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         virtqueue_kick(rq->vq);
 595         return !oom;
 596 }
 597
 598 static void skb_recv_done(struct virtqueue *rvq)
 599 {
 600         struct virtnet_info *vi = rvq->vdev->priv;
 601         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 602
 603         /* Schedule NAPI, Suppress further interrupts if successful. */
 604         if (napi_schedule_prep(&rq->napi)) {
 605                 virtqueue_disable_cb(rvq);
 606                 __napi_schedule(&rq->napi);
 607         }
 608 }
 609
 610 static void virtnet_napi_enable(struct receive_queue *rq)
 611 {
 612         napi_enable(&rq->napi);
 613
 614         /* If all buffers were filled by other side before we napi_enabled, we
 615          * won't get another interrupt, so process any outstanding packets
 616          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 617          * We synchronize against interrupts via NAPI_STATE_SCHED */
 618         if (napi_schedule_prep(&rq->napi)) {
 619                 virtqueue_disable_cb(rq->vq);
 620                 local_bh_disable();
 621                 __napi_schedule(&rq->napi);
 622                 local_bh_enable();
 623         }
 624 }
 625
 626 static void refill_work(struct work_struct *work)
 627 {
 628         struct virtnet_info *vi =
 629                 container_of(work, struct virtnet_info, refill.work);
 630         bool still_empty;
 631         int i;
 632
 633         for (i = 0; i < vi->curr_queue_pairs; i++) {
 634                 struct receive_queue *rq = &vi->rq[i];
 635
 636                 napi_disable(&rq->napi);
 637                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 638                 virtnet_napi_enable(rq);
 639
 640                 /* In theory, this can happen: if we don't get any buffers in
 641                  * we will *never* try to fill again.
 642                  */
 643                 if (still_empty)
 644                         schedule_delayed_work(&vi->refill, HZ/2);
 645         }
 646 }
 647
 648 static int virtnet_poll(struct napi_struct *napi, int budget)
 649 {
 650         struct receive_queue *rq =
 651                 container_of(napi, struct receive_queue, napi);
 652         struct virtnet_info *vi = rq->vq->vdev->priv;
 653         void *buf;
 654         unsigned int r, len, received = 0;
 655
 656 again:
 657         while (received < budget &&
 658                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 659                 receive_buf(rq, buf, len);
 660                 --rq->num;
 661                 received++;
 662         }
 663
 664         if (rq->num < rq->max / 2) {
 665                 if (!try_fill_recv(rq, GFP_ATOMIC))
 666                         schedule_delayed_work(&vi->refill, 0);
 667         }
 668
 669         /* Out of packets? */
 670         if (received < budget) {
 671                 r = virtqueue_enable_cb_prepare(rq->vq);
 672                 napi_complete(napi);
 673                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 674                     napi_schedule_prep(napi)) {
 675                         virtqueue_disable_cb(rq->vq);
 676                         __napi_schedule(napi);
 677                         goto again;
 678                 }
 679         }
 680
 681         return received;
 682 }
 683
 684 static int virtnet_open(struct net_device *dev)
 685 {
 686         struct virtnet_info *vi = netdev_priv(dev);
 687         int i;
 688
 689         for (i = 0; i < vi->max_queue_pairs; i++) {
 690                 if (i < vi->curr_queue_pairs)
 691                         /* Make sure we have some buffers: if oom use wq. */
 692                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 693                                 schedule_delayed_work(&vi->refill, 0);
 694                 virtnet_napi_enable(&vi->rq[i]);
 695         }
 696
 697         return 0;
 698 }
 699
 700 static void free_old_xmit_skbs(struct send_queue *sq)
 701 {
 702         struct sk_buff *skb;
 703         unsigned int len;
 704         struct virtnet_info *vi = sq->vq->vdev->priv;
 705         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 706
 707         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 708                 pr_debug("Sent skb %p\n", skb);
 709
 710                 u64_stats_update_begin(&stats->tx_syncp);
 711                 stats->tx_bytes += skb->len;
 712                 stats->tx_packets++;
 713                 u64_stats_update_end(&stats->tx_syncp);
 714
 715                 dev_kfree_skb_any(skb);
 716         }
 717 }
 718
 719 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 720 {
 721         struct skb_vnet_hdr *hdr;
 722         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 723         struct virtnet_info *vi = sq->vq->vdev->priv;
 724         unsigned num_sg;
 725         unsigned hdr_len;
 726         bool can_push;
 727
 728         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 729         if (vi->mergeable_rx_bufs)
 730                 hdr_len = sizeof hdr->mhdr;
 731         else
 732                 hdr_len = sizeof hdr->hdr;
 733
 734         can_push = vi->any_header_sg &&
 735                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 736                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 737         /* Even if we can, don't push here yet as this would skew
 738          * csum_start offset below. */
 739         if (can_push)
 740                 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
 741         else
 742                 hdr = skb_vnet_hdr(skb);
 743
 744         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 745                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 746                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 747                 hdr->hdr.csum_offset = skb->csum_offset;
 748         } else {
 749                 hdr->hdr.flags = 0;
 750                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 751         }
 752
 753         if (skb_is_gso(skb)) {
 754                 hdr->hdr.hdr_len = skb_headlen(skb);
 755                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 756                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 757                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 758                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 759                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 760                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 761                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 762                 else
 763                         BUG();
 764                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 765                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 766         } else {
 767                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 768                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 769         }
 770
 771         if (vi->mergeable_rx_bufs)
 772                 hdr->mhdr.num_buffers = 0;
 773
 774         if (can_push) {
 775                 __skb_push(skb, hdr_len);
 776                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 777                 /* Pull header back to avoid skew in tx bytes calculations. */
 778                 __skb_pull(skb, hdr_len);
 779         } else {
 780                 sg_set_buf(sq->sg, hdr, hdr_len);
 781                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 782         }
 783         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 784 }
 785
 786 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 787 {
 788         struct virtnet_info *vi = netdev_priv(dev);
 789         int qnum = skb_get_queue_mapping(skb);
 790         struct send_queue *sq = &vi->sq[qnum];
 791         int err;
 792
 793         /* Free up any pending old buffers before queueing new ones. */
 794         free_old_xmit_skbs(sq);
 795
 796         /* Try to transmit */
 797         err = xmit_skb(sq, skb);
 798
 799         /* This should not happen! */
 800         if (unlikely(err)) {
 801                 dev->stats.tx_fifo_errors++;
 802                 if (net_ratelimit())
 803                         dev_warn(&dev->dev,
 804                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 805                 dev->stats.tx_dropped++;
 806                 kfree_skb(skb);
 807                 return NETDEV_TX_OK;
 808         }
 809         virtqueue_kick(sq->vq);
 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         virtqueue_kick(vi->cvq);
 869
 870         /* Spin for a response, the kick causes an ioport write, trapping
 871          * into the hypervisor, so the request should be handled immediately.
 872          */
 873         while (!virtqueue_get_buf(vi->cvq, &tmp))
 874                 cpu_relax();
 875
 876         return status == VIRTIO_NET_OK;
 877 }
 878
 879 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 880 {
 881         struct virtnet_info *vi = netdev_priv(dev);
 882         struct virtio_device *vdev = vi->vdev;
 883         int ret;
 884         struct sockaddr *addr = p;
 885         struct scatterlist sg;
 886
 887         ret = eth_prepare_mac_addr_change(dev, p);
 888         if (ret)
 889                 return ret;
 890
 891         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 892                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 893                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 894                                           VIRTIO_NET_CTRL_MAC_ADDR_SET,
 895                                           &sg, NULL)) {
 896                         dev_warn(&vdev->dev,
 897                                  "Failed to set mac address by vq command.\n");
 898                         return -EINVAL;
 899                 }
 900         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 901                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 902                                   addr->sa_data, dev->addr_len);
 903         }
 904
 905         eth_commit_mac_addr_change(dev, p);
 906
 907         return 0;
 908 }
 909
 910 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 911                                                struct rtnl_link_stats64 *tot)
 912 {
 913         struct virtnet_info *vi = netdev_priv(dev);
 914         int cpu;
 915         unsigned int start;
 916
 917         for_each_possible_cpu(cpu) {
 918                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 919                 u64 tpackets, tbytes, rpackets, rbytes;
 920
 921                 do {
 922                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 923                         tpackets = stats->tx_packets;
 924                         tbytes   = stats->tx_bytes;
 925                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 926
 927                 do {
 928                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 929                         rpackets = stats->rx_packets;
 930                         rbytes   = stats->rx_bytes;
 931                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 932
 933                 tot->rx_packets += rpackets;
 934                 tot->tx_packets += tpackets;
 935                 tot->rx_bytes   += rbytes;
 936                 tot->tx_bytes   += tbytes;
 937         }
 938
 939         tot->tx_dropped = dev->stats.tx_dropped;
 940         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 941         tot->rx_dropped = dev->stats.rx_dropped;
 942         tot->rx_length_errors = dev->stats.rx_length_errors;
 943         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 944
 945         return tot;
 946 }
 947
 948 #ifdef CONFIG_NET_POLL_CONTROLLER
 949 static void virtnet_netpoll(struct net_device *dev)
 950 {
 951         struct virtnet_info *vi = netdev_priv(dev);
 952         int i;
 953
 954         for (i = 0; i < vi->curr_queue_pairs; i++)
 955                 napi_schedule(&vi->rq[i].napi);
 956 }
 957 #endif
 958
 959 static void virtnet_ack_link_announce(struct virtnet_info *vi)
 960 {
 961         rtnl_lock();
 962         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
 963                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
 964                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
 965         rtnl_unlock();
 966 }
 967
 968 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
 969 {
 970         struct scatterlist sg;
 971         struct virtio_net_ctrl_mq s;
 972         struct net_device *dev = vi->dev;
 973
 974         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
 975                 return 0;
 976
 977         s.virtqueue_pairs = queue_pairs;
 978         sg_init_one(&sg, &s, sizeof(s));
 979
 980         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
 981                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
 982                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
 983                          queue_pairs);
 984                 return -EINVAL;
 985         } else {
 986                 vi->curr_queue_pairs = queue_pairs;
 987                 /* virtnet_open() will refill when device is going to up. */
 988                 if (dev->flags & IFF_UP)
 989                         schedule_delayed_work(&vi->refill, 0);
 990         }
 991
 992         return 0;
 993 }
 994
 995 static int virtnet_close(struct net_device *dev)
 996 {
 997         struct virtnet_info *vi = netdev_priv(dev);
 998         int i;
 999
1000         /* Make sure refill_work doesn't re-enable napi! */
1001         cancel_delayed_work_sync(&vi->refill);
1002
1003         for (i = 0; i < vi->max_queue_pairs; i++)
1004                 napi_disable(&vi->rq[i].napi);
1005
1006         return 0;
1007 }
1008
1009 static void virtnet_set_rx_mode(struct net_device *dev)
1010 {
1011         struct virtnet_info *vi = netdev_priv(dev);
1012         struct scatterlist sg[2];
1013         u8 promisc, allmulti;
1014         struct virtio_net_ctrl_mac *mac_data;
1015         struct netdev_hw_addr *ha;
1016         int uc_count;
1017         int mc_count;
1018         void *buf;
1019         int i;
1020
1021         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
1022         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1023                 return;
1024
1025         promisc = ((dev->flags & IFF_PROMISC) != 0);
1026         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1027
1028         sg_init_one(sg, &promisc, sizeof(promisc));
1029
1030         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1031                                   VIRTIO_NET_CTRL_RX_PROMISC,
1032                                   sg, NULL))
1033                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1034                          promisc ? "en" : "dis");
1035
1036         sg_init_one(sg, &allmulti, sizeof(allmulti));
1037
1038         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1039                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
1040                                   sg, NULL))
1041                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1042                          allmulti ? "en" : "dis");
1043
1044         uc_count = netdev_uc_count(dev);
1045         mc_count = netdev_mc_count(dev);
1046         /* MAC filter - use one buffer for both lists */
1047         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1048                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1049         mac_data = buf;
1050         if (!buf)
1051                 return;
1052
1053         sg_init_table(sg, 2);
1054
1055         /* Store the unicast list and count in the front of the buffer */
1056         mac_data->entries = uc_count;
1057         i = 0;
1058         netdev_for_each_uc_addr(ha, dev)
1059                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1060
1061         sg_set_buf(&sg[0], mac_data,
1062                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1063
1064         /* multicast list and count fill the end */
1065         mac_data = (void *)&mac_data->macs[uc_count][0];
1066
1067         mac_data->entries = mc_count;
1068         i = 0;
1069         netdev_for_each_mc_addr(ha, dev)
1070                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1071
1072         sg_set_buf(&sg[1], mac_data,
1073                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1074
1075         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1076                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
1077                                   sg, NULL))
1078                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1079
1080         kfree(buf);
1081 }
1082
1083 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1084                                    __be16 proto, u16 vid)
1085 {
1086         struct virtnet_info *vi = netdev_priv(dev);
1087         struct scatterlist sg;
1088
1089         sg_init_one(&sg, &vid, sizeof(vid));
1090
1091         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1092                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1093                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1094         return 0;
1095 }
1096
1097 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1098                                     __be16 proto, u16 vid)
1099 {
1100         struct virtnet_info *vi = netdev_priv(dev);
1101         struct scatterlist sg;
1102
1103         sg_init_one(&sg, &vid, sizeof(vid));
1104
1105         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1106                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1107                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1108         return 0;
1109 }
1110
1111 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1112 {
1113         int i;
1114
1115         if (vi->affinity_hint_set) {
1116                 for (i = 0; i < vi->max_queue_pairs; i++) {
1117                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1118                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1119                 }
1120
1121                 vi->affinity_hint_set = false;
1122         }
1123 }
1124
1125 static void virtnet_set_affinity(struct virtnet_info *vi)
1126 {
1127         int i;
1128         int cpu;
1129
1130         /* In multiqueue mode, when the number of cpu is equal to the number of
1131          * queue pairs, we let the queue pairs to be private to one cpu by
1132          * setting the affinity hint to eliminate the contention.
1133          */
1134         if (vi->curr_queue_pairs == 1 ||
1135             vi->max_queue_pairs != num_online_cpus()) {
1136                 virtnet_clean_affinity(vi, -1);
1137                 return;
1138         }
1139
1140         i = 0;
1141         for_each_online_cpu(cpu) {
1142                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1143                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1144                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1145                 i++;
1146         }
1147
1148         vi->affinity_hint_set = true;
1149 }
1150
1151 static int virtnet_cpu_callback(struct notifier_block *nfb,
1152                                 unsigned long action, void *hcpu)
1153 {
1154         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1155
1156         switch(action & ~CPU_TASKS_FROZEN) {
1157         case CPU_ONLINE:
1158         case CPU_DOWN_FAILED:
1159         case CPU_DEAD:
1160                 virtnet_set_affinity(vi);
1161                 break;
1162         case CPU_DOWN_PREPARE:
1163                 virtnet_clean_affinity(vi, (long)hcpu);
1164                 break;
1165         default:
1166                 break;
1167         }
1168
1169         return NOTIFY_OK;
1170 }
1171
1172 static void virtnet_get_ringparam(struct net_device *dev,
1173                                 struct ethtool_ringparam *ring)
1174 {
1175         struct virtnet_info *vi = netdev_priv(dev);
1176
1177         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1178         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1179         ring->rx_pending = ring->rx_max_pending;
1180         ring->tx_pending = ring->tx_max_pending;
1181 }
1182
1183
1184 static void virtnet_get_drvinfo(struct net_device *dev,
1185                                 struct ethtool_drvinfo *info)
1186 {
1187         struct virtnet_info *vi = netdev_priv(dev);
1188         struct virtio_device *vdev = vi->vdev;
1189
1190         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1191         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1192         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1193
1194 }
1195
1196 /* TODO: Eliminate OOO packets during switching */
1197 static int virtnet_set_channels(struct net_device *dev,
1198                                 struct ethtool_channels *channels)
1199 {
1200         struct virtnet_info *vi = netdev_priv(dev);
1201         u16 queue_pairs = channels->combined_count;
1202         int err;
1203
1204         /* We don't support separate rx/tx channels.
1205          * We don't allow setting 'other' channels.
1206          */
1207         if (channels->rx_count || channels->tx_count || channels->other_count)
1208                 return -EINVAL;
1209
1210         if (queue_pairs > vi->max_queue_pairs)
1211                 return -EINVAL;
1212
1213         get_online_cpus();
1214         err = virtnet_set_queues(vi, queue_pairs);
1215         if (!err) {
1216                 netif_set_real_num_tx_queues(dev, queue_pairs);
1217                 netif_set_real_num_rx_queues(dev, queue_pairs);
1218
1219                 virtnet_set_affinity(vi);
1220         }
1221         put_online_cpus();
1222
1223         return err;
1224 }
1225
1226 static void virtnet_get_channels(struct net_device *dev,
1227                                  struct ethtool_channels *channels)
1228 {
1229         struct virtnet_info *vi = netdev_priv(dev);
1230
1231         channels->combined_count = vi->curr_queue_pairs;
1232         channels->max_combined = vi->max_queue_pairs;
1233         channels->max_other = 0;
1234         channels->rx_count = 0;
1235         channels->tx_count = 0;
1236         channels->other_count = 0;
1237 }
1238
1239 static const struct ethtool_ops virtnet_ethtool_ops = {
1240         .get_drvinfo = virtnet_get_drvinfo,
1241         .get_link = ethtool_op_get_link,
1242         .get_ringparam = virtnet_get_ringparam,
1243         .set_channels = virtnet_set_channels,
1244         .get_channels = virtnet_get_channels,
1245 };
1246
1247 #define MIN_MTU 68
1248 #define MAX_MTU 65535
1249
1250 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1251 {
1252         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1253                 return -EINVAL;
1254         dev->mtu = new_mtu;
1255         return 0;
1256 }
1257
1258 static const struct net_device_ops virtnet_netdev = {
1259         .ndo_open            = virtnet_open,
1260         .ndo_stop            = virtnet_close,
1261         .ndo_start_xmit      = start_xmit,
1262         .ndo_validate_addr   = eth_validate_addr,
1263         .ndo_set_mac_address = virtnet_set_mac_address,
1264         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1265         .ndo_change_mtu      = virtnet_change_mtu,
1266         .ndo_get_stats64     = virtnet_stats,
1267         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1268         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
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_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1285                               offsetof(struct virtio_net_config, status),
1286                               &v) < 0)
1287                 goto done;
1288
1289         if (v & VIRTIO_NET_S_ANNOUNCE) {
1290                 netdev_notify_peers(vi->dev);
1291                 virtnet_ack_link_announce(vi);
1292         }
1293
1294         /* Ignore unknown (future) status bits */
1295         v &= VIRTIO_NET_S_LINK_UP;
1296
1297         if (vi->status == v)
1298                 goto done;
1299
1300         vi->status = v;
1301
1302         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1303                 netif_carrier_on(vi->dev);
1304                 netif_tx_wake_all_queues(vi->dev);
1305         } else {
1306                 netif_carrier_off(vi->dev);
1307                 netif_tx_stop_all_queues(vi->dev);
1308         }
1309 done:
1310         mutex_unlock(&vi->config_lock);
1311 }
1312
1313 static void virtnet_config_changed(struct virtio_device *vdev)
1314 {
1315         struct virtnet_info *vi = vdev->priv;
1316
1317         schedule_work(&vi->config_work);
1318 }
1319
1320 static void virtnet_free_queues(struct virtnet_info *vi)
1321 {
1322         kfree(vi->rq);
1323         kfree(vi->sq);
1324 }
1325
1326 static void free_receive_bufs(struct virtnet_info *vi)
1327 {
1328         int i;
1329
1330         for (i = 0; i < vi->max_queue_pairs; i++) {
1331                 while (vi->rq[i].pages)
1332                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1333         }
1334 }
1335
1336 static void free_unused_bufs(struct virtnet_info *vi)
1337 {
1338         void *buf;
1339         int i;
1340
1341         for (i = 0; i < vi->max_queue_pairs; i++) {
1342                 struct virtqueue *vq = vi->sq[i].vq;
1343                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1344                         dev_kfree_skb(buf);
1345         }
1346
1347         for (i = 0; i < vi->max_queue_pairs; i++) {
1348                 struct virtqueue *vq = vi->rq[i].vq;
1349
1350                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1351                         if (vi->big_packets)
1352                                 give_pages(&vi->rq[i], buf);
1353                         else if (vi->mergeable_rx_bufs)
1354                                 put_page(virt_to_head_page(buf));
1355                         else
1356                                 dev_kfree_skb(buf);
1357                         --vi->rq[i].num;
1358                 }
1359                 BUG_ON(vi->rq[i].num != 0);
1360         }
1361 }
1362
1363 static void virtnet_del_vqs(struct virtnet_info *vi)
1364 {
1365         struct virtio_device *vdev = vi->vdev;
1366
1367         virtnet_clean_affinity(vi, -1);
1368
1369         vdev->config->del_vqs(vdev);
1370
1371         virtnet_free_queues(vi);
1372 }
1373
1374 static int virtnet_find_vqs(struct virtnet_info *vi)
1375 {
1376         vq_callback_t **callbacks;
1377         struct virtqueue **vqs;
1378         int ret = -ENOMEM;
1379         int i, total_vqs;
1380         const char **names;
1381
1382         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1383          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1384          * possible control vq.
1385          */
1386         total_vqs = vi->max_queue_pairs * 2 +
1387                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1388
1389         /* Allocate space for find_vqs parameters */
1390         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1391         if (!vqs)
1392                 goto err_vq;
1393         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1394         if (!callbacks)
1395                 goto err_callback;
1396         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1397         if (!names)
1398                 goto err_names;
1399
1400         /* Parameters for control virtqueue, if any */
1401         if (vi->has_cvq) {
1402                 callbacks[total_vqs - 1] = NULL;
1403                 names[total_vqs - 1] = "control";
1404         }
1405
1406         /* Allocate/initialize parameters for send/receive virtqueues */
1407         for (i = 0; i < vi->max_queue_pairs; i++) {
1408                 callbacks[rxq2vq(i)] = skb_recv_done;
1409                 callbacks[txq2vq(i)] = skb_xmit_done;
1410                 sprintf(vi->rq[i].name, "input.%d", i);
1411                 sprintf(vi->sq[i].name, "output.%d", i);
1412                 names[rxq2vq(i)] = vi->rq[i].name;
1413                 names[txq2vq(i)] = vi->sq[i].name;
1414         }
1415
1416         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1417                                          names);
1418         if (ret)
1419                 goto err_find;
1420
1421         if (vi->has_cvq) {
1422                 vi->cvq = vqs[total_vqs - 1];
1423                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1424                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1425         }
1426
1427         for (i = 0; i < vi->max_queue_pairs; i++) {
1428                 vi->rq[i].vq = vqs[rxq2vq(i)];
1429                 vi->sq[i].vq = vqs[txq2vq(i)];
1430         }
1431
1432         kfree(names);
1433         kfree(callbacks);
1434         kfree(vqs);
1435
1436         return 0;
1437
1438 err_find:
1439         kfree(names);
1440 err_names:
1441         kfree(callbacks);
1442 err_callback:
1443         kfree(vqs);
1444 err_vq:
1445         return ret;
1446 }
1447
1448 static int virtnet_alloc_queues(struct virtnet_info *vi)
1449 {
1450         int i;
1451
1452         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1453         if (!vi->sq)
1454                 goto err_sq;
1455         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1456         if (!vi->rq)
1457                 goto err_rq;
1458
1459         INIT_DELAYED_WORK(&vi->refill, refill_work);
1460         for (i = 0; i < vi->max_queue_pairs; i++) {
1461                 vi->rq[i].pages = NULL;
1462                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1463                                napi_weight);
1464
1465                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1466                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1467         }
1468
1469         return 0;
1470
1471 err_rq:
1472         kfree(vi->sq);
1473 err_sq:
1474         return -ENOMEM;
1475 }
1476
1477 static int init_vqs(struct virtnet_info *vi)
1478 {
1479         int ret;
1480
1481         /* Allocate send & receive queues */
1482         ret = virtnet_alloc_queues(vi);
1483         if (ret)
1484                 goto err;
1485
1486         ret = virtnet_find_vqs(vi);
1487         if (ret)
1488                 goto err_free;
1489
1490         get_online_cpus();
1491         virtnet_set_affinity(vi);
1492         put_online_cpus();
1493
1494         return 0;
1495
1496 err_free:
1497         virtnet_free_queues(vi);
1498 err:
1499         return ret;
1500 }
1501
1502 static int virtnet_probe(struct virtio_device *vdev)
1503 {
1504         int i, err;
1505         struct net_device *dev;
1506         struct virtnet_info *vi;
1507         u16 max_queue_pairs;
1508
1509         /* Find if host supports multiqueue virtio_net device */
1510         err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1511                                 offsetof(struct virtio_net_config,
1512                                 max_virtqueue_pairs), &max_queue_pairs);
1513
1514         /* We need at least 2 queue's */
1515         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1516             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1517             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1518                 max_queue_pairs = 1;
1519
1520         /* Allocate ourselves a network device with room for our info */
1521         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1522         if (!dev)
1523                 return -ENOMEM;
1524
1525         /* Set up network device as normal. */
1526         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1527         dev->netdev_ops = &virtnet_netdev;
1528         dev->features = NETIF_F_HIGHDMA;
1529
1530         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1531         SET_NETDEV_DEV(dev, &vdev->dev);
1532
1533         /* Do we support "hardware" checksums? */
1534         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1535                 /* This opens up the world of extra features. */
1536                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1537                 if (csum)
1538                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1539
1540                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1541                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1542                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1543                 }
1544                 /* Individual feature bits: what can host handle? */
1545                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1546                         dev->hw_features |= NETIF_F_TSO;
1547                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1548                         dev->hw_features |= NETIF_F_TSO6;
1549                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1550                         dev->hw_features |= NETIF_F_TSO_ECN;
1551                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1552                         dev->hw_features |= NETIF_F_UFO;
1553
1554                 if (gso)
1555                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1556                 /* (!csum && gso) case will be fixed by register_netdev() */
1557         }
1558         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1559                 dev->features |= NETIF_F_RXCSUM;
1560
1561         dev->vlan_features = dev->features;
1562
1563         /* Configuration may specify what MAC to use.  Otherwise random. */
1564         if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1565                                   offsetof(struct virtio_net_config, mac),
1566                                   dev->dev_addr, dev->addr_len) < 0)
1567                 eth_hw_addr_random(dev);
1568
1569         /* Set up our device-specific information */
1570         vi = netdev_priv(dev);
1571         vi->dev = dev;
1572         vi->vdev = vdev;
1573         vdev->priv = vi;
1574         vi->stats = alloc_percpu(struct virtnet_stats);
1575         err = -ENOMEM;
1576         if (vi->stats == NULL)
1577                 goto free;
1578
1579         mutex_init(&vi->config_lock);
1580         vi->config_enable = true;
1581         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1582
1583         /* If we can receive ANY GSO packets, we must allocate large ones. */
1584         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1585             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1586             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1587                 vi->big_packets = true;
1588
1589         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1590                 vi->mergeable_rx_bufs = true;
1591
1592         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1593                 vi->any_header_sg = true;
1594
1595         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1596                 vi->has_cvq = true;
1597
1598         /* Use single tx/rx queue pair as default */
1599         vi->curr_queue_pairs = 1;
1600         vi->max_queue_pairs = max_queue_pairs;
1601
1602         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1603         err = init_vqs(vi);
1604         if (err)
1605                 goto free_stats;
1606
1607         netif_set_real_num_tx_queues(dev, 1);
1608         netif_set_real_num_rx_queues(dev, 1);
1609
1610         err = register_netdev(dev);
1611         if (err) {
1612                 pr_debug("virtio_net: registering device failed\n");
1613                 goto free_vqs;
1614         }
1615
1616         /* Last of all, set up some receive buffers. */
1617         for (i = 0; i < vi->curr_queue_pairs; i++) {
1618                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1619
1620                 /* If we didn't even get one input buffer, we're useless. */
1621                 if (vi->rq[i].num == 0) {
1622                         free_unused_bufs(vi);
1623                         err = -ENOMEM;
1624                         goto free_recv_bufs;
1625                 }
1626         }
1627
1628         vi->nb.notifier_call = &virtnet_cpu_callback;
1629         err = register_hotcpu_notifier(&vi->nb);
1630         if (err) {
1631                 pr_debug("virtio_net: registering cpu notifier failed\n");
1632                 goto free_recv_bufs;
1633         }
1634
1635         /* Assume link up if device can't report link status,
1636            otherwise get link status from config. */
1637         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1638                 netif_carrier_off(dev);
1639                 schedule_work(&vi->config_work);
1640         } else {
1641                 vi->status = VIRTIO_NET_S_LINK_UP;
1642                 netif_carrier_on(dev);
1643         }
1644
1645         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1646                  dev->name, max_queue_pairs);
1647
1648         return 0;
1649
1650 free_recv_bufs:
1651         free_receive_bufs(vi);
1652         unregister_netdev(dev);
1653 free_vqs:
1654         cancel_delayed_work_sync(&vi->refill);
1655         virtnet_del_vqs(vi);
1656         if (vi->alloc_frag.page)
1657                 put_page(vi->alloc_frag.page);
1658 free_stats:
1659         free_percpu(vi->stats);
1660 free:
1661         free_netdev(dev);
1662         return err;
1663 }
1664
1665 static void remove_vq_common(struct virtnet_info *vi)
1666 {
1667         vi->vdev->config->reset(vi->vdev);
1668
1669         /* Free unused buffers in both send and recv, if any. */
1670         free_unused_bufs(vi);
1671
1672         free_receive_bufs(vi);
1673
1674         virtnet_del_vqs(vi);
1675 }
1676
1677 static void virtnet_remove(struct virtio_device *vdev)
1678 {
1679         struct virtnet_info *vi = vdev->priv;
1680
1681         unregister_hotcpu_notifier(&vi->nb);
1682
1683         /* Prevent config work handler from accessing the device. */
1684         mutex_lock(&vi->config_lock);
1685         vi->config_enable = false;
1686         mutex_unlock(&vi->config_lock);
1687
1688         unregister_netdev(vi->dev);
1689
1690         remove_vq_common(vi);
1691         if (vi->alloc_frag.page)
1692                 put_page(vi->alloc_frag.page);
1693
1694         flush_work(&vi->config_work);
1695
1696         free_percpu(vi->stats);
1697         free_netdev(vi->dev);
1698 }
1699
1700 #ifdef CONFIG_PM
1701 static int virtnet_freeze(struct virtio_device *vdev)
1702 {
1703         struct virtnet_info *vi = vdev->priv;
1704         int i;
1705
1706         unregister_hotcpu_notifier(&vi->nb);
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         err = register_hotcpu_notifier(&vi->nb);
1757         if (err)
1758                 return err;
1759
1760         return 0;
1761 }
1762 #endif
1763
1764 static struct virtio_device_id id_table[] = {
1765         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1766         { 0 },
1767 };
1768
1769 static unsigned int features[] = {
1770         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1771         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1772         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1773         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1774         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1775         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1776         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1777         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1778         VIRTIO_NET_F_CTRL_MAC_ADDR,
1779         VIRTIO_F_ANY_LAYOUT,
1780 };
1781
1782 static struct virtio_driver virtio_net_driver = {
1783         .feature_table = features,
1784         .feature_table_size = ARRAY_SIZE(features),
1785         .driver.name =  KBUILD_MODNAME,
1786         .driver.owner = THIS_MODULE,
1787         .id_table =     id_table,
1788         .probe =        virtnet_probe,
1789         .remove =       virtnet_remove,
1790         .config_changed = virtnet_config_changed,
1791 #ifdef CONFIG_PM
1792         .freeze =       virtnet_freeze,
1793         .restore =      virtnet_restore,
1794 #endif
1795 };
1796
1797 module_virtio_driver(virtio_net_driver);
1798
1799 MODULE_DEVICE_TABLE(virtio, id_table);
1800 MODULE_DESCRIPTION("Virtio network driver");
1801 MODULE_LICENSE("GPL");