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         virtqueue_kick(rq->vq);
 597         return !oom;
 598 }
 599
 600 static void skb_recv_done(struct virtqueue *rvq)
 601 {
 602         struct virtnet_info *vi = rvq->vdev->priv;
 603         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 604
 605         /* Schedule NAPI, Suppress further interrupts if successful. */
 606         if (napi_schedule_prep(&rq->napi)) {
 607                 virtqueue_disable_cb(rvq);
 608                 __napi_schedule(&rq->napi);
 609         }
 610 }
 611
 612 static void virtnet_napi_enable(struct receive_queue *rq)
 613 {
 614         napi_enable(&rq->napi);
 615
 616         /* If all buffers were filled by other side before we napi_enabled, we
 617          * won't get another interrupt, so process any outstanding packets
 618          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 619          * We synchronize against interrupts via NAPI_STATE_SCHED */
 620         if (napi_schedule_prep(&rq->napi)) {
 621                 virtqueue_disable_cb(rq->vq);
 622                 local_bh_disable();
 623                 __napi_schedule(&rq->napi);
 624                 local_bh_enable();
 625         }
 626 }
 627
 628 static void refill_work(struct work_struct *work)
 629 {
 630         struct virtnet_info *vi =
 631                 container_of(work, struct virtnet_info, refill.work);
 632         bool still_empty;
 633         int i;
 634
 635         for (i = 0; i < vi->curr_queue_pairs; i++) {
 636                 struct receive_queue *rq = &vi->rq[i];
 637
 638                 napi_disable(&rq->napi);
 639                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 640                 virtnet_napi_enable(rq);
 641
 642                 /* In theory, this can happen: if we don't get any buffers in
 643                  * we will *never* try to fill again.
 644                  */
 645                 if (still_empty)
 646                         schedule_delayed_work(&vi->refill, HZ/2);
 647         }
 648 }
 649
 650 static int virtnet_poll(struct napi_struct *napi, int budget)
 651 {
 652         struct receive_queue *rq =
 653                 container_of(napi, struct receive_queue, napi);
 654         struct virtnet_info *vi = rq->vq->vdev->priv;
 655         void *buf;
 656         unsigned int r, len, received = 0;
 657
 658 again:
 659         while (received < budget &&
 660                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 661                 receive_buf(rq, buf, len);
 662                 --rq->num;
 663                 received++;
 664         }
 665
 666         if (rq->num < rq->max / 2) {
 667                 if (!try_fill_recv(rq, GFP_ATOMIC))
 668                         schedule_delayed_work(&vi->refill, 0);
 669         }
 670
 671         /* Out of packets? */
 672         if (received < budget) {
 673                 r = virtqueue_enable_cb_prepare(rq->vq);
 674                 napi_complete(napi);
 675                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 676                     napi_schedule_prep(napi)) {
 677                         virtqueue_disable_cb(rq->vq);
 678                         __napi_schedule(napi);
 679                         goto again;
 680                 }
 681         }
 682
 683         return received;
 684 }
 685
 686 static int virtnet_open(struct net_device *dev)
 687 {
 688         struct virtnet_info *vi = netdev_priv(dev);
 689         int i;
 690
 691         for (i = 0; i < vi->max_queue_pairs; i++) {
 692                 if (i < vi->curr_queue_pairs)
 693                         /* Make sure we have some buffers: if oom use wq. */
 694                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 695                                 schedule_delayed_work(&vi->refill, 0);
 696                 virtnet_napi_enable(&vi->rq[i]);
 697         }
 698
 699         return 0;
 700 }
 701
 702 static void free_old_xmit_skbs(struct send_queue *sq)
 703 {
 704         struct sk_buff *skb;
 705         unsigned int len;
 706         struct virtnet_info *vi = sq->vq->vdev->priv;
 707         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 708
 709         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 710                 pr_debug("Sent skb %p\n", skb);
 711
 712                 u64_stats_update_begin(&stats->tx_syncp);
 713                 stats->tx_bytes += skb->len;
 714                 stats->tx_packets++;
 715                 u64_stats_update_end(&stats->tx_syncp);
 716
 717                 dev_kfree_skb_any(skb);
 718         }
 719 }
 720
 721 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 722 {
 723         struct skb_vnet_hdr *hdr;
 724         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 725         struct virtnet_info *vi = sq->vq->vdev->priv;
 726         unsigned num_sg;
 727         unsigned hdr_len;
 728         bool can_push;
 729
 730         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 731         if (vi->mergeable_rx_bufs)
 732                 hdr_len = sizeof hdr->mhdr;
 733         else
 734                 hdr_len = sizeof hdr->hdr;
 735
 736         can_push = vi->any_header_sg &&
 737                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 738                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 739         /* Even if we can, don't push here yet as this would skew
 740          * csum_start offset below. */
 741         if (can_push)
 742                 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
 743         else
 744                 hdr = skb_vnet_hdr(skb);
 745
 746         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 747                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 748                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 749                 hdr->hdr.csum_offset = skb->csum_offset;
 750         } else {
 751                 hdr->hdr.flags = 0;
 752                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 753         }
 754
 755         if (skb_is_gso(skb)) {
 756                 hdr->hdr.hdr_len = skb_headlen(skb);
 757                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 758                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 759                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 760                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 761                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 762                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 763                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 764                 else
 765                         BUG();
 766                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 767                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 768         } else {
 769                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 770                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 771         }
 772
 773         if (vi->mergeable_rx_bufs)
 774                 hdr->mhdr.num_buffers = 0;
 775
 776         if (can_push) {
 777                 __skb_push(skb, hdr_len);
 778                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 779                 /* Pull header back to avoid skew in tx bytes calculations. */
 780                 __skb_pull(skb, hdr_len);
 781         } else {
 782                 sg_set_buf(sq->sg, hdr, hdr_len);
 783                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 784         }
 785         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 786 }
 787
 788 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 789 {
 790         struct virtnet_info *vi = netdev_priv(dev);
 791         int qnum = skb_get_queue_mapping(skb);
 792         struct send_queue *sq = &vi->sq[qnum];
 793         int err;
 794
 795         /* Free up any pending old buffers before queueing new ones. */
 796         free_old_xmit_skbs(sq);
 797
 798         /* Try to transmit */
 799         err = xmit_skb(sq, skb);
 800
 801         /* This should not happen! */
 802         if (unlikely(err)) {
 803                 dev->stats.tx_fifo_errors++;
 804                 if (net_ratelimit())
 805                         dev_warn(&dev->dev,
 806                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 807                 dev->stats.tx_dropped++;
 808                 kfree_skb(skb);
 809                 return NETDEV_TX_OK;
 810         }
 811         virtqueue_kick(sq->vq);
 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         virtqueue_kick(vi->cvq);
 871
 872         /* Spin for a response, the kick causes an ioport write, trapping
 873          * into the hypervisor, so the request should be handled immediately.
 874          */
 875         while (!virtqueue_get_buf(vi->cvq, &tmp))
 876                 cpu_relax();
 877
 878         return status == VIRTIO_NET_OK;
 879 }
 880
 881 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 882 {
 883         struct virtnet_info *vi = netdev_priv(dev);
 884         struct virtio_device *vdev = vi->vdev;
 885         int ret;
 886         struct sockaddr *addr = p;
 887         struct scatterlist sg;
 888
 889         ret = eth_prepare_mac_addr_change(dev, p);
 890         if (ret)
 891                 return ret;
 892
 893         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 894                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 895                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 896                                           VIRTIO_NET_CTRL_MAC_ADDR_SET,
 897                                           &sg, NULL)) {
 898                         dev_warn(&vdev->dev,
 899                                  "Failed to set mac address by vq command.\n");
 900                         return -EINVAL;
 901                 }
 902         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 903                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 904                                   addr->sa_data, dev->addr_len);
 905         }
 906
 907         eth_commit_mac_addr_change(dev, p);
 908
 909         return 0;
 910 }
 911
 912 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 913                                                struct rtnl_link_stats64 *tot)
 914 {
 915         struct virtnet_info *vi = netdev_priv(dev);
 916         int cpu;
 917         unsigned int start;
 918
 919         for_each_possible_cpu(cpu) {
 920                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 921                 u64 tpackets, tbytes, rpackets, rbytes;
 922
 923                 do {
 924                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 925                         tpackets = stats->tx_packets;
 926                         tbytes   = stats->tx_bytes;
 927                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 928
 929                 do {
 930                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 931                         rpackets = stats->rx_packets;
 932                         rbytes   = stats->rx_bytes;
 933                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 934
 935                 tot->rx_packets += rpackets;
 936                 tot->tx_packets += tpackets;
 937                 tot->rx_bytes   += rbytes;
 938                 tot->tx_bytes   += tbytes;
 939         }
 940
 941         tot->tx_dropped = dev->stats.tx_dropped;
 942         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 943         tot->rx_dropped = dev->stats.rx_dropped;
 944         tot->rx_length_errors = dev->stats.rx_length_errors;
 945         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 946
 947         return tot;
 948 }
 949
 950 #ifdef CONFIG_NET_POLL_CONTROLLER
 951 static void virtnet_netpoll(struct net_device *dev)
 952 {
 953         struct virtnet_info *vi = netdev_priv(dev);
 954         int i;
 955
 956         for (i = 0; i < vi->curr_queue_pairs; i++)
 957                 napi_schedule(&vi->rq[i].napi);
 958 }
 959 #endif
 960
 961 static void virtnet_ack_link_announce(struct virtnet_info *vi)
 962 {
 963         rtnl_lock();
 964         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
 965                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL, NULL))
 966                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
 967         rtnl_unlock();
 968 }
 969
 970 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
 971 {
 972         struct scatterlist sg;
 973         struct virtio_net_ctrl_mq s;
 974         struct net_device *dev = vi->dev;
 975
 976         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
 977                 return 0;
 978
 979         s.virtqueue_pairs = queue_pairs;
 980         sg_init_one(&sg, &s, sizeof(s));
 981
 982         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
 983                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, NULL)) {
 984                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
 985                          queue_pairs);
 986                 return -EINVAL;
 987         } else {
 988                 vi->curr_queue_pairs = queue_pairs;
 989                 /* virtnet_open() will refill when device is going to up. */
 990                 if (dev->flags & IFF_UP)
 991                         schedule_delayed_work(&vi->refill, 0);
 992         }
 993
 994         return 0;
 995 }
 996
 997 static int virtnet_close(struct net_device *dev)
 998 {
 999         struct virtnet_info *vi = netdev_priv(dev);
1000         int i;
1001
1002         /* Make sure refill_work doesn't re-enable napi! */
1003         cancel_delayed_work_sync(&vi->refill);
1004
1005         for (i = 0; i < vi->max_queue_pairs; i++)
1006                 napi_disable(&vi->rq[i].napi);
1007
1008         return 0;
1009 }
1010
1011 static void virtnet_set_rx_mode(struct net_device *dev)
1012 {
1013         struct virtnet_info *vi = netdev_priv(dev);
1014         struct scatterlist sg[2];
1015         u8 promisc, allmulti;
1016         struct virtio_net_ctrl_mac *mac_data;
1017         struct netdev_hw_addr *ha;
1018         int uc_count;
1019         int mc_count;
1020         void *buf;
1021         int i;
1022
1023         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
1024         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1025                 return;
1026
1027         promisc = ((dev->flags & IFF_PROMISC) != 0);
1028         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1029
1030         sg_init_one(sg, &promisc, sizeof(promisc));
1031
1032         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1033                                   VIRTIO_NET_CTRL_RX_PROMISC,
1034                                   sg, NULL))
1035                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1036                          promisc ? "en" : "dis");
1037
1038         sg_init_one(sg, &allmulti, sizeof(allmulti));
1039
1040         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1041                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
1042                                   sg, NULL))
1043                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1044                          allmulti ? "en" : "dis");
1045
1046         uc_count = netdev_uc_count(dev);
1047         mc_count = netdev_mc_count(dev);
1048         /* MAC filter - use one buffer for both lists */
1049         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1050                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1051         mac_data = buf;
1052         if (!buf)
1053                 return;
1054
1055         sg_init_table(sg, 2);
1056
1057         /* Store the unicast list and count in the front of the buffer */
1058         mac_data->entries = uc_count;
1059         i = 0;
1060         netdev_for_each_uc_addr(ha, dev)
1061                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1062
1063         sg_set_buf(&sg[0], mac_data,
1064                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1065
1066         /* multicast list and count fill the end */
1067         mac_data = (void *)&mac_data->macs[uc_count][0];
1068
1069         mac_data->entries = mc_count;
1070         i = 0;
1071         netdev_for_each_mc_addr(ha, dev)
1072                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1073
1074         sg_set_buf(&sg[1], mac_data,
1075                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1076
1077         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1078                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
1079                                   sg, NULL))
1080                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1081
1082         kfree(buf);
1083 }
1084
1085 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1086                                    __be16 proto, u16 vid)
1087 {
1088         struct virtnet_info *vi = netdev_priv(dev);
1089         struct scatterlist sg;
1090
1091         sg_init_one(&sg, &vid, sizeof(vid));
1092
1093         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1094                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, NULL))
1095                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1096         return 0;
1097 }
1098
1099 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1100                                     __be16 proto, u16 vid)
1101 {
1102         struct virtnet_info *vi = netdev_priv(dev);
1103         struct scatterlist sg;
1104
1105         sg_init_one(&sg, &vid, sizeof(vid));
1106
1107         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1108                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, NULL))
1109                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1110         return 0;
1111 }
1112
1113 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1114 {
1115         int i;
1116
1117         if (vi->affinity_hint_set) {
1118                 for (i = 0; i < vi->max_queue_pairs; i++) {
1119                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1120                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1121                 }
1122
1123                 vi->affinity_hint_set = false;
1124         }
1125 }
1126
1127 static void virtnet_set_affinity(struct virtnet_info *vi)
1128 {
1129         int i;
1130         int cpu;
1131
1132         /* In multiqueue mode, when the number of cpu is equal to the number of
1133          * queue pairs, we let the queue pairs to be private to one cpu by
1134          * setting the affinity hint to eliminate the contention.
1135          */
1136         if (vi->curr_queue_pairs == 1 ||
1137             vi->max_queue_pairs != num_online_cpus()) {
1138                 virtnet_clean_affinity(vi, -1);
1139                 return;
1140         }
1141
1142         i = 0;
1143         for_each_online_cpu(cpu) {
1144                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1145                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1146                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1147                 i++;
1148         }
1149
1150         vi->affinity_hint_set = true;
1151 }
1152
1153 static int virtnet_cpu_callback(struct notifier_block *nfb,
1154                                 unsigned long action, void *hcpu)
1155 {
1156         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1157
1158         switch(action & ~CPU_TASKS_FROZEN) {
1159         case CPU_ONLINE:
1160         case CPU_DOWN_FAILED:
1161         case CPU_DEAD:
1162                 virtnet_set_affinity(vi);
1163                 break;
1164         case CPU_DOWN_PREPARE:
1165                 virtnet_clean_affinity(vi, (long)hcpu);
1166                 break;
1167         default:
1168                 break;
1169         }
1170
1171         return NOTIFY_OK;
1172 }
1173
1174 static void virtnet_get_ringparam(struct net_device *dev,
1175                                 struct ethtool_ringparam *ring)
1176 {
1177         struct virtnet_info *vi = netdev_priv(dev);
1178
1179         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1180         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1181         ring->rx_pending = ring->rx_max_pending;
1182         ring->tx_pending = ring->tx_max_pending;
1183 }
1184
1185
1186 static void virtnet_get_drvinfo(struct net_device *dev,
1187                                 struct ethtool_drvinfo *info)
1188 {
1189         struct virtnet_info *vi = netdev_priv(dev);
1190         struct virtio_device *vdev = vi->vdev;
1191
1192         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1193         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1194         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1195
1196 }
1197
1198 /* TODO: Eliminate OOO packets during switching */
1199 static int virtnet_set_channels(struct net_device *dev,
1200                                 struct ethtool_channels *channels)
1201 {
1202         struct virtnet_info *vi = netdev_priv(dev);
1203         u16 queue_pairs = channels->combined_count;
1204         int err;
1205
1206         /* We don't support separate rx/tx channels.
1207          * We don't allow setting 'other' channels.
1208          */
1209         if (channels->rx_count || channels->tx_count || channels->other_count)
1210                 return -EINVAL;
1211
1212         if (queue_pairs > vi->max_queue_pairs)
1213                 return -EINVAL;
1214
1215         get_online_cpus();
1216         err = virtnet_set_queues(vi, queue_pairs);
1217         if (!err) {
1218                 netif_set_real_num_tx_queues(dev, queue_pairs);
1219                 netif_set_real_num_rx_queues(dev, queue_pairs);
1220
1221                 virtnet_set_affinity(vi);
1222         }
1223         put_online_cpus();
1224
1225         return err;
1226 }
1227
1228 static void virtnet_get_channels(struct net_device *dev,
1229                                  struct ethtool_channels *channels)
1230 {
1231         struct virtnet_info *vi = netdev_priv(dev);
1232
1233         channels->combined_count = vi->curr_queue_pairs;
1234         channels->max_combined = vi->max_queue_pairs;
1235         channels->max_other = 0;
1236         channels->rx_count = 0;
1237         channels->tx_count = 0;
1238         channels->other_count = 0;
1239 }
1240
1241 static const struct ethtool_ops virtnet_ethtool_ops = {
1242         .get_drvinfo = virtnet_get_drvinfo,
1243         .get_link = ethtool_op_get_link,
1244         .get_ringparam = virtnet_get_ringparam,
1245         .set_channels = virtnet_set_channels,
1246         .get_channels = virtnet_get_channels,
1247 };
1248
1249 #define MIN_MTU 68
1250 #define MAX_MTU 65535
1251
1252 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1253 {
1254         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1255                 return -EINVAL;
1256         dev->mtu = new_mtu;
1257         return 0;
1258 }
1259
1260 static const struct net_device_ops virtnet_netdev = {
1261         .ndo_open            = virtnet_open,
1262         .ndo_stop            = virtnet_close,
1263         .ndo_start_xmit      = start_xmit,
1264         .ndo_validate_addr   = eth_validate_addr,
1265         .ndo_set_mac_address = virtnet_set_mac_address,
1266         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1267         .ndo_change_mtu      = virtnet_change_mtu,
1268         .ndo_get_stats64     = virtnet_stats,
1269         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1270         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1271 #ifdef CONFIG_NET_POLL_CONTROLLER
1272         .ndo_poll_controller = virtnet_netpoll,
1273 #endif
1274 };
1275
1276 static void virtnet_config_changed_work(struct work_struct *work)
1277 {
1278         struct virtnet_info *vi =
1279                 container_of(work, struct virtnet_info, config_work);
1280         u16 v;
1281
1282         mutex_lock(&vi->config_lock);
1283         if (!vi->config_enable)
1284                 goto done;
1285
1286         if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1287                               offsetof(struct virtio_net_config, status),
1288                               &v) < 0)
1289                 goto done;
1290
1291         if (v & VIRTIO_NET_S_ANNOUNCE) {
1292                 netdev_notify_peers(vi->dev);
1293                 virtnet_ack_link_announce(vi);
1294         }
1295
1296         /* Ignore unknown (future) status bits */
1297         v &= VIRTIO_NET_S_LINK_UP;
1298
1299         if (vi->status == v)
1300                 goto done;
1301
1302         vi->status = v;
1303
1304         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1305                 netif_carrier_on(vi->dev);
1306                 netif_tx_wake_all_queues(vi->dev);
1307         } else {
1308                 netif_carrier_off(vi->dev);
1309                 netif_tx_stop_all_queues(vi->dev);
1310         }
1311 done:
1312         mutex_unlock(&vi->config_lock);
1313 }
1314
1315 static void virtnet_config_changed(struct virtio_device *vdev)
1316 {
1317         struct virtnet_info *vi = vdev->priv;
1318
1319         schedule_work(&vi->config_work);
1320 }
1321
1322 static void virtnet_free_queues(struct virtnet_info *vi)
1323 {
1324         kfree(vi->rq);
1325         kfree(vi->sq);
1326 }
1327
1328 static void free_receive_bufs(struct virtnet_info *vi)
1329 {
1330         int i;
1331
1332         for (i = 0; i < vi->max_queue_pairs; i++) {
1333                 while (vi->rq[i].pages)
1334                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1335         }
1336 }
1337
1338 static void free_unused_bufs(struct virtnet_info *vi)
1339 {
1340         void *buf;
1341         int i;
1342
1343         for (i = 0; i < vi->max_queue_pairs; i++) {
1344                 struct virtqueue *vq = vi->sq[i].vq;
1345                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1346                         dev_kfree_skb(buf);
1347         }
1348
1349         for (i = 0; i < vi->max_queue_pairs; i++) {
1350                 struct virtqueue *vq = vi->rq[i].vq;
1351
1352                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1353                         if (vi->big_packets)
1354                                 give_pages(&vi->rq[i], buf);
1355                         else if (vi->mergeable_rx_bufs)
1356                                 put_page(virt_to_head_page(buf));
1357                         else
1358                                 dev_kfree_skb(buf);
1359                         --vi->rq[i].num;
1360                 }
1361                 BUG_ON(vi->rq[i].num != 0);
1362         }
1363 }
1364
1365 static void virtnet_del_vqs(struct virtnet_info *vi)
1366 {
1367         struct virtio_device *vdev = vi->vdev;
1368
1369         virtnet_clean_affinity(vi, -1);
1370
1371         vdev->config->del_vqs(vdev);
1372
1373         virtnet_free_queues(vi);
1374 }
1375
1376 static int virtnet_find_vqs(struct virtnet_info *vi)
1377 {
1378         vq_callback_t **callbacks;
1379         struct virtqueue **vqs;
1380         int ret = -ENOMEM;
1381         int i, total_vqs;
1382         const char **names;
1383
1384         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1385          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1386          * possible control vq.
1387          */
1388         total_vqs = vi->max_queue_pairs * 2 +
1389                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1390
1391         /* Allocate space for find_vqs parameters */
1392         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1393         if (!vqs)
1394                 goto err_vq;
1395         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1396         if (!callbacks)
1397                 goto err_callback;
1398         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1399         if (!names)
1400                 goto err_names;
1401
1402         /* Parameters for control virtqueue, if any */
1403         if (vi->has_cvq) {
1404                 callbacks[total_vqs - 1] = NULL;
1405                 names[total_vqs - 1] = "control";
1406         }
1407
1408         /* Allocate/initialize parameters for send/receive virtqueues */
1409         for (i = 0; i < vi->max_queue_pairs; i++) {
1410                 callbacks[rxq2vq(i)] = skb_recv_done;
1411                 callbacks[txq2vq(i)] = skb_xmit_done;
1412                 sprintf(vi->rq[i].name, "input.%d", i);
1413                 sprintf(vi->sq[i].name, "output.%d", i);
1414                 names[rxq2vq(i)] = vi->rq[i].name;
1415                 names[txq2vq(i)] = vi->sq[i].name;
1416         }
1417
1418         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1419                                          names);
1420         if (ret)
1421                 goto err_find;
1422
1423         if (vi->has_cvq) {
1424                 vi->cvq = vqs[total_vqs - 1];
1425                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1426                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1427         }
1428
1429         for (i = 0; i < vi->max_queue_pairs; i++) {
1430                 vi->rq[i].vq = vqs[rxq2vq(i)];
1431                 vi->sq[i].vq = vqs[txq2vq(i)];
1432         }
1433
1434         kfree(names);
1435         kfree(callbacks);
1436         kfree(vqs);
1437
1438         return 0;
1439
1440 err_find:
1441         kfree(names);
1442 err_names:
1443         kfree(callbacks);
1444 err_callback:
1445         kfree(vqs);
1446 err_vq:
1447         return ret;
1448 }
1449
1450 static int virtnet_alloc_queues(struct virtnet_info *vi)
1451 {
1452         int i;
1453
1454         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1455         if (!vi->sq)
1456                 goto err_sq;
1457         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1458         if (!vi->rq)
1459                 goto err_rq;
1460
1461         INIT_DELAYED_WORK(&vi->refill, refill_work);
1462         for (i = 0; i < vi->max_queue_pairs; i++) {
1463                 vi->rq[i].pages = NULL;
1464                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1465                                napi_weight);
1466
1467                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1468                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1469         }
1470
1471         return 0;
1472
1473 err_rq:
1474         kfree(vi->sq);
1475 err_sq:
1476         return -ENOMEM;
1477 }
1478
1479 static int init_vqs(struct virtnet_info *vi)
1480 {
1481         int ret;
1482
1483         /* Allocate send & receive queues */
1484         ret = virtnet_alloc_queues(vi);
1485         if (ret)
1486                 goto err;
1487
1488         ret = virtnet_find_vqs(vi);
1489         if (ret)
1490                 goto err_free;
1491
1492         get_online_cpus();
1493         virtnet_set_affinity(vi);
1494         put_online_cpus();
1495
1496         return 0;
1497
1498 err_free:
1499         virtnet_free_queues(vi);
1500 err:
1501         return ret;
1502 }
1503
1504 static int virtnet_probe(struct virtio_device *vdev)
1505 {
1506         int i, err;
1507         struct net_device *dev;
1508         struct virtnet_info *vi;
1509         u16 max_queue_pairs;
1510
1511         /* Find if host supports multiqueue virtio_net device */
1512         err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1513                                 offsetof(struct virtio_net_config,
1514                                 max_virtqueue_pairs), &max_queue_pairs);
1515
1516         /* We need at least 2 queue's */
1517         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1518             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1519             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1520                 max_queue_pairs = 1;
1521
1522         /* Allocate ourselves a network device with room for our info */
1523         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1524         if (!dev)
1525                 return -ENOMEM;
1526
1527         /* Set up network device as normal. */
1528         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1529         dev->netdev_ops = &virtnet_netdev;
1530         dev->features = NETIF_F_HIGHDMA;
1531
1532         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1533         SET_NETDEV_DEV(dev, &vdev->dev);
1534
1535         /* Do we support "hardware" checksums? */
1536         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1537                 /* This opens up the world of extra features. */
1538                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1539                 if (csum)
1540                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1541
1542                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1543                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1544                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1545                 }
1546                 /* Individual feature bits: what can host handle? */
1547                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1548                         dev->hw_features |= NETIF_F_TSO;
1549                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1550                         dev->hw_features |= NETIF_F_TSO6;
1551                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1552                         dev->hw_features |= NETIF_F_TSO_ECN;
1553                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1554                         dev->hw_features |= NETIF_F_UFO;
1555
1556                 if (gso)
1557                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1558                 /* (!csum && gso) case will be fixed by register_netdev() */
1559         }
1560         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1561                 dev->features |= NETIF_F_RXCSUM;
1562
1563         dev->vlan_features = dev->features;
1564
1565         /* Configuration may specify what MAC to use.  Otherwise random. */
1566         if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1567                                   offsetof(struct virtio_net_config, mac),
1568                                   dev->dev_addr, dev->addr_len) < 0)
1569                 eth_hw_addr_random(dev);
1570
1571         /* Set up our device-specific information */
1572         vi = netdev_priv(dev);
1573         vi->dev = dev;
1574         vi->vdev = vdev;
1575         vdev->priv = vi;
1576         vi->stats = alloc_percpu(struct virtnet_stats);
1577         err = -ENOMEM;
1578         if (vi->stats == NULL)
1579                 goto free;
1580
1581         mutex_init(&vi->config_lock);
1582         vi->config_enable = true;
1583         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1584
1585         /* If we can receive ANY GSO packets, we must allocate large ones. */
1586         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1587             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1588             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1589                 vi->big_packets = true;
1590
1591         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1592                 vi->mergeable_rx_bufs = true;
1593
1594         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1595                 vi->any_header_sg = true;
1596
1597         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1598                 vi->has_cvq = true;
1599
1600         /* Use single tx/rx queue pair as default */
1601         vi->curr_queue_pairs = 1;
1602         vi->max_queue_pairs = max_queue_pairs;
1603
1604         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1605         err = init_vqs(vi);
1606         if (err)
1607                 goto free_stats;
1608
1609         netif_set_real_num_tx_queues(dev, 1);
1610         netif_set_real_num_rx_queues(dev, 1);
1611
1612         err = register_netdev(dev);
1613         if (err) {
1614                 pr_debug("virtio_net: registering device failed\n");
1615                 goto free_vqs;
1616         }
1617
1618         /* Last of all, set up some receive buffers. */
1619         for (i = 0; i < vi->curr_queue_pairs; i++) {
1620                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1621
1622                 /* If we didn't even get one input buffer, we're useless. */
1623                 if (vi->rq[i].num == 0) {
1624                         free_unused_bufs(vi);
1625                         err = -ENOMEM;
1626                         goto free_recv_bufs;
1627                 }
1628         }
1629
1630         vi->nb.notifier_call = &virtnet_cpu_callback;
1631         err = register_hotcpu_notifier(&vi->nb);
1632         if (err) {
1633                 pr_debug("virtio_net: registering cpu notifier failed\n");
1634                 goto free_recv_bufs;
1635         }
1636
1637         /* Assume link up if device can't report link status,
1638            otherwise get link status from config. */
1639         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1640                 netif_carrier_off(dev);
1641                 schedule_work(&vi->config_work);
1642         } else {
1643                 vi->status = VIRTIO_NET_S_LINK_UP;
1644                 netif_carrier_on(dev);
1645         }
1646
1647         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1648                  dev->name, max_queue_pairs);
1649
1650         return 0;
1651
1652 free_recv_bufs:
1653         free_receive_bufs(vi);
1654         unregister_netdev(dev);
1655 free_vqs:
1656         cancel_delayed_work_sync(&vi->refill);
1657         virtnet_del_vqs(vi);
1658         if (vi->alloc_frag.page)
1659                 put_page(vi->alloc_frag.page);
1660 free_stats:
1661         free_percpu(vi->stats);
1662 free:
1663         free_netdev(dev);
1664         return err;
1665 }
1666
1667 static void remove_vq_common(struct virtnet_info *vi)
1668 {
1669         vi->vdev->config->reset(vi->vdev);
1670
1671         /* Free unused buffers in both send and recv, if any. */
1672         free_unused_bufs(vi);
1673
1674         free_receive_bufs(vi);
1675
1676         virtnet_del_vqs(vi);
1677 }
1678
1679 static void virtnet_remove(struct virtio_device *vdev)
1680 {
1681         struct virtnet_info *vi = vdev->priv;
1682
1683         unregister_hotcpu_notifier(&vi->nb);
1684
1685         /* Prevent config work handler from accessing the device. */
1686         mutex_lock(&vi->config_lock);
1687         vi->config_enable = false;
1688         mutex_unlock(&vi->config_lock);
1689
1690         unregister_netdev(vi->dev);
1691
1692         remove_vq_common(vi);
1693         if (vi->alloc_frag.page)
1694                 put_page(vi->alloc_frag.page);
1695
1696         flush_work(&vi->config_work);
1697
1698         free_percpu(vi->stats);
1699         free_netdev(vi->dev);
1700 }
1701
1702 #ifdef CONFIG_PM
1703 static int virtnet_freeze(struct virtio_device *vdev)
1704 {
1705         struct virtnet_info *vi = vdev->priv;
1706         int i;
1707
1708         unregister_hotcpu_notifier(&vi->nb);
1709
1710         /* Prevent config work handler from accessing the device */
1711         mutex_lock(&vi->config_lock);
1712         vi->config_enable = false;
1713         mutex_unlock(&vi->config_lock);
1714
1715         netif_device_detach(vi->dev);
1716         cancel_delayed_work_sync(&vi->refill);
1717
1718         if (netif_running(vi->dev))
1719                 for (i = 0; i < vi->max_queue_pairs; i++) {
1720                         napi_disable(&vi->rq[i].napi);
1721                         netif_napi_del(&vi->rq[i].napi);
1722                 }
1723
1724         remove_vq_common(vi);
1725
1726         flush_work(&vi->config_work);
1727
1728         return 0;
1729 }
1730
1731 static int virtnet_restore(struct virtio_device *vdev)
1732 {
1733         struct virtnet_info *vi = vdev->priv;
1734         int err, i;
1735
1736         err = init_vqs(vi);
1737         if (err)
1738                 return err;
1739
1740         if (netif_running(vi->dev))
1741                 for (i = 0; i < vi->max_queue_pairs; i++)
1742                         virtnet_napi_enable(&vi->rq[i]);
1743
1744         netif_device_attach(vi->dev);
1745
1746         for (i = 0; i < vi->curr_queue_pairs; i++)
1747                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1748                         schedule_delayed_work(&vi->refill, 0);
1749
1750         mutex_lock(&vi->config_lock);
1751         vi->config_enable = true;
1752         mutex_unlock(&vi->config_lock);
1753
1754         rtnl_lock();
1755         virtnet_set_queues(vi, vi->curr_queue_pairs);
1756         rtnl_unlock();
1757
1758         err = register_hotcpu_notifier(&vi->nb);
1759         if (err)
1760                 return err;
1761
1762         return 0;
1763 }
1764 #endif
1765
1766 static struct virtio_device_id id_table[] = {
1767         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1768         { 0 },
1769 };
1770
1771 static unsigned int features[] = {
1772         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1773         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1774         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1775         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1776         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1777         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1778         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1779         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1780         VIRTIO_NET_F_CTRL_MAC_ADDR,
1781         VIRTIO_F_ANY_LAYOUT,
1782 };
1783
1784 static struct virtio_driver virtio_net_driver = {
1785         .feature_table = features,
1786         .feature_table_size = ARRAY_SIZE(features),
1787         .driver.name =  KBUILD_MODNAME,
1788         .driver.owner = THIS_MODULE,
1789         .id_table =     id_table,
1790         .probe =        virtnet_probe,
1791         .remove =       virtnet_remove,
1792         .config_changed = virtnet_config_changed,
1793 #ifdef CONFIG_PM
1794         .freeze =       virtnet_freeze,
1795         .restore =      virtnet_restore,
1796 #endif
1797 };
1798
1799 module_virtio_driver(virtio_net_driver);
1800
1801 MODULE_DEVICE_TABLE(virtio, id_table);
1802 MODULE_DESCRIPTION("Virtio network driver");
1803 MODULE_LICENSE("GPL");