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 GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  40 #define MERGE_BUFFER_LEN (ALIGN(GOOD_PACKET_LEN + \
  41                                 sizeof(struct virtio_net_hdr_mrg_rxbuf), \
  42                                 L1_CACHE_BYTES))
  43 #define GOOD_COPY_LEN   128
  44
  45 #define VIRTNET_DRIVER_VERSION "1.0.0"
  46
  47 struct virtnet_stats {
  48         struct u64_stats_sync tx_syncp;
  49         struct u64_stats_sync rx_syncp;
  50         u64 tx_bytes;
  51         u64 tx_packets;
  52
  53         u64 rx_bytes;
  54         u64 rx_packets;
  55 };
  56
  57 /* Internal representation of a send virtqueue */
  58 struct send_queue {
  59         /* Virtqueue associated with this send _queue */
  60         struct virtqueue *vq;
  61
  62         /* TX: fragments + linear part + virtio header */
  63         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  64
  65         /* Name of the send queue: output.$index */
  66         char name[40];
  67 };
  68
  69 /* Internal representation of a receive virtqueue */
  70 struct receive_queue {
  71         /* Virtqueue associated with this receive_queue */
  72         struct virtqueue *vq;
  73
  74         struct napi_struct napi;
  75
  76         /* Number of input buffers, and max we've ever had. */
  77         unsigned int num, max;
  78
  79         /* Chain pages by the private ptr. */
  80         struct page *pages;
  81
  82         /* RX: fragments + linear part + virtio header */
  83         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  84
  85         /* Name of this receive queue: input.$index */
  86         char name[40];
  87 };
  88
  89 struct virtnet_info {
  90         struct virtio_device *vdev;
  91         struct virtqueue *cvq;
  92         struct net_device *dev;
  93         struct send_queue *sq;
  94         struct receive_queue *rq;
  95         unsigned int status;
  96
  97         /* Max # of queue pairs supported by the device */
  98         u16 max_queue_pairs;
  99
 100         /* # of queue pairs currently used by the driver */
 101         u16 curr_queue_pairs;
 102
 103         /* I like... big packets and I cannot lie! */
 104         bool big_packets;
 105
 106         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 107         bool mergeable_rx_bufs;
 108
 109         /* Has control virtqueue */
 110         bool has_cvq;
 111
 112         /* Host can handle any s/g split between our header and packet data */
 113         bool any_header_sg;
 114
 115         /* enable config space updates */
 116         bool config_enable;
 117
 118         /* Active statistics */
 119         struct virtnet_stats __percpu *stats;
 120
 121         /* Work struct for refilling if we run low on memory. */
 122         struct delayed_work refill;
 123
 124         /* Work struct for config space updates */
 125         struct work_struct config_work;
 126
 127         /* Lock for config space updates */
 128         struct mutex config_lock;
 129
 130         /* Page_frag for GFP_KERNEL packet buffer allocation when we run
 131          * low on memory.
 132          */
 133         struct page_frag alloc_frag;
 134
 135         /* Does the affinity hint is set for virtqueues? */
 136         bool affinity_hint_set;
 137
 138         /* CPU hot plug notifier */
 139         struct notifier_block nb;
 140 };
 141
 142 struct skb_vnet_hdr {
 143         union {
 144                 struct virtio_net_hdr hdr;
 145                 struct virtio_net_hdr_mrg_rxbuf mhdr;
 146         };
 147 };
 148
 149 struct padded_vnet_hdr {
 150         struct virtio_net_hdr hdr;
 151         /*
 152          * virtio_net_hdr should be in a separated sg buffer because of a
 153          * QEMU bug, and data sg buffer shares same page with this header sg.
 154          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
 155          */
 156         char padding[6];
 157 };
 158
 159 /* Converting between virtqueue no. and kernel tx/rx queue no.
 160  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 161  */
 162 static int vq2txq(struct virtqueue *vq)
 163 {
 164         return (vq->index - 1) / 2;
 165 }
 166
 167 static int txq2vq(int txq)
 168 {
 169         return txq * 2 + 1;
 170 }
 171
 172 static int vq2rxq(struct virtqueue *vq)
 173 {
 174         return vq->index / 2;
 175 }
 176
 177 static int rxq2vq(int rxq)
 178 {
 179         return rxq * 2;
 180 }
 181
 182 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 183 {
 184         return (struct skb_vnet_hdr *)skb->cb;
 185 }
 186
 187 /*
 188  * private is used to chain pages for big packets, put the whole
 189  * most recent used list in the beginning for reuse
 190  */
 191 static void give_pages(struct receive_queue *rq, struct page *page)
 192 {
 193         struct page *end;
 194
 195         /* Find end of list, sew whole thing into vi->rq.pages. */
 196         for (end = page; end->private; end = (struct page *)end->private);
 197         end->private = (unsigned long)rq->pages;
 198         rq->pages = page;
 199 }
 200
 201 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 202 {
 203         struct page *p = rq->pages;
 204
 205         if (p) {
 206                 rq->pages = (struct page *)p->private;
 207                 /* clear private here, it is used to chain pages */
 208                 p->private = 0;
 209         } else
 210                 p = alloc_page(gfp_mask);
 211         return p;
 212 }
 213
 214 static void skb_xmit_done(struct virtqueue *vq)
 215 {
 216         struct virtnet_info *vi = vq->vdev->priv;
 217
 218         /* Suppress further interrupts. */
 219         virtqueue_disable_cb(vq);
 220
 221         /* We were probably waiting for more output buffers. */
 222         netif_wake_subqueue(vi->dev, vq2txq(vq));
 223 }
 224
 225 /* Called from bottom half context */
 226 static struct sk_buff *page_to_skb(struct receive_queue *rq,
 227                                    struct page *page, unsigned int offset,
 228                                    unsigned int len, unsigned int truesize)
 229 {
 230         struct virtnet_info *vi = rq->vq->vdev->priv;
 231         struct sk_buff *skb;
 232         struct skb_vnet_hdr *hdr;
 233         unsigned int copy, hdr_len, hdr_padded_len;
 234         char *p;
 235
 236         p = page_address(page) + offset;
 237
 238         /* copy small packet so we can reuse these pages for small data */
 239         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 240         if (unlikely(!skb))
 241                 return NULL;
 242
 243         hdr = skb_vnet_hdr(skb);
 244
 245         if (vi->mergeable_rx_bufs) {
 246                 hdr_len = sizeof hdr->mhdr;
 247                 hdr_padded_len = sizeof hdr->mhdr;
 248         } else {
 249                 hdr_len = sizeof hdr->hdr;
 250                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 251         }
 252
 253         memcpy(hdr, p, hdr_len);
 254
 255         len -= hdr_len;
 256         offset += hdr_padded_len;
 257         p += hdr_padded_len;
 258
 259         copy = len;
 260         if (copy > skb_tailroom(skb))
 261                 copy = skb_tailroom(skb);
 262         memcpy(skb_put(skb, copy), p, copy);
 263
 264         len -= copy;
 265         offset += copy;
 266
 267         if (vi->mergeable_rx_bufs) {
 268                 if (len)
 269                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 270                 else
 271                         put_page(page);
 272                 return skb;
 273         }
 274
 275         /*
 276          * Verify that we can indeed put this data into a skb.
 277          * This is here to handle cases when the device erroneously
 278          * tries to receive more than is possible. This is usually
 279          * the case of a broken device.
 280          */
 281         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 282                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 283                 dev_kfree_skb(skb);
 284                 return NULL;
 285         }
 286         BUG_ON(offset >= PAGE_SIZE);
 287         while (len) {
 288                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 289                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 290                                 frag_size, truesize);
 291                 len -= frag_size;
 292                 page = (struct page *)page->private;
 293                 offset = 0;
 294         }
 295
 296         if (page)
 297                 give_pages(rq, page);
 298
 299         return skb;
 300 }
 301
 302 static int receive_mergeable(struct receive_queue *rq, struct sk_buff *head_skb)
 303 {
 304         struct skb_vnet_hdr *hdr = skb_vnet_hdr(head_skb);
 305         struct sk_buff *curr_skb = head_skb;
 306         char *buf;
 307         struct page *page;
 308         int num_buf, len, offset;
 309
 310         num_buf = hdr->mhdr.num_buffers;
 311         while (--num_buf) {
 312                 int num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 313                 buf = virtqueue_get_buf(rq->vq, &len);
 314                 if (unlikely(!buf)) {
 315                         pr_debug("%s: rx error: %d buffers missing\n",
 316                                  head_skb->dev->name, hdr->mhdr.num_buffers);
 317                         head_skb->dev->stats.rx_length_errors++;
 318                         return -EINVAL;
 319                 }
 320                 if (unlikely(len > MERGE_BUFFER_LEN)) {
 321                         pr_debug("%s: rx error: merge buffer too long\n",
 322                                  head_skb->dev->name);
 323                         len = MERGE_BUFFER_LEN;
 324                 }
 325                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 326                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 327                         if (unlikely(!nskb)) {
 328                                 head_skb->dev->stats.rx_dropped++;
 329                                 return -ENOMEM;
 330                         }
 331                         if (curr_skb == head_skb)
 332                                 skb_shinfo(curr_skb)->frag_list = nskb;
 333                         else
 334                                 curr_skb->next = nskb;
 335                         curr_skb = nskb;
 336                         head_skb->truesize += nskb->truesize;
 337                         num_skb_frags = 0;
 338                 }
 339                 if (curr_skb != head_skb) {
 340                         head_skb->data_len += len;
 341                         head_skb->len += len;
 342                         head_skb->truesize += MERGE_BUFFER_LEN;
 343                 }
 344                 page = virt_to_head_page(buf);
 345                 offset = buf - (char *)page_address(page);
 346                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 347                         put_page(page);
 348                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 349                                              len, MERGE_BUFFER_LEN);
 350                 } else {
 351                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 352                                         offset, len, MERGE_BUFFER_LEN);
 353                 }
 354                 --rq->num;
 355         }
 356         return 0;
 357 }
 358
 359 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
 360 {
 361         struct virtnet_info *vi = rq->vq->vdev->priv;
 362         struct net_device *dev = vi->dev;
 363         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 364         struct sk_buff *skb;
 365         struct page *page;
 366         struct skb_vnet_hdr *hdr;
 367
 368         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 369                 pr_debug("%s: short packet %i\n", dev->name, len);
 370                 dev->stats.rx_length_errors++;
 371                 if (vi->big_packets)
 372                         give_pages(rq, buf);
 373                 else if (vi->mergeable_rx_bufs)
 374                         put_page(virt_to_head_page(buf));
 375                 else
 376                         dev_kfree_skb(buf);
 377                 return;
 378         }
 379
 380         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
 381                 skb = buf;
 382                 len -= sizeof(struct virtio_net_hdr);
 383                 skb_trim(skb, len);
 384         } else if (vi->mergeable_rx_bufs) {
 385                 struct page *page = virt_to_head_page(buf);
 386                 skb = page_to_skb(rq, page,
 387                                   (char *)buf - (char *)page_address(page),
 388                                   len, MERGE_BUFFER_LEN);
 389                 if (unlikely(!skb)) {
 390                         dev->stats.rx_dropped++;
 391                         put_page(page);
 392                         return;
 393                 }
 394                 if (receive_mergeable(rq, skb)) {
 395                         dev_kfree_skb(skb);
 396                         return;
 397                 }
 398         } else {
 399                 page = buf;
 400                 skb = page_to_skb(rq, page, 0, len, PAGE_SIZE);
 401                 if (unlikely(!skb)) {
 402                         dev->stats.rx_dropped++;
 403                         give_pages(rq, page);
 404                         return;
 405                 }
 406         }
 407
 408         hdr = skb_vnet_hdr(skb);
 409
 410         u64_stats_update_begin(&stats->rx_syncp);
 411         stats->rx_bytes += skb->len;
 412         stats->rx_packets++;
 413         u64_stats_update_end(&stats->rx_syncp);
 414
 415         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 416                 pr_debug("Needs csum!\n");
 417                 if (!skb_partial_csum_set(skb,
 418                                           hdr->hdr.csum_start,
 419                                           hdr->hdr.csum_offset))
 420                         goto frame_err;
 421         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 422                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 423         }
 424
 425         skb->protocol = eth_type_trans(skb, dev);
 426         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 427                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 428
 429         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 430                 pr_debug("GSO!\n");
 431                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 432                 case VIRTIO_NET_HDR_GSO_TCPV4:
 433                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 434                         break;
 435                 case VIRTIO_NET_HDR_GSO_UDP:
 436                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 437                         break;
 438                 case VIRTIO_NET_HDR_GSO_TCPV6:
 439                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 440                         break;
 441                 default:
 442                         net_warn_ratelimited("%s: bad gso type %u.\n",
 443                                              dev->name, hdr->hdr.gso_type);
 444                         goto frame_err;
 445                 }
 446
 447                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 448                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 449
 450                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 451                 if (skb_shinfo(skb)->gso_size == 0) {
 452                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 453                         goto frame_err;
 454                 }
 455
 456                 /* Header must be checked, and gso_segs computed. */
 457                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 458                 skb_shinfo(skb)->gso_segs = 0;
 459         }
 460
 461         netif_receive_skb(skb);
 462         return;
 463
 464 frame_err:
 465         dev->stats.rx_frame_errors++;
 466         dev_kfree_skb(skb);
 467 }
 468
 469 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
 470 {
 471         struct virtnet_info *vi = rq->vq->vdev->priv;
 472         struct sk_buff *skb;
 473         struct skb_vnet_hdr *hdr;
 474         int err;
 475
 476         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 477         if (unlikely(!skb))
 478                 return -ENOMEM;
 479
 480         skb_put(skb, GOOD_PACKET_LEN);
 481
 482         hdr = skb_vnet_hdr(skb);
 483         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
 484
 485         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 486
 487         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 488         if (err < 0)
 489                 dev_kfree_skb(skb);
 490
 491         return err;
 492 }
 493
 494 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
 495 {
 496         struct page *first, *list = NULL;
 497         char *p;
 498         int i, err, offset;
 499
 500         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 501         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 502                 first = get_a_page(rq, gfp);
 503                 if (!first) {
 504                         if (list)
 505                                 give_pages(rq, list);
 506                         return -ENOMEM;
 507                 }
 508                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 509
 510                 /* chain new page in list head to match sg */
 511                 first->private = (unsigned long)list;
 512                 list = first;
 513         }
 514
 515         first = get_a_page(rq, gfp);
 516         if (!first) {
 517                 give_pages(rq, list);
 518                 return -ENOMEM;
 519         }
 520         p = page_address(first);
 521
 522         /* rq->sg[0], rq->sg[1] share the same page */
 523         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
 524         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
 525
 526         /* rq->sg[1] for data packet, from offset */
 527         offset = sizeof(struct padded_vnet_hdr);
 528         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 529
 530         /* chain first in list head */
 531         first->private = (unsigned long)list;
 532         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 533                                   first, gfp);
 534         if (err < 0)
 535                 give_pages(rq, first);
 536
 537         return err;
 538 }
 539
 540 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 541 {
 542         struct virtnet_info *vi = rq->vq->vdev->priv;
 543         char *buf = NULL;
 544         int err;
 545
 546         if (gfp & __GFP_WAIT) {
 547                 if (skb_page_frag_refill(MERGE_BUFFER_LEN, &vi->alloc_frag,
 548                                          gfp)) {
 549                         buf = (char *)page_address(vi->alloc_frag.page) +
 550                               vi->alloc_frag.offset;
 551                         get_page(vi->alloc_frag.page);
 552                         vi->alloc_frag.offset += MERGE_BUFFER_LEN;
 553                 }
 554         } else {
 555                 buf = netdev_alloc_frag(MERGE_BUFFER_LEN);
 556         }
 557         if (!buf)
 558                 return -ENOMEM;
 559
 560         sg_init_one(rq->sg, buf, MERGE_BUFFER_LEN);
 561         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
 562         if (err < 0)
 563                 put_page(virt_to_head_page(buf));
 564
 565         return err;
 566 }
 567
 568 /*
 569  * Returns false if we couldn't fill entirely (OOM).
 570  *
 571  * Normally run in the receive path, but can also be run from ndo_open
 572  * before we're receiving packets, or from refill_work which is
 573  * careful to disable receiving (using napi_disable).
 574  */
 575 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
 576 {
 577         struct virtnet_info *vi = rq->vq->vdev->priv;
 578         int err;
 579         bool oom;
 580
 581         do {
 582                 if (vi->mergeable_rx_bufs)
 583                         err = add_recvbuf_mergeable(rq, gfp);
 584                 else if (vi->big_packets)
 585                         err = add_recvbuf_big(rq, gfp);
 586                 else
 587                         err = add_recvbuf_small(rq, gfp);
 588
 589                 oom = err == -ENOMEM;
 590                 if (err)
 591                         break;
 592                 ++rq->num;
 593         } while (rq->vq->num_free);
 594         if (unlikely(rq->num > rq->max))
 595                 rq->max = rq->num;
 596         if (unlikely(!virtqueue_kick(rq->vq)))
 597                 return false;
 598         return !oom;
 599 }
 600
 601 static void skb_recv_done(struct virtqueue *rvq)
 602 {
 603         struct virtnet_info *vi = rvq->vdev->priv;
 604         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 605
 606         /* Schedule NAPI, Suppress further interrupts if successful. */
 607         if (napi_schedule_prep(&rq->napi)) {
 608                 virtqueue_disable_cb(rvq);
 609                 __napi_schedule(&rq->napi);
 610         }
 611 }
 612
 613 static void virtnet_napi_enable(struct receive_queue *rq)
 614 {
 615         napi_enable(&rq->napi);
 616
 617         /* If all buffers were filled by other side before we napi_enabled, we
 618          * won't get another interrupt, so process any outstanding packets
 619          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 620          * We synchronize against interrupts via NAPI_STATE_SCHED */
 621         if (napi_schedule_prep(&rq->napi)) {
 622                 virtqueue_disable_cb(rq->vq);
 623                 local_bh_disable();
 624                 __napi_schedule(&rq->napi);
 625                 local_bh_enable();
 626         }
 627 }
 628
 629 static void refill_work(struct work_struct *work)
 630 {
 631         struct virtnet_info *vi =
 632                 container_of(work, struct virtnet_info, refill.work);
 633         bool still_empty;
 634         int i;
 635
 636         for (i = 0; i < vi->curr_queue_pairs; i++) {
 637                 struct receive_queue *rq = &vi->rq[i];
 638
 639                 napi_disable(&rq->napi);
 640                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 641                 virtnet_napi_enable(rq);
 642
 643                 /* In theory, this can happen: if we don't get any buffers in
 644                  * we will *never* try to fill again.
 645                  */
 646                 if (still_empty)
 647                         schedule_delayed_work(&vi->refill, HZ/2);
 648         }
 649 }
 650
 651 static int virtnet_poll(struct napi_struct *napi, int budget)
 652 {
 653         struct receive_queue *rq =
 654                 container_of(napi, struct receive_queue, napi);
 655         struct virtnet_info *vi = rq->vq->vdev->priv;
 656         void *buf;
 657         unsigned int r, len, received = 0;
 658
 659 again:
 660         while (received < budget &&
 661                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 662                 receive_buf(rq, buf, len);
 663                 --rq->num;
 664                 received++;
 665         }
 666
 667         if (rq->num < rq->max / 2) {
 668                 if (!try_fill_recv(rq, GFP_ATOMIC))
 669                         schedule_delayed_work(&vi->refill, 0);
 670         }
 671
 672         /* Out of packets? */
 673         if (received < budget) {
 674                 r = virtqueue_enable_cb_prepare(rq->vq);
 675                 napi_complete(napi);
 676                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 677                     napi_schedule_prep(napi)) {
 678                         virtqueue_disable_cb(rq->vq);
 679                         __napi_schedule(napi);
 680                         goto again;
 681                 }
 682         }
 683
 684         return received;
 685 }
 686
 687 static int virtnet_open(struct net_device *dev)
 688 {
 689         struct virtnet_info *vi = netdev_priv(dev);
 690         int i;
 691
 692         for (i = 0; i < vi->max_queue_pairs; i++) {
 693                 if (i < vi->curr_queue_pairs)
 694                         /* Make sure we have some buffers: if oom use wq. */
 695                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 696                                 schedule_delayed_work(&vi->refill, 0);
 697                 virtnet_napi_enable(&vi->rq[i]);
 698         }
 699
 700         return 0;
 701 }
 702
 703 static void free_old_xmit_skbs(struct send_queue *sq)
 704 {
 705         struct sk_buff *skb;
 706         unsigned int len;
 707         struct virtnet_info *vi = sq->vq->vdev->priv;
 708         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 709
 710         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 711                 pr_debug("Sent skb %p\n", skb);
 712
 713                 u64_stats_update_begin(&stats->tx_syncp);
 714                 stats->tx_bytes += skb->len;
 715                 stats->tx_packets++;
 716                 u64_stats_update_end(&stats->tx_syncp);
 717
 718                 dev_kfree_skb_any(skb);
 719         }
 720 }
 721
 722 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 723 {
 724         struct skb_vnet_hdr *hdr;
 725         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 726         struct virtnet_info *vi = sq->vq->vdev->priv;
 727         unsigned num_sg;
 728         unsigned hdr_len;
 729         bool can_push;
 730
 731         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 732         if (vi->mergeable_rx_bufs)
 733                 hdr_len = sizeof hdr->mhdr;
 734         else
 735                 hdr_len = sizeof hdr->hdr;
 736
 737         can_push = vi->any_header_sg &&
 738                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 739                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 740         /* Even if we can, don't push here yet as this would skew
 741          * csum_start offset below. */
 742         if (can_push)
 743                 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
 744         else
 745                 hdr = skb_vnet_hdr(skb);
 746
 747         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 748                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 749                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 750                 hdr->hdr.csum_offset = skb->csum_offset;
 751         } else {
 752                 hdr->hdr.flags = 0;
 753                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 754         }
 755
 756         if (skb_is_gso(skb)) {
 757                 hdr->hdr.hdr_len = skb_headlen(skb);
 758                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 759                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 760                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 761                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 762                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 763                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 764                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 765                 else
 766                         BUG();
 767                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 768                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 769         } else {
 770                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 771                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 772         }
 773
 774         if (vi->mergeable_rx_bufs)
 775                 hdr->mhdr.num_buffers = 0;
 776
 777         if (can_push) {
 778                 __skb_push(skb, hdr_len);
 779                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 780                 /* Pull header back to avoid skew in tx bytes calculations. */
 781                 __skb_pull(skb, hdr_len);
 782         } else {
 783                 sg_set_buf(sq->sg, hdr, hdr_len);
 784                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 785         }
 786         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 787 }
 788
 789 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 790 {
 791         struct virtnet_info *vi = netdev_priv(dev);
 792         int qnum = skb_get_queue_mapping(skb);
 793         struct send_queue *sq = &vi->sq[qnum];
 794         int err;
 795
 796         /* Free up any pending old buffers before queueing new ones. */
 797         free_old_xmit_skbs(sq);
 798
 799         /* Try to transmit */
 800         err = xmit_skb(sq, skb);
 801
 802         /* This should not happen! */
 803         if (unlikely(err) || unlikely(!virtqueue_kick(sq->vq))) {
 804                 dev->stats.tx_fifo_errors++;
 805                 if (net_ratelimit())
 806                         dev_warn(&dev->dev,
 807                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 808                 dev->stats.tx_dropped++;
 809                 kfree_skb(skb);
 810                 return NETDEV_TX_OK;
 811         }
 812
 813         /* Don't wait up for transmitted skbs to be freed. */
 814         skb_orphan(skb);
 815         nf_reset(skb);
 816
 817         /* Apparently nice girls don't return TX_BUSY; stop the queue
 818          * before it gets out of hand.  Naturally, this wastes entries. */
 819         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 820                 netif_stop_subqueue(dev, qnum);
 821                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 822                         /* More just got used, free them then recheck. */
 823                         free_old_xmit_skbs(sq);
 824                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 825                                 netif_start_subqueue(dev, qnum);
 826                                 virtqueue_disable_cb(sq->vq);
 827                         }
 828                 }
 829         }
 830
 831         return NETDEV_TX_OK;
 832 }
 833
 834 /*
 835  * Send command via the control virtqueue and check status.  Commands
 836  * supported by the hypervisor, as indicated by feature bits, should
 837  * never fail unless improperly formated.
 838  */
 839 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 840                                  struct scatterlist *out,
 841                                  struct scatterlist *in)
 842 {
 843         struct scatterlist *sgs[4], hdr, stat;
 844         struct virtio_net_ctrl_hdr ctrl;
 845         virtio_net_ctrl_ack status = ~0;
 846         unsigned out_num = 0, in_num = 0, tmp;
 847
 848         /* Caller should know better */
 849         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 850
 851         ctrl.class = class;
 852         ctrl.cmd = cmd;
 853         /* Add header */
 854         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 855         sgs[out_num++] = &hdr;
 856
 857         if (out)
 858                 sgs[out_num++] = out;
 859         if (in)
 860                 sgs[out_num + in_num++] = in;
 861
 862         /* Add return status. */
 863         sg_init_one(&stat, &status, sizeof(status));
 864         sgs[out_num + in_num++] = &stat;
 865
 866         BUG_ON(out_num + in_num > ARRAY_SIZE(sgs));
 867         BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, in_num, vi, GFP_ATOMIC)
 868                < 0);
 869
 870         if (unlikely(!virtqueue_kick(vi->cvq)))
 871                 return status == VIRTIO_NET_OK;
 872
 873         /* Spin for a response, the kick causes an ioport write, trapping
 874          * into the hypervisor, so the request should be handled immediately.
 875          */
 876         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
 877                !virtqueue_is_broken(vi->cvq))
 878                 cpu_relax();
 879
 880         return status == VIRTIO_NET_OK;
 881 }
 882
 883 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 884 {
 885         struct virtnet_info *vi = netdev_priv(dev);
 886         struct virtio_device *vdev = vi->vdev;
 887         int ret;
 888         struct sockaddr *addr = p;
 889         struct scatterlist sg;
 890
 891         ret = eth_prepare_mac_addr_change(dev, p);
 892         if (ret)
 893                 return ret;
 894
 895         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 896                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 897                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 898                                           VIRTIO_NET_CTRL_MAC_ADDR_SET,
 899                                           &sg, NULL)) {
 900                         dev_warn(&vdev->dev,
 901                                  "Failed to set mac address by vq command.\n");
 902                         return -EINVAL;
 903                 }
 904         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 905                 unsigned int i;
 906
 907                 /* Naturally, this has an atomicity problem. */
 908                 for (i = 0; i < dev->addr_len; i++)
 909                         virtio_cwrite8(vdev,
 910                                        offsetof(struct virtio_net_config, mac) +
 911                                        i, addr->sa_data[i]);
 912         }
 913
 914         eth_commit_mac_addr_change(dev, p);
 915
 916         return 0;
 917 }
 918
 919 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 920                                                struct rtnl_link_stats64 *tot)
 921 {
 922         struct virtnet_info *vi = netdev_priv(dev);
 923         int cpu;
 924         unsigned int start;
 925
 926         for_each_possible_cpu(cpu) {
 927                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 928                 u64 tpackets, tbytes, rpackets, rbytes;
 929
 930                 do {
 931                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 932                         tpackets = stats->tx_packets;
 933                         tbytes   = stats->tx_bytes;
 934                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 935
 936                 do {
 937                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 938                         rpackets = stats->rx_packets;
 939                         rbytes   = stats->rx_bytes;
 940                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 941
 942                 tot->rx_packets += rpackets;
 943                 tot->tx_packets += tpackets;
 944                 tot->rx_bytes   += rbytes;
 945                 tot->tx_bytes   += tbytes;
 946         }
 947
 948         tot->tx_dropped = dev->stats.tx_dropped;
 949         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 950         tot->rx_dropped = dev->stats.rx_dropped;
 951         tot->rx_length_errors = dev->stats.rx_length_errors;
 952         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 953
 954         return tot;
 955 }
 956
 957 #ifdef CONFIG_NET_POLL_CONTROLLER
 958 static void virtnet_netpoll(struct net_device *dev)
 959 {
 960         struct virtnet_info *vi = netdev_priv(dev);
 961         int i;
 962
 963         for (i = 0; i < vi->curr_queue_pairs; i++)
 964                 napi_schedule(&vi->rq[i].napi);
 965 }
 966 #endif
 967
 968 static void virtnet_ack_link_announce(struct virtnet_info *vi)
 969 {
 970         rtnl_lock();
 971         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
 972                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
 973                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
 974         rtnl_unlock();
 975 }
 976
 977 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
 978 {
 979         struct scatterlist sg;
 980         struct virtio_net_ctrl_mq s;
 981         struct net_device *dev = vi->dev;
 982
 983         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
 984                 return 0;
 985
 986         s.virtqueue_pairs = queue_pairs;
 987         sg_init_one(&sg, &s, sizeof(s));
 988
 989         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
 990                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
 991                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
 992                          queue_pairs);
 993                 return -EINVAL;
 994         } else {
 995                 vi->curr_queue_pairs = queue_pairs;
 996                 /* virtnet_open() will refill when device is going to up. */
 997                 if (dev->flags & IFF_UP)
 998                         schedule_delayed_work(&vi->refill, 0);
 999         }
1000
1001         return 0;
1002 }
1003
1004 static int virtnet_close(struct net_device *dev)
1005 {
1006         struct virtnet_info *vi = netdev_priv(dev);
1007         int i;
1008
1009         /* Make sure refill_work doesn't re-enable napi! */
1010         cancel_delayed_work_sync(&vi->refill);
1011
1012         for (i = 0; i < vi->max_queue_pairs; i++)
1013                 napi_disable(&vi->rq[i].napi);
1014
1015         return 0;
1016 }
1017
1018 static void virtnet_set_rx_mode(struct net_device *dev)
1019 {
1020         struct virtnet_info *vi = netdev_priv(dev);
1021         struct scatterlist sg[2];
1022         u8 promisc, allmulti;
1023         struct virtio_net_ctrl_mac *mac_data;
1024         struct netdev_hw_addr *ha;
1025         int uc_count;
1026         int mc_count;
1027         void *buf;
1028         int i;
1029
1030         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
1031         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1032                 return;
1033
1034         promisc = ((dev->flags & IFF_PROMISC) != 0);
1035         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1036
1037         sg_init_one(sg, &promisc, sizeof(promisc));
1038
1039         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1040                                   VIRTIO_NET_CTRL_RX_PROMISC,
1041                                   sg, NULL))
1042                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1043                          promisc ? "en" : "dis");
1044
1045         sg_init_one(sg, &allmulti, sizeof(allmulti));
1046
1047         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1048                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
1049                                   sg, NULL))
1050                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1051                          allmulti ? "en" : "dis");
1052
1053         uc_count = netdev_uc_count(dev);
1054         mc_count = netdev_mc_count(dev);
1055         /* MAC filter - use one buffer for both lists */
1056         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1057                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1058         mac_data = buf;
1059         if (!buf)
1060                 return;
1061
1062         sg_init_table(sg, 2);
1063
1064         /* Store the unicast list and count in the front of the buffer */
1065         mac_data->entries = uc_count;
1066         i = 0;
1067         netdev_for_each_uc_addr(ha, dev)
1068                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1069
1070         sg_set_buf(&sg[0], mac_data,
1071                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1072
1073         /* multicast list and count fill the end */
1074         mac_data = (void *)&mac_data->macs[uc_count][0];
1075
1076         mac_data->entries = mc_count;
1077         i = 0;
1078         netdev_for_each_mc_addr(ha, dev)
1079                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1080
1081         sg_set_buf(&sg[1], mac_data,
1082                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1083
1084         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1085                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
1086                                   sg, NULL))
1087                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1088
1089         kfree(buf);
1090 }
1091
1092 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1093                                    __be16 proto, u16 vid)
1094 {
1095         struct virtnet_info *vi = netdev_priv(dev);
1096         struct scatterlist sg;
1097
1098         sg_init_one(&sg, &vid, sizeof(vid));
1099
1100         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1101                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1102                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1103         return 0;
1104 }
1105
1106 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1107                                     __be16 proto, u16 vid)
1108 {
1109         struct virtnet_info *vi = netdev_priv(dev);
1110         struct scatterlist sg;
1111
1112         sg_init_one(&sg, &vid, sizeof(vid));
1113
1114         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1115                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1116                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1117         return 0;
1118 }
1119
1120 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1121 {
1122         int i;
1123
1124         if (vi->affinity_hint_set) {
1125                 for (i = 0; i < vi->max_queue_pairs; i++) {
1126                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1127                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1128                 }
1129
1130                 vi->affinity_hint_set = false;
1131         }
1132 }
1133
1134 static void virtnet_set_affinity(struct virtnet_info *vi)
1135 {
1136         int i;
1137         int cpu;
1138
1139         /* In multiqueue mode, when the number of cpu is equal to the number of
1140          * queue pairs, we let the queue pairs to be private to one cpu by
1141          * setting the affinity hint to eliminate the contention.
1142          */
1143         if (vi->curr_queue_pairs == 1 ||
1144             vi->max_queue_pairs != num_online_cpus()) {
1145                 virtnet_clean_affinity(vi, -1);
1146                 return;
1147         }
1148
1149         i = 0;
1150         for_each_online_cpu(cpu) {
1151                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1152                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1153                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1154                 i++;
1155         }
1156
1157         vi->affinity_hint_set = true;
1158 }
1159
1160 static int virtnet_cpu_callback(struct notifier_block *nfb,
1161                                 unsigned long action, void *hcpu)
1162 {
1163         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1164
1165         switch(action & ~CPU_TASKS_FROZEN) {
1166         case CPU_ONLINE:
1167         case CPU_DOWN_FAILED:
1168         case CPU_DEAD:
1169                 virtnet_set_affinity(vi);
1170                 break;
1171         case CPU_DOWN_PREPARE:
1172                 virtnet_clean_affinity(vi, (long)hcpu);
1173                 break;
1174         default:
1175                 break;
1176         }
1177
1178         return NOTIFY_OK;
1179 }
1180
1181 static void virtnet_get_ringparam(struct net_device *dev,
1182                                 struct ethtool_ringparam *ring)
1183 {
1184         struct virtnet_info *vi = netdev_priv(dev);
1185
1186         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1187         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1188         ring->rx_pending = ring->rx_max_pending;
1189         ring->tx_pending = ring->tx_max_pending;
1190 }
1191
1192
1193 static void virtnet_get_drvinfo(struct net_device *dev,
1194                                 struct ethtool_drvinfo *info)
1195 {
1196         struct virtnet_info *vi = netdev_priv(dev);
1197         struct virtio_device *vdev = vi->vdev;
1198
1199         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1200         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1201         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1202
1203 }
1204
1205 /* TODO: Eliminate OOO packets during switching */
1206 static int virtnet_set_channels(struct net_device *dev,
1207                                 struct ethtool_channels *channels)
1208 {
1209         struct virtnet_info *vi = netdev_priv(dev);
1210         u16 queue_pairs = channels->combined_count;
1211         int err;
1212
1213         /* We don't support separate rx/tx channels.
1214          * We don't allow setting 'other' channels.
1215          */
1216         if (channels->rx_count || channels->tx_count || channels->other_count)
1217                 return -EINVAL;
1218
1219         if (queue_pairs > vi->max_queue_pairs)
1220                 return -EINVAL;
1221
1222         get_online_cpus();
1223         err = virtnet_set_queues(vi, queue_pairs);
1224         if (!err) {
1225                 netif_set_real_num_tx_queues(dev, queue_pairs);
1226                 netif_set_real_num_rx_queues(dev, queue_pairs);
1227
1228                 virtnet_set_affinity(vi);
1229         }
1230         put_online_cpus();
1231
1232         return err;
1233 }
1234
1235 static void virtnet_get_channels(struct net_device *dev,
1236                                  struct ethtool_channels *channels)
1237 {
1238         struct virtnet_info *vi = netdev_priv(dev);
1239
1240         channels->combined_count = vi->curr_queue_pairs;
1241         channels->max_combined = vi->max_queue_pairs;
1242         channels->max_other = 0;
1243         channels->rx_count = 0;
1244         channels->tx_count = 0;
1245         channels->other_count = 0;
1246 }
1247
1248 static const struct ethtool_ops virtnet_ethtool_ops = {
1249         .get_drvinfo = virtnet_get_drvinfo,
1250         .get_link = ethtool_op_get_link,
1251         .get_ringparam = virtnet_get_ringparam,
1252         .set_channels = virtnet_set_channels,
1253         .get_channels = virtnet_get_channels,
1254 };
1255
1256 #define MIN_MTU 68
1257 #define MAX_MTU 65535
1258
1259 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1260 {
1261         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1262                 return -EINVAL;
1263         dev->mtu = new_mtu;
1264         return 0;
1265 }
1266
1267 static const struct net_device_ops virtnet_netdev = {
1268         .ndo_open            = virtnet_open,
1269         .ndo_stop            = virtnet_close,
1270         .ndo_start_xmit      = start_xmit,
1271         .ndo_validate_addr   = eth_validate_addr,
1272         .ndo_set_mac_address = virtnet_set_mac_address,
1273         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1274         .ndo_change_mtu      = virtnet_change_mtu,
1275         .ndo_get_stats64     = virtnet_stats,
1276         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1277         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1278 #ifdef CONFIG_NET_POLL_CONTROLLER
1279         .ndo_poll_controller = virtnet_netpoll,
1280 #endif
1281 };
1282
1283 static void virtnet_config_changed_work(struct work_struct *work)
1284 {
1285         struct virtnet_info *vi =
1286                 container_of(work, struct virtnet_info, config_work);
1287         u16 v;
1288
1289         mutex_lock(&vi->config_lock);
1290         if (!vi->config_enable)
1291                 goto done;
1292
1293         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1294                                  struct virtio_net_config, status, &v) < 0)
1295                 goto done;
1296
1297         if (v & VIRTIO_NET_S_ANNOUNCE) {
1298                 netdev_notify_peers(vi->dev);
1299                 virtnet_ack_link_announce(vi);
1300         }
1301
1302         /* Ignore unknown (future) status bits */
1303         v &= VIRTIO_NET_S_LINK_UP;
1304
1305         if (vi->status == v)
1306                 goto done;
1307
1308         vi->status = v;
1309
1310         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1311                 netif_carrier_on(vi->dev);
1312                 netif_tx_wake_all_queues(vi->dev);
1313         } else {
1314                 netif_carrier_off(vi->dev);
1315                 netif_tx_stop_all_queues(vi->dev);
1316         }
1317 done:
1318         mutex_unlock(&vi->config_lock);
1319 }
1320
1321 static void virtnet_config_changed(struct virtio_device *vdev)
1322 {
1323         struct virtnet_info *vi = vdev->priv;
1324
1325         schedule_work(&vi->config_work);
1326 }
1327
1328 static void virtnet_free_queues(struct virtnet_info *vi)
1329 {
1330         kfree(vi->rq);
1331         kfree(vi->sq);
1332 }
1333
1334 static void free_receive_bufs(struct virtnet_info *vi)
1335 {
1336         int i;
1337
1338         for (i = 0; i < vi->max_queue_pairs; i++) {
1339                 while (vi->rq[i].pages)
1340                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1341         }
1342 }
1343
1344 static void free_unused_bufs(struct virtnet_info *vi)
1345 {
1346         void *buf;
1347         int i;
1348
1349         for (i = 0; i < vi->max_queue_pairs; i++) {
1350                 struct virtqueue *vq = vi->sq[i].vq;
1351                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1352                         dev_kfree_skb(buf);
1353         }
1354
1355         for (i = 0; i < vi->max_queue_pairs; i++) {
1356                 struct virtqueue *vq = vi->rq[i].vq;
1357
1358                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1359                         if (vi->big_packets)
1360                                 give_pages(&vi->rq[i], buf);
1361                         else if (vi->mergeable_rx_bufs)
1362                                 put_page(virt_to_head_page(buf));
1363                         else
1364                                 dev_kfree_skb(buf);
1365                         --vi->rq[i].num;
1366                 }
1367                 BUG_ON(vi->rq[i].num != 0);
1368         }
1369 }
1370
1371 static void virtnet_del_vqs(struct virtnet_info *vi)
1372 {
1373         struct virtio_device *vdev = vi->vdev;
1374
1375         virtnet_clean_affinity(vi, -1);
1376
1377         vdev->config->del_vqs(vdev);
1378
1379         virtnet_free_queues(vi);
1380 }
1381
1382 static int virtnet_find_vqs(struct virtnet_info *vi)
1383 {
1384         vq_callback_t **callbacks;
1385         struct virtqueue **vqs;
1386         int ret = -ENOMEM;
1387         int i, total_vqs;
1388         const char **names;
1389
1390         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1391          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1392          * possible control vq.
1393          */
1394         total_vqs = vi->max_queue_pairs * 2 +
1395                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1396
1397         /* Allocate space for find_vqs parameters */
1398         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1399         if (!vqs)
1400                 goto err_vq;
1401         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1402         if (!callbacks)
1403                 goto err_callback;
1404         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1405         if (!names)
1406                 goto err_names;
1407
1408         /* Parameters for control virtqueue, if any */
1409         if (vi->has_cvq) {
1410                 callbacks[total_vqs - 1] = NULL;
1411                 names[total_vqs - 1] = "control";
1412         }
1413
1414         /* Allocate/initialize parameters for send/receive virtqueues */
1415         for (i = 0; i < vi->max_queue_pairs; i++) {
1416                 callbacks[rxq2vq(i)] = skb_recv_done;
1417                 callbacks[txq2vq(i)] = skb_xmit_done;
1418                 sprintf(vi->rq[i].name, "input.%d", i);
1419                 sprintf(vi->sq[i].name, "output.%d", i);
1420                 names[rxq2vq(i)] = vi->rq[i].name;
1421                 names[txq2vq(i)] = vi->sq[i].name;
1422         }
1423
1424         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1425                                          names);
1426         if (ret)
1427                 goto err_find;
1428
1429         if (vi->has_cvq) {
1430                 vi->cvq = vqs[total_vqs - 1];
1431                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1432                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1433         }
1434
1435         for (i = 0; i < vi->max_queue_pairs; i++) {
1436                 vi->rq[i].vq = vqs[rxq2vq(i)];
1437                 vi->sq[i].vq = vqs[txq2vq(i)];
1438         }
1439
1440         kfree(names);
1441         kfree(callbacks);
1442         kfree(vqs);
1443
1444         return 0;
1445
1446 err_find:
1447         kfree(names);
1448 err_names:
1449         kfree(callbacks);
1450 err_callback:
1451         kfree(vqs);
1452 err_vq:
1453         return ret;
1454 }
1455
1456 static int virtnet_alloc_queues(struct virtnet_info *vi)
1457 {
1458         int i;
1459
1460         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1461         if (!vi->sq)
1462                 goto err_sq;
1463         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1464         if (!vi->rq)
1465                 goto err_rq;
1466
1467         INIT_DELAYED_WORK(&vi->refill, refill_work);
1468         for (i = 0; i < vi->max_queue_pairs; i++) {
1469                 vi->rq[i].pages = NULL;
1470                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1471                                napi_weight);
1472
1473                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1474                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1475         }
1476
1477         return 0;
1478
1479 err_rq:
1480         kfree(vi->sq);
1481 err_sq:
1482         return -ENOMEM;
1483 }
1484
1485 static int init_vqs(struct virtnet_info *vi)
1486 {
1487         int ret;
1488
1489         /* Allocate send & receive queues */
1490         ret = virtnet_alloc_queues(vi);
1491         if (ret)
1492                 goto err;
1493
1494         ret = virtnet_find_vqs(vi);
1495         if (ret)
1496                 goto err_free;
1497
1498         get_online_cpus();
1499         virtnet_set_affinity(vi);
1500         put_online_cpus();
1501
1502         return 0;
1503
1504 err_free:
1505         virtnet_free_queues(vi);
1506 err:
1507         return ret;
1508 }
1509
1510 static int virtnet_probe(struct virtio_device *vdev)
1511 {
1512         int i, err;
1513         struct net_device *dev;
1514         struct virtnet_info *vi;
1515         u16 max_queue_pairs;
1516
1517         /* Find if host supports multiqueue virtio_net device */
1518         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1519                                    struct virtio_net_config,
1520                                    max_virtqueue_pairs, &max_queue_pairs);
1521
1522         /* We need at least 2 queue's */
1523         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1524             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1525             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1526                 max_queue_pairs = 1;
1527
1528         /* Allocate ourselves a network device with room for our info */
1529         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1530         if (!dev)
1531                 return -ENOMEM;
1532
1533         /* Set up network device as normal. */
1534         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1535         dev->netdev_ops = &virtnet_netdev;
1536         dev->features = NETIF_F_HIGHDMA;
1537
1538         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1539         SET_NETDEV_DEV(dev, &vdev->dev);
1540
1541         /* Do we support "hardware" checksums? */
1542         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1543                 /* This opens up the world of extra features. */
1544                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1545                 if (csum)
1546                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1547
1548                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1549                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1550                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1551                 }
1552                 /* Individual feature bits: what can host handle? */
1553                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1554                         dev->hw_features |= NETIF_F_TSO;
1555                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1556                         dev->hw_features |= NETIF_F_TSO6;
1557                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1558                         dev->hw_features |= NETIF_F_TSO_ECN;
1559                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1560                         dev->hw_features |= NETIF_F_UFO;
1561
1562                 if (gso)
1563                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1564                 /* (!csum && gso) case will be fixed by register_netdev() */
1565         }
1566         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1567                 dev->features |= NETIF_F_RXCSUM;
1568
1569         dev->vlan_features = dev->features;
1570
1571         /* Configuration may specify what MAC to use.  Otherwise random. */
1572         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1573                 virtio_cread_bytes(vdev,
1574                                    offsetof(struct virtio_net_config, mac),
1575                                    dev->dev_addr, dev->addr_len);
1576         else
1577                 eth_hw_addr_random(dev);
1578
1579         /* Set up our device-specific information */
1580         vi = netdev_priv(dev);
1581         vi->dev = dev;
1582         vi->vdev = vdev;
1583         vdev->priv = vi;
1584         vi->stats = alloc_percpu(struct virtnet_stats);
1585         err = -ENOMEM;
1586         if (vi->stats == NULL)
1587                 goto free;
1588
1589         for_each_possible_cpu(i) {
1590                 struct virtnet_stats *virtnet_stats;
1591                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1592                 u64_stats_init(&virtnet_stats->tx_syncp);
1593                 u64_stats_init(&virtnet_stats->rx_syncp);
1594         }
1595
1596         mutex_init(&vi->config_lock);
1597         vi->config_enable = true;
1598         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1599
1600         /* If we can receive ANY GSO packets, we must allocate large ones. */
1601         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1602             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1603             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1604                 vi->big_packets = true;
1605
1606         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1607                 vi->mergeable_rx_bufs = true;
1608
1609         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1610                 vi->any_header_sg = true;
1611
1612         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1613                 vi->has_cvq = true;
1614
1615         /* Use single tx/rx queue pair as default */
1616         vi->curr_queue_pairs = 1;
1617         vi->max_queue_pairs = max_queue_pairs;
1618
1619         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1620         err = init_vqs(vi);
1621         if (err)
1622                 goto free_stats;
1623
1624         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1625         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1626
1627         err = register_netdev(dev);
1628         if (err) {
1629                 pr_debug("virtio_net: registering device failed\n");
1630                 goto free_vqs;
1631         }
1632
1633         /* Last of all, set up some receive buffers. */
1634         for (i = 0; i < vi->curr_queue_pairs; i++) {
1635                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1636
1637                 /* If we didn't even get one input buffer, we're useless. */
1638                 if (vi->rq[i].num == 0) {
1639                         free_unused_bufs(vi);
1640                         err = -ENOMEM;
1641                         goto free_recv_bufs;
1642                 }
1643         }
1644
1645         vi->nb.notifier_call = &virtnet_cpu_callback;
1646         err = register_hotcpu_notifier(&vi->nb);
1647         if (err) {
1648                 pr_debug("virtio_net: registering cpu notifier failed\n");
1649                 goto free_recv_bufs;
1650         }
1651
1652         /* Assume link up if device can't report link status,
1653            otherwise get link status from config. */
1654         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1655                 netif_carrier_off(dev);
1656                 schedule_work(&vi->config_work);
1657         } else {
1658                 vi->status = VIRTIO_NET_S_LINK_UP;
1659                 netif_carrier_on(dev);
1660         }
1661
1662         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1663                  dev->name, max_queue_pairs);
1664
1665         return 0;
1666
1667 free_recv_bufs:
1668         free_receive_bufs(vi);
1669         unregister_netdev(dev);
1670 free_vqs:
1671         cancel_delayed_work_sync(&vi->refill);
1672         virtnet_del_vqs(vi);
1673         if (vi->alloc_frag.page)
1674                 put_page(vi->alloc_frag.page);
1675 free_stats:
1676         free_percpu(vi->stats);
1677 free:
1678         free_netdev(dev);
1679         return err;
1680 }
1681
1682 static void remove_vq_common(struct virtnet_info *vi)
1683 {
1684         vi->vdev->config->reset(vi->vdev);
1685
1686         /* Free unused buffers in both send and recv, if any. */
1687         free_unused_bufs(vi);
1688
1689         free_receive_bufs(vi);
1690
1691         virtnet_del_vqs(vi);
1692 }
1693
1694 static void virtnet_remove(struct virtio_device *vdev)
1695 {
1696         struct virtnet_info *vi = vdev->priv;
1697
1698         unregister_hotcpu_notifier(&vi->nb);
1699
1700         /* Prevent config work handler from accessing the device. */
1701         mutex_lock(&vi->config_lock);
1702         vi->config_enable = false;
1703         mutex_unlock(&vi->config_lock);
1704
1705         unregister_netdev(vi->dev);
1706
1707         remove_vq_common(vi);
1708         if (vi->alloc_frag.page)
1709                 put_page(vi->alloc_frag.page);
1710
1711         flush_work(&vi->config_work);
1712
1713         free_percpu(vi->stats);
1714         free_netdev(vi->dev);
1715 }
1716
1717 #ifdef CONFIG_PM_SLEEP
1718 static int virtnet_freeze(struct virtio_device *vdev)
1719 {
1720         struct virtnet_info *vi = vdev->priv;
1721         int i;
1722
1723         unregister_hotcpu_notifier(&vi->nb);
1724
1725         /* Prevent config work handler from accessing the device */
1726         mutex_lock(&vi->config_lock);
1727         vi->config_enable = false;
1728         mutex_unlock(&vi->config_lock);
1729
1730         netif_device_detach(vi->dev);
1731         cancel_delayed_work_sync(&vi->refill);
1732
1733         if (netif_running(vi->dev))
1734                 for (i = 0; i < vi->max_queue_pairs; i++) {
1735                         napi_disable(&vi->rq[i].napi);
1736                         netif_napi_del(&vi->rq[i].napi);
1737                 }
1738
1739         remove_vq_common(vi);
1740
1741         flush_work(&vi->config_work);
1742
1743         return 0;
1744 }
1745
1746 static int virtnet_restore(struct virtio_device *vdev)
1747 {
1748         struct virtnet_info *vi = vdev->priv;
1749         int err, i;
1750
1751         err = init_vqs(vi);
1752         if (err)
1753                 return err;
1754
1755         if (netif_running(vi->dev))
1756                 for (i = 0; i < vi->max_queue_pairs; i++)
1757                         virtnet_napi_enable(&vi->rq[i]);
1758
1759         netif_device_attach(vi->dev);
1760
1761         for (i = 0; i < vi->curr_queue_pairs; i++)
1762                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1763                         schedule_delayed_work(&vi->refill, 0);
1764
1765         mutex_lock(&vi->config_lock);
1766         vi->config_enable = true;
1767         mutex_unlock(&vi->config_lock);
1768
1769         rtnl_lock();
1770         virtnet_set_queues(vi, vi->curr_queue_pairs);
1771         rtnl_unlock();
1772
1773         err = register_hotcpu_notifier(&vi->nb);
1774         if (err)
1775                 return err;
1776
1777         return 0;
1778 }
1779 #endif
1780
1781 static struct virtio_device_id id_table[] = {
1782         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1783         { 0 },
1784 };
1785
1786 static unsigned int features[] = {
1787         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1788         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1789         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1790         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1791         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1792         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1793         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1794         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1795         VIRTIO_NET_F_CTRL_MAC_ADDR,
1796         VIRTIO_F_ANY_LAYOUT,
1797 };
1798
1799 static struct virtio_driver virtio_net_driver = {
1800         .feature_table = features,
1801         .feature_table_size = ARRAY_SIZE(features),
1802         .driver.name =  KBUILD_MODNAME,
1803         .driver.owner = THIS_MODULE,
1804         .id_table =     id_table,
1805         .probe =        virtnet_probe,
1806         .remove =       virtnet_remove,
1807         .config_changed = virtnet_config_changed,
1808 #ifdef CONFIG_PM_SLEEP
1809         .freeze =       virtnet_freeze,
1810         .restore =      virtnet_restore,
1811 #endif
1812 };
1813
1814 module_virtio_driver(virtio_net_driver);
1815
1816 MODULE_DEVICE_TABLE(virtio, id_table);
1817 MODULE_DESCRIPTION("Virtio network driver");
1818 MODULE_LICENSE("GPL");