid, sizeof(vid));
1097
1098         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1099                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1100                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1101         return 0;
1102 }
1103
1104 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1105                                     __be16 proto, u16 vid)
1106 {
1107         struct virtnet_info *vi = netdev_priv(dev);
1108         struct scatterlist sg;
1109
1110         sg_init_one(&sg, &vid, sizeof(vid));
1111
1112         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1113                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1114                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1115         return 0;
1116 }
1117
1118 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1119 {
1120         int i;
1121
1122         if (vi->affinity_hint_set) {
1123                 for (i = 0; i < vi->max_queue_pairs; i++) {
1124                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1125                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1126                 }
1127
1128                 vi->affinity_hint_set = false;
1129         }
1130 }
1131
1132 static void virtnet_set_affinity(struct virtnet_info *vi)
1133 {
1134         int i;
1135         int cpu;
1136
1137         /* In multiqueue mode, when the number of cpu is equal to the number of
1138          * queue pairs, we let the queue pairs to be private to one cpu by
1139          * setting the affinity hint to eliminate the contention.
1140          */
1141         if (vi->curr_queue_pairs == 1 ||
1142             vi->max_queue_pairs != num_online_cpus()) {
1143                 virtnet_clean_affinity(vi, -1);
1144                 return;
1145         }
1146
1147         i = 0;
1148         for_each_online_cpu(cpu) {
1149                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1150                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1151                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1152                 i++;
1153         }
1154
1155         vi->affinity_hint_set = true;
1156 }
1157
1158 static int virtnet_cpu_callback(struct notifier_block *nfb,
1159                                 unsigned long action, void *hcpu)
1160 {
1161         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1162
1163         switch(action & ~CPU_TASKS_FROZEN) {
1164         case CPU_ONLINE:
1165         case CPU_DOWN_FAILED:
1166         case CPU_DEAD:
1167                 virtnet_set_affinity(vi);
1168                 break;
1169         case CPU_DOWN_PREPARE:
1170                 virtnet_clean_affinity(vi, (long)hcpu);
1171                 break;
1172         default:
1173                 break;
1174         }
1175
1176         return NOTIFY_OK;
1177 }
1178
1179 static void virtnet_get_ringparam(struct net_device *dev,
1180                                 struct ethtool_ringparam *ring)
1181 {
1182         struct virtnet_info *vi = netdev_priv(dev);
1183
1184         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1185         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1186         ring->rx_pending = ring->rx_max_pending;
1187         ring->tx_pending = ring->tx_max_pending;
1188 }
1189
1190
1191 static void virtnet_get_drvinfo(struct net_device *dev,
1192                                 struct ethtool_drvinfo *info)
1193 {
1194         struct virtnet_info *vi = netdev_priv(dev);
1195         struct virtio_device *vdev = vi->vdev;
1196
1197         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1198         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1199         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1200
1201 }
1202
1203 /* TODO: Eliminate OOO packets during switching */
1204 static int virtnet_set_channels(struct net_device *dev,
1205                                 struct ethtool_channels *channels)
1206 {
1207         struct virtnet_info *vi = netdev_priv(dev);
1208         u16 queue_pairs = channels->combined_count;
1209         int err;
1210
1211         /* We don't support separate rx/tx channels.
1212          * We don't allow setting 'other' channels.
1213          */
1214         if (channels->rx_count || channels->tx_count || channels->other_count)
1215                 return -EINVAL;
1216
1217         if (queue_pairs > vi->max_queue_pairs)
1218                 return -EINVAL;
1219
1220         get_online_cpus();
1221         err = virtnet_set_queues(vi, queue_pairs);
1222         if (!err) {
1223                 netif_set_real_num_tx_queues(dev, queue_pairs);
1224                 netif_set_real_num_rx_queues(dev, queue_pairs);
1225
1226                 virtnet_set_affinity(vi);
1227         }
1228         put_online_cpus();
1229
1230         return err;
1231 }
1232
1233 static void virtnet_get_channels(struct net_device *dev,
1234                                  struct ethtool_channels *channels)
1235 {
1236         struct virtnet_info *vi = netdev_priv(dev);
1237
1238         channels->combined_count = vi->curr_queue_pairs;
1239         channels->max_combined = vi->max_queue_pairs;
1240         channels->max_other = 0;
1241         channels->rx_count = 0;
1242         channels->tx_count = 0;
1243         channels->other_count = 0;
1244 }
1245
1246 static const struct ethtool_ops virtnet_ethtool_ops = {
1247         .get_drvinfo = virtnet_get_drvinfo,
1248         .get_link = ethtool_op_get_link,
1249         .get_ringparam = virtnet_get_ringparam,
1250         .set_channels = virtnet_set_channels,
1251         .get_channels = virtnet_get_channels,
1252 };
1253
1254 #define MIN_MTU 68
1255 #define MAX_MTU 65535
1256
1257 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1258 {
1259         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1260                 return -EINVAL;
1261         dev->mtu = new_mtu;
1262         return 0;
1263 }
1264
1265 static const struct net_device_ops virtnet_netdev = {
1266         .ndo_open            = virtnet_open,
1267         .ndo_stop            = virtnet_close,
1268         .ndo_start_xmit      = start_xmit,
1269         .ndo_validate_addr   = eth_validate_addr,
1270         .ndo_set_mac_address = virtnet_set_mac_address,
1271         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1272         .ndo_change_mtu      = virtnet_change_mtu,
1273         .ndo_get_stats64     = virtnet_stats,
1274         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1275         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1276 #ifdef CONFIG_NET_POLL_CONTROLLER
1277         .ndo_poll_controller = virtnet_netpoll,
1278 #endif
1279 };
1280
1281 static void virtnet_config_changed_work(struct work_struct *work)
1282 {
1283         struct virtnet_info *vi =
1284                 container_of(work, struct virtnet_info, config_work);
1285         u16 v;
1286
1287         mutex_lock(&vi->config_lock);
1288         if (!vi->config_enable)
1289                 goto done;
1290
1291         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1292                                  struct virtio_net_config, status, &v) < 0)
1293                 goto done;
1294
1295         if (v & VIRTIO_NET_S_ANNOUNCE) {
1296                 netdev_notify_peers(vi->dev);
1297                 virtnet_ack_link_announce(vi);
1298         }
1299
1300         /* Ignore unknown (future) status bits */
1301         v &= VIRTIO_NET_S_LINK_UP;
1302
1303         if (vi->status == v)
1304                 goto done;
1305
1306         vi->status = v;
1307
1308         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1309                 netif_carrier_on(vi->dev);
1310                 netif_tx_wake_all_queues(vi->dev);
1311         } else {
1312                 netif_carrier_off(vi->dev);
1313                 netif_tx_stop_all_queues(vi->dev);
1314         }
1315 done:
1316         mutex_unlock(&vi->config_lock);
1317 }
1318
1319 static void virtnet_config_changed(struct virtio_device *vdev)
1320 {
1321         struct virtnet_info *vi = vdev->priv;
1322
1323         schedule_work(&vi->config_work);
1324 }
1325
1326 static void virtnet_free_queues(struct virtnet_info *vi)
1327 {
1328         kfree(vi->rq);
1329         kfree(vi->sq);
1330 }
1331
1332 static void free_receive_bufs(struct virtnet_info *vi)
1333 {
1334         int i;
1335
1336         for (i = 0; i < vi->max_queue_pairs; i++) {
1337                 while (vi->rq[i].pages)
1338                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1339         }
1340 }
1341
1342 static void free_unused_bufs(struct virtnet_info *vi)
1343 {
1344         void *buf;
1345         int i;
1346
1347         for (i = 0; i < vi->max_queue_pairs; i++) {
1348                 struct virtqueue *vq = vi->sq[i].vq;
1349                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1350                         dev_kfree_skb(buf);
1351         }
1352
1353         for (i = 0; i < vi->max_queue_pairs; i++) {
1354                 struct virtqueue *vq = vi->rq[i].vq;
1355
1356                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1357                         if (vi->big_packets)
1358                                 give_pages(&vi->rq[i], buf);
1359                         else if (vi->mergeable_rx_bufs)
1360                                 put_page(virt_to_head_page(buf));
1361                         else
1362                                 dev_kfree_skb(buf);
1363                         --vi->rq[i].num;
1364                 }
1365                 BUG_ON(vi->rq[i].num != 0);
1366         }
1367 }
1368
1369 static void virtnet_del_vqs(struct virtnet_info *vi)
1370 {
1371         struct virtio_device *vdev = vi->vdev;
1372
1373         virtnet_clean_affinity(vi, -1);
1374
1375         vdev->config->del_vqs(vdev);
1376
1377         virtnet_free_queues(vi);
1378 }
1379
1380 static int virtnet_find_vqs(struct virtnet_info *vi)
1381 {
1382         vq_callback_t **callbacks;
1383         struct virtqueue **vqs;
1384         int ret = -ENOMEM;
1385         int i, total_vqs;
1386         const char **names;
1387
1388         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1389          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1390          * possible control vq.
1391          */
1392         total_vqs = vi->max_queue_pairs * 2 +
1393                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1394
1395         /* Allocate space for find_vqs parameters */
1396         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1397         if (!vqs)
1398                 goto err_vq;
1399         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1400         if (!callbacks)
1401                 goto err_callback;
1402         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1403         if (!names)
1404                 goto err_names;
1405
1406         /* Parameters for control virtqueue, if any */
1407         if (vi->has_cvq) {
1408                 callbacks[total_vqs - 1] = NULL;
1409                 names[total_vqs - 1] = "control";
1410         }
1411
1412         /* Allocate/initialize parameters for send/receive virtqueues */
1413         for (i = 0; i < vi->max_queue_pairs; i++) {
1414                 callbacks[rxq2vq(i)] = skb_recv_done;
1415                 callbacks[txq2vq(i)] = skb_xmit_done;
1416                 sprintf(vi->rq[i].name, "input.%d", i);
1417                 sprintf(vi->sq[i].name, "output.%d", i);
1418                 names[rxq2vq(i)] = vi->rq[i].name;
1419                 names[txq2vq(i)] = vi->sq[i].name;
1420         }
1421
1422         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1423                                          names);
1424         if (ret)
1425                 goto err_find;
1426
1427         if (vi->has_cvq) {
1428                 vi->cvq = vqs[total_vqs - 1];
1429                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1430                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1431         }
1432
1433         for (i = 0; i < vi->max_queue_pairs; i++) {
1434                 vi->rq[i].vq = vqs[rxq2vq(i)];
1435                 vi->sq[i].vq = vqs[txq2vq(i)];
1436         }
1437
1438         kfree(names);
1439         kfree(callbacks);
1440         kfree(vqs);
1441
1442         return 0;
1443
1444 err_find:
1445         kfree(names);
1446 err_names:
1447         kfree(callbacks);
1448 err_callback:
1449         kfree(vqs);
1450 err_vq:
1451         return ret;
1452 }
1453
1454 static int virtnet_alloc_queues(struct virtnet_info *vi)
1455 {
1456         int i;
1457
1458         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1459         if (!vi->sq)
1460                 goto err_sq;
1461         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1462         if (!vi->rq)
1463                 goto err_rq;
1464
1465         INIT_DELAYED_WORK(&vi->refill, refill_work);
1466         for (i = 0; i < vi->max_queue_pairs; i++) {
1467                 vi->rq[i].pages = NULL;
1468                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1469                                napi_weight);
1470
1471                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1472                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1473         }
1474
1475         return 0;
1476
1477 err_rq:
1478         kfree(vi->sq);
1479 err_sq:
1480         return -ENOMEM;
1481 }
1482
1483 static int init_vqs(struct virtnet_info *vi)
1484 {
1485         int ret;
1486
1487         /* Allocate send & receive queues */
1488         ret = virtnet_alloc_queues(vi);
1489         if (ret)
1490                 goto err;
1491
1492         ret = virtnet_find_vqs(vi);
1493         if (ret)
1494                 goto err_free;
1495
1496         get_online_cpus();
1497         virtnet_set_affinity(vi);
1498         put_online_cpus();
1499
1500         return 0;
1501
1502 err_free:
1503         virtnet_free_queues(vi);
1504 err:
1505         return ret;
1506 }
1507
1508 static int virtnet_probe(struct virtio_device *vdev)
1509 {
1510         int i, err;
1511         struct net_device *dev;
1512         struct virtnet_info *vi;
1513         u16 max_queue_pairs;
1514
1515         /* Find if host supports multiqueue virtio_net device */
1516         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1517                                    struct virtio_net_config,
1518                                    max_virtqueue_pairs, &max_queue_pairs);
1519
1520         /* We need at least 2 queue's */
1521         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1522             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1523             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1524                 max_queue_pairs = 1;
1525
1526         /* Allocate ourselves a network device with room for our info */
1527         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1528         if (!dev)
1529                 return -ENOMEM;
1530
1531         /* Set up network device as normal. */
1532         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1533         dev->netdev_ops = &virtnet_netdev;
1534         dev->features = NETIF_F_HIGHDMA;
1535
1536         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1537         SET_NETDEV_DEV(dev, &vdev->dev);
1538
1539         /* Do we support "hardware" checksums? */
1540         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1541                 /* This opens up the world of extra features. */
1542                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1543                 if (csum)
1544                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1545
1546                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1547                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1548                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1549                 }
1550                 /* Individual feature bits: what can host handle? */
1551                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1552                         dev->hw_features |= NETIF_F_TSO;
1553                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1554                         dev->hw_features |= NETIF_F_TSO6;
1555                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1556                         dev->hw_features |= NETIF_F_TSO_ECN;
1557                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1558                         dev->hw_features |= NETIF_F_UFO;
1559
1560                 if (gso)
1561                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1562                 /* (!csum && gso) case will be fixed by register_netdev() */
1563         }
1564         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1565                 dev->features |= NETIF_F_RXCSUM;
1566
1567         dev->vlan_features = dev->features;
1568
1569         /* Configuration may specify what MAC to use.  Otherwise random. */
1570         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1571                 virtio_cread_bytes(vdev,
1572                                    offsetof(struct virtio_net_config, mac),
1573                                    dev->dev_addr, dev->addr_len);
1574         else
1575                 eth_hw_addr_random(dev);
1576
1577         /* Set up our device-specific information */
1578         vi = netdev_priv(dev);
1579         vi->dev = dev;
1580         vi->vdev = vdev;
1581         vdev->priv = vi;
1582         vi->stats = alloc_percpu(struct virtnet_stats);
1583         err = -ENOMEM;
1584         if (vi->stats == NULL)
1585                 goto free;
1586
1587         for_each_possible_cpu(i) {
1588                 struct virtnet_stats *virtnet_stats;
1589                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1590                 u64_stats_init(&virtnet_stats->tx_syncp);
1591                 u64_stats_init(&virtnet_stats->rx_syncp);
1592         }
1593
1594         mutex_init(&vi->config_lock);
1595         vi->config_enable = true;
1596         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1597
1598         /* If we can receive ANY GSO packets, we must allocate large ones. */
1599         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1600             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1601             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1602                 vi->big_packets = true;
1603
1604         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1605                 vi->mergeable_rx_bufs = true;
1606
1607         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1608                 vi->any_header_sg = true;
1609
1610         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1611                 vi->has_cvq = true;
1612
1613         /* Use single tx/rx queue pair as default */
1614         vi->curr_queue_pairs = 1;
1615         vi->max_queue_pairs = max_queue_pairs;
1616
1617         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1618         err = init_vqs(vi);
1619         if (err)
1620                 goto free_stats;
1621
1622         netif_set_real_num_tx_queues(dev, 1);
1623         netif_set_real_num_rx_queues(dev, 1);
1624
1625         err = register_netdev(dev);
1626         if (err) {
1627                 pr_debug("virtio_net: registering device failed\n");
1628                 goto free_vqs;
1629         }
1630
1631         /* Last of all, set up some receive buffers. */
1632         for (i = 0; i < vi->curr_queue_pairs; i++) {
1633                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1634
1635                 /* If we didn't even get one input buffer, we're useless. */
1636                 if (vi->rq[i].num == 0) {
1637                         free_unused_bufs(vi);
1638                         err = -ENOMEM;
1639                         goto free_recv_bufs;
1640                 }
1641         }
1642
1643         vi->nb.notifier_call = &virtnet_cpu_callback;
1644         err = register_hotcpu_notifier(&vi->nb);
1645         if (err) {
1646                 pr_debug("virtio_net: registering cpu notifier failed\n");
1647                 goto free_recv_bufs;
1648         }
1649
1650         /* Assume link up if device can't report link status,
1651            otherwise get link status from config. */
1652         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1653                 netif_carrier_off(dev);
1654                 schedule_work(&vi->config_work);
1655         } else {
1656                 vi->status = VIRTIO_NET_S_LINK_UP;
1657                 netif_carrier_on(dev);
1658         }
1659
1660         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1661                  dev->name, max_queue_pairs);
1662
1663         return 0;
1664
1665 free_recv_bufs:
1666         free_receive_bufs(vi);
1667         unregister_netdev(dev);
1668 free_vqs:
1669         cancel_delayed_work_sync(&vi->refill);
1670         virtnet_del_vqs(vi);
1671         if (vi->alloc_frag.page)
1672                 put_page(vi->alloc_frag.page);
1673 free_stats:
1674         free_percpu(vi->stats);
1675 free:
1676         free_netdev(dev);
1677         return err;
1678 }
1679
1680 static void remove_vq_common(struct virtnet_info *vi)
1681 {
1682         vi->vdev->config->reset(vi->vdev);
1683
1684         /* Free unused buffers in both send and recv, if any. */
1685         free_unused_bufs(vi);
1686
1687         free_receive_bufs(vi);
1688
1689         virtnet_del_vqs(vi);
1690 }
1691
1692 static void virtnet_remove(struct virtio_device *vdev)
1693 {
1694         struct virtnet_info *vi = vdev->priv;
1695
1696         unregister_hotcpu_notifier(&vi->nb);
1697
1698         /* Prevent config work handler from accessing the device. */
1699         mutex_lock(&vi->config_lock);
1700         vi->config_enable = false;
1701         mutex_unlock(&vi->config_lock);
1702
1703         unregister_netdev(vi->dev);
1704
1705         remove_vq_common(vi);
1706         if (vi->alloc_frag.page)
1707                 put_page(vi->alloc_frag.page);
1708
1709         flush_work(&vi->config_work);
1710
1711         free_percpu(vi->stats);
1712         free_netdev(vi->dev);
1713 }
1714
1715 #ifdef CONFIG_PM_SLEEP
1716 static int virtnet_freeze(struct virtio_device *vdev)
1717 {
1718         struct virtnet_info *vi = vdev->priv;
1719         int i;
1720
1721         unregister_hotcpu_notifier(&vi->nb);
1722
1723         /* Prevent config work handler from accessing the device */
1724         mutex_lock(&vi->config_lock);
1725         vi->config_enable = false;
1726         mutex_unlock(&vi->config_lock);
1727
1728         netif_device_detach(vi->dev);
1729         cancel_delayed_work_sync(&vi->refill);
1730
1731         if (netif_running(vi->dev))
1732                 for (i = 0; i < vi->max_queue_pairs; i++) {
1733                         napi_disable(&vi->rq[i].napi);
1734                         netif_napi_del(&vi->rq[i].napi);
1735                 }
1736
1737         remove_vq_common(vi);
1738
1739         flush_work(&vi->config_work);
1740
1741         return 0;
1742 }
1743
1744 static int virtnet_restore(struct virtio_device *vdev)
1745 {
1746         struct virtnet_info *vi = vdev->priv;
1747         int err, i;
1748
1749         err = init_vqs(vi);
1750         if (err)
1751                 return err;
1752
1753         if (netif_running(vi->dev))
1754                 for (i = 0; i < vi->max_queue_pairs; i++)
1755                         virtnet_napi_enable(&vi->rq[i]);
1756
1757         netif_device_attach(vi->dev);
1758
1759         for (i = 0; i < vi->curr_queue_pairs; i++)
1760                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1761                         schedule_delayed_work(&vi->refill, 0);
1762
1763         mutex_lock(&vi->config_lock);
1764         vi->config_enable = true;
1765         mutex_unlock(&vi->config_lock);
1766
1767         rtnl_lock();
1768         virtnet_set_queues(vi, vi->curr_queue_pairs);
1769         rtnl_unlock();
1770
1771         err = register_hotcpu_notifier(&vi->nb);
1772         if (err)
1773                 return err;
1774
1775         return 0;
1776 }
1777 #endif
1778
1779 static struct virtio_device_id id_table[] = {
1780         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1781         { 0 },
1782 };
1783
1784 static unsigned int features[] = {
1785         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1786         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1787         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1788         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1789         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1790         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1791         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1792         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1793         VIRTIO_NET_F_CTRL_MAC_ADDR,
1794         VIRTIO_F_ANY_LAYOUT,
1795 };
1796
1797 static struct virtio_driver virtio_net_driver = {
1798         .feature_table = features,
1799         .feature_table_size = ARRAY_SIZE(features),
1800         .driver.name =  KBUILD_MODNAME,
1801         .driver.owner = THIS_MODULE,
1802         .id_table =     id_table,
1803         .probe =        virtnet_probe,
1804         .remove =       virtnet_remove,
1805         .config_changed = virtnet_config_changed,
1806 #ifdef CONFIG_PM_SLEEP
1807         .freeze =       virtnet_freeze,
1808         .restore =      virtnet_restore,
1809 #endif
1810 };
1811
1812 module_virtio_driver(virtio_net_driver);
1813
1814 MODULE_DEVICE_TABLE(virtio, id_table);
1815 MODULE_DESCRIPTION("Virtio network driver");
1816 MODULE_LICENSE("GPL");