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 = 128;
  32 module_param(napi_weight, int, 0444);
  33
  34 static bool csum = true, gso = true;
  35 module_param(csum, bool, 0444);
  36 module_param(gso, bool, 0444);
  37
  38 /* FIXME: MTU in config. */
  39 #define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  40 #define GOOD_COPY_LEN   128
  41
  42 #define VIRTNET_SEND_COMMAND_SG_MAX    2
  43 #define VIRTNET_DRIVER_VERSION "1.0.0"
  44
  45 struct virtnet_stats {
  46         struct u64_stats_sync tx_syncp;
  47         struct u64_stats_sync rx_syncp;
  48         u64 tx_bytes;
  49         u64 tx_packets;
  50
  51         u64 rx_bytes;
  52         u64 rx_packets;
  53 };
  54
  55 /* Internal representation of a send virtqueue */
  56 struct send_queue {
  57         /* Virtqueue associated with this send _queue */
  58         struct virtqueue *vq;
  59
  60         /* TX: fragments + linear part + virtio header */
  61         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  62
  63         /* Name of the send queue: output.$index */
  64         char name[40];
  65 };
  66
  67 /* Internal representation of a receive virtqueue */
  68 struct receive_queue {
  69         /* Virtqueue associated with this receive_queue */
  70         struct virtqueue *vq;
  71
  72         struct napi_struct napi;
  73
  74         /* Number of input buffers, and max we've ever had. */
  75         unsigned int num, max;
  76
  77         /* Chain pages by the private ptr. */
  78         struct page *pages;
  79
  80         /* RX: fragments + linear part + virtio header */
  81         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  82
  83         /* Name of this receive queue: input.$index */
  84         char name[40];
  85 };
  86
  87 struct virtnet_info {
  88         struct virtio_device *vdev;
  89         struct virtqueue *cvq;
  90         struct net_device *dev;
  91         struct send_queue *sq;
  92         struct receive_queue *rq;
  93         unsigned int status;
  94
  95         /* Max # of queue pairs supported by the device */
  96         u16 max_queue_pairs;
  97
  98         /* # of queue pairs currently used by the driver */
  99         u16 curr_queue_pairs;
 100
 101         /* I like... big packets and I cannot lie! */
 102         bool big_packets;
 103
 104         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 105         bool mergeable_rx_bufs;
 106
 107         /* Has control virtqueue */
 108         bool has_cvq;
 109
 110         /* enable config space updates */
 111         bool config_enable;
 112
 113         /* Active statistics */
 114         struct virtnet_stats __percpu *stats;
 115
 116         /* Work struct for refilling if we run low on memory. */
 117         struct delayed_work refill;
 118
 119         /* Work struct for config space updates */
 120         struct work_struct config_work;
 121
 122         /* Lock for config space updates */
 123         struct mutex config_lock;
 124
 125         /* Does the affinity hint is set for virtqueues? */
 126         bool affinity_hint_set;
 127
 128         /* Per-cpu variable to show the mapping from CPU to virtqueue */
 129         int __percpu *vq_index;
 130
 131         /* CPU hot plug notifier */
 132         struct notifier_block nb;
 133 };
 134
 135 struct skb_vnet_hdr {
 136         union {
 137                 struct virtio_net_hdr hdr;
 138                 struct virtio_net_hdr_mrg_rxbuf mhdr;
 139         };
 140 };
 141
 142 struct padded_vnet_hdr {
 143         struct virtio_net_hdr hdr;
 144         /*
 145          * virtio_net_hdr should be in a separated sg buffer because of a
 146          * QEMU bug, and data sg buffer shares same page with this header sg.
 147          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
 148          */
 149         char padding[6];
 150 };
 151
 152 /* Converting between virtqueue no. and kernel tx/rx queue no.
 153  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 154  */
 155 static int vq2txq(struct virtqueue *vq)
 156 {
 157         return (virtqueue_get_queue_index(vq) - 1) / 2;
 158 }
 159
 160 static int txq2vq(int txq)
 161 {
 162         return txq * 2 + 1;
 163 }
 164
 165 static int vq2rxq(struct virtqueue *vq)
 166 {
 167         return virtqueue_get_queue_index(vq) / 2;
 168 }
 169
 170 static int rxq2vq(int rxq)
 171 {
 172         return rxq * 2;
 173 }
 174
 175 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 176 {
 177         return (struct skb_vnet_hdr *)skb->cb;
 178 }
 179
 180 /*
 181  * private is used to chain pages for big packets, put the whole
 182  * most recent used list in the beginning for reuse
 183  */
 184 static void give_pages(struct receive_queue *rq, struct page *page)
 185 {
 186         struct page *end;
 187
 188         /* Find end of list, sew whole thing into vi->rq.pages. */
 189         for (end = page; end->private; end = (struct page *)end->private);
 190         end->private = (unsigned long)rq->pages;
 191         rq->pages = page;
 192 }
 193
 194 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 195 {
 196         struct page *p = rq->pages;
 197
 198         if (p) {
 199                 rq->pages = (struct page *)p->private;
 200                 /* clear private here, it is used to chain pages */
 201                 p->private = 0;
 202         } else
 203                 p = alloc_page(gfp_mask);
 204         return p;
 205 }
 206
 207 static void skb_xmit_done(struct virtqueue *vq)
 208 {
 209         struct virtnet_info *vi = vq->vdev->priv;
 210
 211         /* Suppress further interrupts. */
 212         virtqueue_disable_cb(vq);
 213
 214         /* We were probably waiting for more output buffers. */
 215         netif_wake_subqueue(vi->dev, vq2txq(vq));
 216 }
 217
 218 static void set_skb_frag(struct sk_buff *skb, struct page *page,
 219                          unsigned int offset, unsigned int *len)
 220 {
 221         int size = min((unsigned)PAGE_SIZE - offset, *len);
 222         int i = skb_shinfo(skb)->nr_frags;
 223
 224         __skb_fill_page_desc(skb, i, page, offset, size);
 225
 226         skb->data_len += size;
 227         skb->len += size;
 228         skb->truesize += PAGE_SIZE;
 229         skb_shinfo(skb)->nr_frags++;
 230         skb_shinfo(skb)->gso_type |= SKB_GSO_SHARED_FRAG;
 231         *len -= size;
 232 }
 233
 234 /* Called from bottom half context */
 235 static struct sk_buff *page_to_skb(struct receive_queue *rq,
 236                                    struct page *page, unsigned int len)
 237 {
 238         struct virtnet_info *vi = rq->vq->vdev->priv;
 239         struct sk_buff *skb;
 240         struct skb_vnet_hdr *hdr;
 241         unsigned int copy, hdr_len, offset;
 242         char *p;
 243
 244         p = page_address(page);
 245
 246         /* copy small packet so we can reuse these pages for small data */
 247         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 248         if (unlikely(!skb))
 249                 return NULL;
 250
 251         hdr = skb_vnet_hdr(skb);
 252
 253         if (vi->mergeable_rx_bufs) {
 254                 hdr_len = sizeof hdr->mhdr;
 255                 offset = hdr_len;
 256         } else {
 257                 hdr_len = sizeof hdr->hdr;
 258                 offset = sizeof(struct padded_vnet_hdr);
 259         }
 260
 261         memcpy(hdr, p, hdr_len);
 262
 263         len -= hdr_len;
 264         p += offset;
 265
 266         copy = len;
 267         if (copy > skb_tailroom(skb))
 268                 copy = skb_tailroom(skb);
 269         memcpy(skb_put(skb, copy), p, copy);
 270
 271         len -= copy;
 272         offset += copy;
 273
 274         /*
 275          * Verify that we can indeed put this data into a skb.
 276          * This is here to handle cases when the device erroneously
 277          * tries to receive more than is possible. This is usually
 278          * the case of a broken device.
 279          */
 280         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 281                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 282                 dev_kfree_skb(skb);
 283                 return NULL;
 284         }
 285
 286         while (len) {
 287                 set_skb_frag(skb, page, offset, &len);
 288                 page = (struct page *)page->private;
 289                 offset = 0;
 290         }
 291
 292         if (page)
 293                 give_pages(rq, page);
 294
 295         return skb;
 296 }
 297
 298 static int receive_mergeable(struct receive_queue *rq, struct sk_buff *skb)
 299 {
 300         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 301         struct page *page;
 302         int num_buf, i, len;
 303
 304         num_buf = hdr->mhdr.num_buffers;
 305         while (--num_buf) {
 306                 i = skb_shinfo(skb)->nr_frags;
 307                 if (i >= MAX_SKB_FRAGS) {
 308                         pr_debug("%s: packet too long\n", skb->dev->name);
 309                         skb->dev->stats.rx_length_errors++;
 310                         return -EINVAL;
 311                 }
 312                 page = virtqueue_get_buf(rq->vq, &len);
 313                 if (!page) {
 314                         pr_debug("%s: rx error: %d buffers missing\n",
 315                                  skb->dev->name, hdr->mhdr.num_buffers);
 316                         skb->dev->stats.rx_length_errors++;
 317                         return -EINVAL;
 318                 }
 319
 320                 if (len > PAGE_SIZE)
 321                         len = PAGE_SIZE;
 322
 323                 set_skb_frag(skb, page, 0, &len);
 324
 325                 --rq->num;
 326         }
 327         return 0;
 328 }
 329
 330 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
 331 {
 332         struct virtnet_info *vi = rq->vq->vdev->priv;
 333         struct net_device *dev = vi->dev;
 334         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 335         struct sk_buff *skb;
 336         struct page *page;
 337         struct skb_vnet_hdr *hdr;
 338
 339         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 340                 pr_debug("%s: short packet %i\n", dev->name, len);
 341                 dev->stats.rx_length_errors++;
 342                 if (vi->mergeable_rx_bufs || vi->big_packets)
 343                         give_pages(rq, buf);
 344                 else
 345                         dev_kfree_skb(buf);
 346                 return;
 347         }
 348
 349         if (!vi->mergeable_rx_bufs && !vi->big_packets) {
 350                 skb = buf;
 351                 len -= sizeof(struct virtio_net_hdr);
 352                 skb_trim(skb, len);
 353         } else {
 354                 page = buf;
 355                 skb = page_to_skb(rq, page, len);
 356                 if (unlikely(!skb)) {
 357                         dev->stats.rx_dropped++;
 358                         give_pages(rq, page);
 359                         return;
 360                 }
 361                 if (vi->mergeable_rx_bufs)
 362                         if (receive_mergeable(rq, skb)) {
 363                                 dev_kfree_skb(skb);
 364                                 return;
 365                         }
 366         }
 367
 368         hdr = skb_vnet_hdr(skb);
 369
 370         u64_stats_update_begin(&stats->rx_syncp);
 371         stats->rx_bytes += skb->len;
 372         stats->rx_packets++;
 373         u64_stats_update_end(&stats->rx_syncp);
 374
 375         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 376                 pr_debug("Needs csum!\n");
 377                 if (!skb_partial_csum_set(skb,
 378                                           hdr->hdr.csum_start,
 379                                           hdr->hdr.csum_offset))
 380                         goto frame_err;
 381         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 382                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 383         }
 384
 385         skb->protocol = eth_type_trans(skb, dev);
 386         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 387                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 388
 389         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 390                 unsigned short gso_type = 0;
 391
 392                 pr_debug("GSO!\n");
 393                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 394                 case VIRTIO_NET_HDR_GSO_TCPV4:
 395                         gso_type = SKB_GSO_TCPV4;
 396                         break;
 397                 case VIRTIO_NET_HDR_GSO_UDP:
 398                         gso_type = SKB_GSO_UDP;
 399                         break;
 400                 case VIRTIO_NET_HDR_GSO_TCPV6:
 401                         gso_type = SKB_GSO_TCPV6;
 402                         break;
 403                 default:
 404                         net_warn_ratelimited("%s: bad gso type %u.\n",
 405                                              dev->name, hdr->hdr.gso_type);
 406                         goto frame_err;
 407                 }
 408
 409                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 410                         gso_type |= SKB_GSO_TCP_ECN;
 411
 412                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 413                 if (skb_shinfo(skb)->gso_size == 0) {
 414                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 415                         goto frame_err;
 416                 }
 417
 418                 skb_shinfo(skb)->gso_type |= gso_type;
 419                 /* Header must be checked, and gso_segs computed. */
 420                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 421                 skb_shinfo(skb)->gso_segs = 0;
 422         }
 423
 424         netif_receive_skb(skb);
 425         return;
 426
 427 frame_err:
 428         dev->stats.rx_frame_errors++;
 429         dev_kfree_skb(skb);
 430 }
 431
 432 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
 433 {
 434         struct virtnet_info *vi = rq->vq->vdev->priv;
 435         struct sk_buff *skb;
 436         struct skb_vnet_hdr *hdr;
 437         int err;
 438
 439         skb = __netdev_alloc_skb_ip_align(vi->dev, MAX_PACKET_LEN, gfp);
 440         if (unlikely(!skb))
 441                 return -ENOMEM;
 442
 443         skb_put(skb, MAX_PACKET_LEN);
 444
 445         hdr = skb_vnet_hdr(skb);
 446         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
 447
 448         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 449
 450         err = virtqueue_add_buf(rq->vq, rq->sg, 0, 2, skb, gfp);
 451         if (err < 0)
 452                 dev_kfree_skb(skb);
 453
 454         return err;
 455 }
 456
 457 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
 458 {
 459         struct page *first, *list = NULL;
 460         char *p;
 461         int i, err, offset;
 462
 463         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 464         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 465                 first = get_a_page(rq, gfp);
 466                 if (!first) {
 467                         if (list)
 468                                 give_pages(rq, list);
 469                         return -ENOMEM;
 470                 }
 471                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 472
 473                 /* chain new page in list head to match sg */
 474                 first->private = (unsigned long)list;
 475                 list = first;
 476         }
 477
 478         first = get_a_page(rq, gfp);
 479         if (!first) {
 480                 give_pages(rq, list);
 481                 return -ENOMEM;
 482         }
 483         p = page_address(first);
 484
 485         /* rq->sg[0], rq->sg[1] share the same page */
 486         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
 487         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
 488
 489         /* rq->sg[1] for data packet, from offset */
 490         offset = sizeof(struct padded_vnet_hdr);
 491         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 492
 493         /* chain first in list head */
 494         first->private = (unsigned long)list;
 495         err = virtqueue_add_buf(rq->vq, rq->sg, 0, MAX_SKB_FRAGS + 2,
 496                                 first, gfp);
 497         if (err < 0)
 498                 give_pages(rq, first);
 499
 500         return err;
 501 }
 502
 503 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 504 {
 505         struct page *page;
 506         int err;
 507
 508         page = get_a_page(rq, gfp);
 509         if (!page)
 510                 return -ENOMEM;
 511
 512         sg_init_one(rq->sg, page_address(page), PAGE_SIZE);
 513
 514         err = virtqueue_add_buf(rq->vq, rq->sg, 0, 1, page, gfp);
 515         if (err < 0)
 516                 give_pages(rq, page);
 517
 518         return err;
 519 }
 520
 521 /*
 522  * Returns false if we couldn't fill entirely (OOM).
 523  *
 524  * Normally run in the receive path, but can also be run from ndo_open
 525  * before we're receiving packets, or from refill_work which is
 526  * careful to disable receiving (using napi_disable).
 527  */
 528 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
 529 {
 530         struct virtnet_info *vi = rq->vq->vdev->priv;
 531         int err;
 532         bool oom;
 533
 534         do {
 535                 if (vi->mergeable_rx_bufs)
 536                         err = add_recvbuf_mergeable(rq, gfp);
 537                 else if (vi->big_packets)
 538                         err = add_recvbuf_big(rq, gfp);
 539                 else
 540                         err = add_recvbuf_small(rq, gfp);
 541
 542                 oom = err == -ENOMEM;
 543                 if (err)
 544                         break;
 545                 ++rq->num;
 546         } while (rq->vq->num_free);
 547         if (unlikely(rq->num > rq->max))
 548                 rq->max = rq->num;
 549         virtqueue_kick(rq->vq);
 550         return !oom;
 551 }
 552
 553 static void skb_recv_done(struct virtqueue *rvq)
 554 {
 555         struct virtnet_info *vi = rvq->vdev->priv;
 556         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 557
 558         /* Schedule NAPI, Suppress further interrupts if successful. */
 559         if (napi_schedule_prep(&rq->napi)) {
 560                 virtqueue_disable_cb(rvq);
 561                 __napi_schedule(&rq->napi);
 562         }
 563 }
 564
 565 static void virtnet_napi_enable(struct receive_queue *rq)
 566 {
 567         napi_enable(&rq->napi);
 568
 569         /* If all buffers were filled by other side before we napi_enabled, we
 570          * won't get another interrupt, so process any outstanding packets
 571          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 572          * We synchronize against interrupts via NAPI_STATE_SCHED */
 573         if (napi_schedule_prep(&rq->napi)) {
 574                 virtqueue_disable_cb(rq->vq);
 575                 local_bh_disable();
 576                 __napi_schedule(&rq->napi);
 577                 local_bh_enable();
 578         }
 579 }
 580
 581 static void refill_work(struct work_struct *work)
 582 {
 583         struct virtnet_info *vi =
 584                 container_of(work, struct virtnet_info, refill.work);
 585         bool still_empty;
 586         int i;
 587
 588         for (i = 0; i < vi->max_queue_pairs; i++) {
 589                 struct receive_queue *rq = &vi->rq[i];
 590
 591                 napi_disable(&rq->napi);
 592                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 593                 virtnet_napi_enable(rq);
 594
 595                 /* In theory, this can happen: if we don't get any buffers in
 596                  * we will *never* try to fill again.
 597                  */
 598                 if (still_empty)
 599                         schedule_delayed_work(&vi->refill, HZ/2);
 600         }
 601 }
 602
 603 static int virtnet_poll(struct napi_struct *napi, int budget)
 604 {
 605         struct receive_queue *rq =
 606                 container_of(napi, struct receive_queue, napi);
 607         struct virtnet_info *vi = rq->vq->vdev->priv;
 608         void *buf;
 609         unsigned int len, received = 0;
 610
 611 again:
 612         while (received < budget &&
 613                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 614                 receive_buf(rq, buf, len);
 615                 --rq->num;
 616                 received++;
 617         }
 618
 619         if (rq->num < rq->max / 2) {
 620                 if (!try_fill_recv(rq, GFP_ATOMIC))
 621                         schedule_delayed_work(&vi->refill, 0);
 622         }
 623
 624         /* Out of packets? */
 625         if (received < budget) {
 626                 napi_complete(napi);
 627                 if (unlikely(!virtqueue_enable_cb(rq->vq)) &&
 628                     napi_schedule_prep(napi)) {
 629                         virtqueue_disable_cb(rq->vq);
 630                         __napi_schedule(napi);
 631                         goto again;
 632                 }
 633         }
 634
 635         return received;
 636 }
 637
 638 static int virtnet_open(struct net_device *dev)
 639 {
 640         struct virtnet_info *vi = netdev_priv(dev);
 641         int i;
 642
 643         for (i = 0; i < vi->max_queue_pairs; i++) {
 644                 /* Make sure we have some buffers: if oom use wq. */
 645                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 646                         schedule_delayed_work(&vi->refill, 0);
 647                 virtnet_napi_enable(&vi->rq[i]);
 648         }
 649
 650         return 0;
 651 }
 652
 653 static void free_old_xmit_skbs(struct send_queue *sq)
 654 {
 655         struct sk_buff *skb;
 656         unsigned int len;
 657         struct virtnet_info *vi = sq->vq->vdev->priv;
 658         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 659
 660         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 661                 pr_debug("Sent skb %p\n", skb);
 662
 663                 u64_stats_update_begin(&stats->tx_syncp);
 664                 stats->tx_bytes += skb->len;
 665                 stats->tx_packets++;
 666                 u64_stats_update_end(&stats->tx_syncp);
 667
 668                 dev_kfree_skb_any(skb);
 669         }
 670 }
 671
 672 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 673 {
 674         struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
 675         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 676         struct virtnet_info *vi = sq->vq->vdev->priv;
 677         unsigned num_sg;
 678
 679         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 680
 681         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 682                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 683                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 684                 hdr->hdr.csum_offset = skb->csum_offset;
 685         } else {
 686                 hdr->hdr.flags = 0;
 687                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 688         }
 689
 690         if (skb_is_gso(skb)) {
 691                 hdr->hdr.hdr_len = skb_headlen(skb);
 692                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 693                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 694                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 695                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 696                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 697                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 698                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 699                 else
 700                         BUG();
 701                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 702                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 703         } else {
 704                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 705                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 706         }
 707
 708         hdr->mhdr.num_buffers = 0;
 709
 710         /* Encode metadata header at front. */
 711         if (vi->mergeable_rx_bufs)
 712                 sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
 713         else
 714                 sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
 715
 716         num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 717         return virtqueue_add_buf(sq->vq, sq->sg, num_sg,
 718                                  0, skb, GFP_ATOMIC);
 719 }
 720
 721 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 722 {
 723         struct virtnet_info *vi = netdev_priv(dev);
 724         int qnum = skb_get_queue_mapping(skb);
 725         struct send_queue *sq = &vi->sq[qnum];
 726         int err;
 727
 728         /* Free up any pending old buffers before queueing new ones. */
 729         free_old_xmit_skbs(sq);
 730
 731         /* Try to transmit */
 732         err = xmit_skb(sq, skb);
 733
 734         /* This should not happen! */
 735         if (unlikely(err)) {
 736                 dev->stats.tx_fifo_errors++;
 737                 if (net_ratelimit())
 738                         dev_warn(&dev->dev,
 739                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 740                 dev->stats.tx_dropped++;
 741                 kfree_skb(skb);
 742                 return NETDEV_TX_OK;
 743         }
 744         virtqueue_kick(sq->vq);
 745
 746         /* Don't wait up for transmitted skbs to be freed. */
 747         skb_orphan(skb);
 748         nf_reset(skb);
 749
 750         /* Apparently nice girls don't return TX_BUSY; stop the queue
 751          * before it gets out of hand.  Naturally, this wastes entries. */
 752         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 753                 netif_stop_subqueue(dev, qnum);
 754                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 755                         /* More just got used, free them then recheck. */
 756                         free_old_xmit_skbs(sq);
 757                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 758                                 netif_start_subqueue(dev, qnum);
 759                                 virtqueue_disable_cb(sq->vq);
 760                         }
 761                 }
 762         }
 763
 764         return NETDEV_TX_OK;
 765 }
 766
 767 /*
 768  * Send command via the control virtqueue and check status.  Commands
 769  * supported by the hypervisor, as indicated by feature bits, should
 770  * never fail unless improperly formated.
 771  */
 772 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 773                                  struct scatterlist *data, int out, int in)
 774 {
 775         struct scatterlist *s, sg[VIRTNET_SEND_COMMAND_SG_MAX + 2];
 776         struct virtio_net_ctrl_hdr ctrl;
 777         virtio_net_ctrl_ack status = ~0;
 778         unsigned int tmp;
 779         int i;
 780
 781         /* Caller should know better */
 782         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ) ||
 783                 (out + in > VIRTNET_SEND_COMMAND_SG_MAX));
 784
 785         out++; /* Add header */
 786         in++; /* Add return status */
 787
 788         ctrl.class = class;
 789         ctrl.cmd = cmd;
 790
 791         sg_init_table(sg, out + in);
 792
 793         sg_set_buf(&sg[0], &ctrl, sizeof(ctrl));
 794         for_each_sg(data, s, out + in - 2, i)
 795                 sg_set_buf(&sg[i + 1], sg_virt(s), s->length);
 796         sg_set_buf(&sg[out + in - 1], &status, sizeof(status));
 797
 798         BUG_ON(virtqueue_add_buf(vi->cvq, sg, out, in, vi, GFP_ATOMIC) < 0);
 799
 800         virtqueue_kick(vi->cvq);
 801
 802         /* Spin for a response, the kick causes an ioport write, trapping
 803          * into the hypervisor, so the request should be handled immediately.
 804          */
 805         while (!virtqueue_get_buf(vi->cvq, &tmp))
 806                 cpu_relax();
 807
 808         return status == VIRTIO_NET_OK;
 809 }
 810
 811 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 812 {
 813         struct virtnet_info *vi = netdev_priv(dev);
 814         struct virtio_device *vdev = vi->vdev;
 815         int ret;
 816         struct sockaddr *addr = p;
 817         struct scatterlist sg;
 818
 819         ret = eth_prepare_mac_addr_change(dev, p);
 820         if (ret)
 821                 return ret;
 822
 823         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 824                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 825                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 826                                           VIRTIO_NET_CTRL_MAC_ADDR_SET,
 827                                           &sg, 1, 0)) {
 828                         dev_warn(&vdev->dev,
 829                                  "Failed to set mac address by vq command.\n");
 830                         return -EINVAL;
 831                 }
 832         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 833                 vdev->config->set(vdev, offsetof(struct virtio_net_config, mac),
 834                                   addr->sa_data, dev->addr_len);
 835         }
 836
 837         eth_commit_mac_addr_change(dev, p);
 838
 839         return 0;
 840 }
 841
 842 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 843                                                struct rtnl_link_stats64 *tot)
 844 {
 845         struct virtnet_info *vi = netdev_priv(dev);
 846         int cpu;
 847         unsigned int start;
 848
 849         for_each_possible_cpu(cpu) {
 850                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 851                 u64 tpackets, tbytes, rpackets, rbytes;
 852
 853                 do {
 854                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 855                         tpackets = stats->tx_packets;
 856                         tbytes   = stats->tx_bytes;
 857                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 858
 859                 do {
 860                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 861                         rpackets = stats->rx_packets;
 862                         rbytes   = stats->rx_bytes;
 863                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 864
 865                 tot->rx_packets += rpackets;
 866                 tot->tx_packets += tpackets;
 867                 tot->rx_bytes   += rbytes;
 868                 tot->tx_bytes   += tbytes;
 869         }
 870
 871         tot->tx_dropped = dev->stats.tx_dropped;
 872         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 873         tot->rx_dropped = dev->stats.rx_dropped;
 874         tot->rx_length_errors = dev->stats.rx_length_errors;
 875         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 876
 877         return tot;
 878 }
 879
 880 #ifdef CONFIG_NET_POLL_CONTROLLER
 881 static void virtnet_netpoll(struct net_device *dev)
 882 {
 883         struct virtnet_info *vi = netdev_priv(dev);
 884         int i;
 885
 886         for (i = 0; i < vi->curr_queue_pairs; i++)
 887                 napi_schedule(&vi->rq[i].napi);
 888 }
 889 #endif
 890
 891 static void virtnet_ack_link_announce(struct virtnet_info *vi)
 892 {
 893         rtnl_lock();
 894         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
 895                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL,
 896                                   0, 0))
 897                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
 898         rtnl_unlock();
 899 }
 900
 901 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
 902 {
 903         struct scatterlist sg;
 904         struct virtio_net_ctrl_mq s;
 905         struct net_device *dev = vi->dev;
 906
 907         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
 908                 return 0;
 909
 910         s.virtqueue_pairs = queue_pairs;
 911         sg_init_one(&sg, &s, sizeof(s));
 912
 913         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
 914                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg, 1, 0)){
 915                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
 916                          queue_pairs);
 917                 return -EINVAL;
 918         } else
 919                 vi->curr_queue_pairs = queue_pairs;
 920
 921         return 0;
 922 }
 923
 924 static int virtnet_close(struct net_device *dev)
 925 {
 926         struct virtnet_info *vi = netdev_priv(dev);
 927         int i;
 928
 929         /* Make sure refill_work doesn't re-enable napi! */
 930         cancel_delayed_work_sync(&vi->refill);
 931
 932         for (i = 0; i < vi->max_queue_pairs; i++)
 933                 napi_disable(&vi->rq[i].napi);
 934
 935         return 0;
 936 }
 937
 938 static void virtnet_set_rx_mode(struct net_device *dev)
 939 {
 940         struct virtnet_info *vi = netdev_priv(dev);
 941         struct scatterlist sg[2];
 942         u8 promisc, allmulti;
 943         struct virtio_net_ctrl_mac *mac_data;
 944         struct netdev_hw_addr *ha;
 945         int uc_count;
 946         int mc_count;
 947         void *buf;
 948         int i;
 949
 950         /* We can't dynamicaly set ndo_set_rx_mode, so return gracefully */
 951         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
 952                 return;
 953
 954         promisc = ((dev->flags & IFF_PROMISC) != 0);
 955         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
 956
 957         sg_init_one(sg, &promisc, sizeof(promisc));
 958
 959         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 960                                   VIRTIO_NET_CTRL_RX_PROMISC,
 961                                   sg, 1, 0))
 962                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
 963                          promisc ? "en" : "dis");
 964
 965         sg_init_one(sg, &allmulti, sizeof(allmulti));
 966
 967         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
 968                                   VIRTIO_NET_CTRL_RX_ALLMULTI,
 969                                   sg, 1, 0))
 970                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
 971                          allmulti ? "en" : "dis");
 972
 973         uc_count = netdev_uc_count(dev);
 974         mc_count = netdev_mc_count(dev);
 975         /* MAC filter - use one buffer for both lists */
 976         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
 977                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
 978         mac_data = buf;
 979         if (!buf) {
 980                 dev_warn(&dev->dev, "No memory for MAC address buffer\n");
 981                 return;
 982         }
 983
 984         sg_init_table(sg, 2);
 985
 986         /* Store the unicast list and count in the front of the buffer */
 987         mac_data->entries = uc_count;
 988         i = 0;
 989         netdev_for_each_uc_addr(ha, dev)
 990                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
 991
 992         sg_set_buf(&sg[0], mac_data,
 993                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
 994
 995         /* multicast list and count fill the end */
 996         mac_data = (void *)&mac_data->macs[uc_count][0];
 997
 998         mac_data->entries = mc_count;
 999         i = 0;
1000         netdev_for_each_mc_addr(ha, dev)
1001                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1002
1003         sg_set_buf(&sg[1], mac_data,
1004                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1005
1006         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1007                                   VIRTIO_NET_CTRL_MAC_TABLE_SET,
1008                                   sg, 2, 0))
1009                 dev_warn(&dev->dev, "Failed to set MAC fitler table.\n");
1010
1011         kfree(buf);
1012 }
1013
1014 static int virtnet_vlan_rx_add_vid(struct net_device *dev, u16 vid)
1015 {
1016         struct virtnet_info *vi = netdev_priv(dev);
1017         struct scatterlist sg;
1018
1019         sg_init_one(&sg, &vid, sizeof(vid));
1020
1021         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1022                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg, 1, 0))
1023                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1024         return 0;
1025 }
1026
1027 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, u16 vid)
1028 {
1029         struct virtnet_info *vi = netdev_priv(dev);
1030         struct scatterlist sg;
1031
1032         sg_init_one(&sg, &vid, sizeof(vid));
1033
1034         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1035                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg, 1, 0))
1036                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1037         return 0;
1038 }
1039
1040 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1041 {
1042         int i;
1043         int cpu;
1044
1045         if (vi->affinity_hint_set) {
1046                 for (i = 0; i < vi->max_queue_pairs; i++) {
1047                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1048                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1049                 }
1050
1051                 vi->affinity_hint_set = false;
1052         }
1053
1054         i = 0;
1055         for_each_online_cpu(cpu) {
1056                 if (cpu == hcpu) {
1057                         *per_cpu_ptr(vi->vq_index, cpu) = -1;
1058                 } else {
1059                         *per_cpu_ptr(vi->vq_index, cpu) =
1060                                 ++i % vi->curr_queue_pairs;
1061                 }
1062         }
1063 }
1064
1065 static void virtnet_set_affinity(struct virtnet_info *vi)
1066 {
1067         int i;
1068         int cpu;
1069
1070         /* In multiqueue mode, when the number of cpu is equal to the number of
1071          * queue pairs, we let the queue pairs to be private to one cpu by
1072          * setting the affinity hint to eliminate the contention.
1073          */
1074         if (vi->curr_queue_pairs == 1 ||
1075             vi->max_queue_pairs != num_online_cpus()) {
1076                 virtnet_clean_affinity(vi, -1);
1077                 return;
1078         }
1079
1080         i = 0;
1081         for_each_online_cpu(cpu) {
1082                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1083                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1084                 *per_cpu_ptr(vi->vq_index, cpu) = i;
1085                 i++;
1086         }
1087
1088         vi->affinity_hint_set = true;
1089 }
1090
1091 static int virtnet_cpu_callback(struct notifier_block *nfb,
1092                                 unsigned long action, void *hcpu)
1093 {
1094         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1095
1096         switch(action & ~CPU_TASKS_FROZEN) {
1097         case CPU_ONLINE:
1098         case CPU_DOWN_FAILED:
1099         case CPU_DEAD:
1100                 virtnet_set_affinity(vi);
1101                 break;
1102         case CPU_DOWN_PREPARE:
1103                 virtnet_clean_affinity(vi, (long)hcpu);
1104                 break;
1105         default:
1106                 break;
1107         }
1108         return NOTIFY_OK;
1109 }
1110
1111 static void virtnet_get_ringparam(struct net_device *dev,
1112                                 struct ethtool_ringparam *ring)
1113 {
1114         struct virtnet_info *vi = netdev_priv(dev);
1115
1116         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1117         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1118         ring->rx_pending = ring->rx_max_pending;
1119         ring->tx_pending = ring->tx_max_pending;
1120 }
1121
1122
1123 static void virtnet_get_drvinfo(struct net_device *dev,
1124                                 struct ethtool_drvinfo *info)
1125 {
1126         struct virtnet_info *vi = netdev_priv(dev);
1127         struct virtio_device *vdev = vi->vdev;
1128
1129         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1130         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1131         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1132
1133 }
1134
1135 /* TODO: Eliminate OOO packets during switching */
1136 static int virtnet_set_channels(struct net_device *dev,
1137                                 struct ethtool_channels *channels)
1138 {
1139         struct virtnet_info *vi = netdev_priv(dev);
1140         u16 queue_pairs = channels->combined_count;
1141         int err;
1142
1143         /* We don't support separate rx/tx channels.
1144          * We don't allow setting 'other' channels.
1145          */
1146         if (channels->rx_count || channels->tx_count || channels->other_count)
1147                 return -EINVAL;
1148
1149         if (queue_pairs > vi->max_queue_pairs)
1150                 return -EINVAL;
1151
1152         get_online_cpus();
1153         err = virtnet_set_queues(vi, queue_pairs);
1154         if (!err) {
1155                 netif_set_real_num_tx_queues(dev, queue_pairs);
1156                 netif_set_real_num_rx_queues(dev, queue_pairs);
1157
1158                 virtnet_set_affinity(vi);
1159         }
1160         put_online_cpus();
1161
1162         return err;
1163 }
1164
1165 static void virtnet_get_channels(struct net_device *dev,
1166                                  struct ethtool_channels *channels)
1167 {
1168         struct virtnet_info *vi = netdev_priv(dev);
1169
1170         channels->combined_count = vi->curr_queue_pairs;
1171         channels->max_combined = vi->max_queue_pairs;
1172         channels->max_other = 0;
1173         channels->rx_count = 0;
1174         channels->tx_count = 0;
1175         channels->other_count = 0;
1176 }
1177
1178 static const struct ethtool_ops virtnet_ethtool_ops = {
1179         .get_drvinfo = virtnet_get_drvinfo,
1180         .get_link = ethtool_op_get_link,
1181         .get_ringparam = virtnet_get_ringparam,
1182         .set_channels = virtnet_set_channels,
1183         .get_channels = virtnet_get_channels,
1184 };
1185
1186 #define MIN_MTU 68
1187 #define MAX_MTU 65535
1188
1189 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1190 {
1191         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1192                 return -EINVAL;
1193         dev->mtu = new_mtu;
1194         return 0;
1195 }
1196
1197 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
1198  * txq based on the processor id.
1199  */
1200 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
1201 {
1202         int txq;
1203         struct virtnet_info *vi = netdev_priv(dev);
1204
1205         if (skb_rx_queue_recorded(skb)) {
1206                 txq = skb_get_rx_queue(skb);
1207         } else {
1208                 txq = *__this_cpu_ptr(vi->vq_index);
1209                 if (txq == -1)
1210                         txq = 0;
1211         }
1212
1213         while (unlikely(txq >= dev->real_num_tx_queues))
1214                 txq -= dev->real_num_tx_queues;
1215
1216         return txq;
1217 }
1218
1219 static const struct net_device_ops virtnet_netdev = {
1220         .ndo_open            = virtnet_open,
1221         .ndo_stop            = virtnet_close,
1222         .ndo_start_xmit      = start_xmit,
1223         .ndo_validate_addr   = eth_validate_addr,
1224         .ndo_set_mac_address = virtnet_set_mac_address,
1225         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1226         .ndo_change_mtu      = virtnet_change_mtu,
1227         .ndo_get_stats64     = virtnet_stats,
1228         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1229         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1230         .ndo_select_queue     = virtnet_select_queue,
1231 #ifdef CONFIG_NET_POLL_CONTROLLER
1232         .ndo_poll_controller = virtnet_netpoll,
1233 #endif
1234 };
1235
1236 static void virtnet_config_changed_work(struct work_struct *work)
1237 {
1238         struct virtnet_info *vi =
1239                 container_of(work, struct virtnet_info, config_work);
1240         u16 v;
1241
1242         mutex_lock(&vi->config_lock);
1243         if (!vi->config_enable)
1244                 goto done;
1245
1246         if (virtio_config_val(vi->vdev, VIRTIO_NET_F_STATUS,
1247                               offsetof(struct virtio_net_config, status),
1248                               &v) < 0)
1249                 goto done;
1250
1251         if (v & VIRTIO_NET_S_ANNOUNCE) {
1252                 netdev_notify_peers(vi->dev);
1253                 virtnet_ack_link_announce(vi);
1254         }
1255
1256         /* Ignore unknown (future) status bits */
1257         v &= VIRTIO_NET_S_LINK_UP;
1258
1259         if (vi->status == v)
1260                 goto done;
1261
1262         vi->status = v;
1263
1264         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1265                 netif_carrier_on(vi->dev);
1266                 netif_tx_wake_all_queues(vi->dev);
1267         } else {
1268                 netif_carrier_off(vi->dev);
1269                 netif_tx_stop_all_queues(vi->dev);
1270         }
1271 done:
1272         mutex_unlock(&vi->config_lock);
1273 }
1274
1275 static void virtnet_config_changed(struct virtio_device *vdev)
1276 {
1277         struct virtnet_info *vi = vdev->priv;
1278
1279         schedule_work(&vi->config_work);
1280 }
1281
1282 static void virtnet_free_queues(struct virtnet_info *vi)
1283 {
1284         kfree(vi->rq);
1285         kfree(vi->sq);
1286 }
1287
1288 static void free_receive_bufs(struct virtnet_info *vi)
1289 {
1290         int i;
1291
1292         for (i = 0; i < vi->max_queue_pairs; i++) {
1293                 while (vi->rq[i].pages)
1294                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1295         }
1296 }
1297
1298 static void free_unused_bufs(struct virtnet_info *vi)
1299 {
1300         void *buf;
1301         int i;
1302
1303         for (i = 0; i < vi->max_queue_pairs; i++) {
1304                 struct virtqueue *vq = vi->sq[i].vq;
1305                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1306                         dev_kfree_skb(buf);
1307         }
1308
1309         for (i = 0; i < vi->max_queue_pairs; i++) {
1310                 struct virtqueue *vq = vi->rq[i].vq;
1311
1312                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1313                         if (vi->mergeable_rx_bufs || vi->big_packets)
1314                                 give_pages(&vi->rq[i], buf);
1315                         else
1316                                 dev_kfree_skb(buf);
1317                         --vi->rq[i].num;
1318                 }
1319                 BUG_ON(vi->rq[i].num != 0);
1320         }
1321 }
1322
1323 static void virtnet_del_vqs(struct virtnet_info *vi)
1324 {
1325         struct virtio_device *vdev = vi->vdev;
1326
1327         virtnet_clean_affinity(vi, -1);
1328
1329         vdev->config->del_vqs(vdev);
1330
1331         virtnet_free_queues(vi);
1332 }
1333
1334 static int virtnet_find_vqs(struct virtnet_info *vi)
1335 {
1336         vq_callback_t **callbacks;
1337         struct virtqueue **vqs;
1338         int ret = -ENOMEM;
1339         int i, total_vqs;
1340         const char **names;
1341
1342         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1343          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1344          * possible control vq.
1345          */
1346         total_vqs = vi->max_queue_pairs * 2 +
1347                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1348
1349         /* Allocate space for find_vqs parameters */
1350         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1351         if (!vqs)
1352                 goto err_vq;
1353         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1354         if (!callbacks)
1355                 goto err_callback;
1356         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1357         if (!names)
1358                 goto err_names;
1359
1360         /* Parameters for control virtqueue, if any */
1361         if (vi->has_cvq) {
1362                 callbacks[total_vqs - 1] = NULL;
1363                 names[total_vqs - 1] = "control";
1364         }
1365
1366         /* Allocate/initialize parameters for send/receive virtqueues */
1367         for (i = 0; i < vi->max_queue_pairs; i++) {
1368                 callbacks[rxq2vq(i)] = skb_recv_done;
1369                 callbacks[txq2vq(i)] = skb_xmit_done;
1370                 sprintf(vi->rq[i].name, "input.%d", i);
1371                 sprintf(vi->sq[i].name, "output.%d", i);
1372                 names[rxq2vq(i)] = vi->rq[i].name;
1373                 names[txq2vq(i)] = vi->sq[i].name;
1374         }
1375
1376         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1377                                          names);
1378         if (ret)
1379                 goto err_find;
1380
1381         if (vi->has_cvq) {
1382                 vi->cvq = vqs[total_vqs - 1];
1383                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1384                         vi->dev->features |= NETIF_F_HW_VLAN_FILTER;
1385         }
1386
1387         for (i = 0; i < vi->max_queue_pairs; i++) {
1388                 vi->rq[i].vq = vqs[rxq2vq(i)];
1389                 vi->sq[i].vq = vqs[txq2vq(i)];
1390         }
1391
1392         kfree(names);
1393         kfree(callbacks);
1394         kfree(vqs);
1395
1396         return 0;
1397
1398 err_find:
1399         kfree(names);
1400 err_names:
1401         kfree(callbacks);
1402 err_callback:
1403         kfree(vqs);
1404 err_vq:
1405         return ret;
1406 }
1407
1408 static int virtnet_alloc_queues(struct virtnet_info *vi)
1409 {
1410         int i;
1411
1412         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1413         if (!vi->sq)
1414                 goto err_sq;
1415         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1416         if (!vi->rq)
1417                 goto err_rq;
1418
1419         INIT_DELAYED_WORK(&vi->refill, refill_work);
1420         for (i = 0; i < vi->max_queue_pairs; i++) {
1421                 vi->rq[i].pages = NULL;
1422                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1423                                napi_weight);
1424
1425                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1426                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1427         }
1428
1429         return 0;
1430
1431 err_rq:
1432         kfree(vi->sq);
1433 err_sq:
1434         return -ENOMEM;
1435 }
1436
1437 static int init_vqs(struct virtnet_info *vi)
1438 {
1439         int ret;
1440
1441         /* Allocate send & receive queues */
1442         ret = virtnet_alloc_queues(vi);
1443         if (ret)
1444                 goto err;
1445
1446         ret = virtnet_find_vqs(vi);
1447         if (ret)
1448                 goto err_free;
1449
1450         get_online_cpus();
1451         virtnet_set_affinity(vi);
1452         put_online_cpus();
1453
1454         return 0;
1455
1456 err_free:
1457         virtnet_free_queues(vi);
1458 err:
1459         return ret;
1460 }
1461
1462 static int virtnet_probe(struct virtio_device *vdev)
1463 {
1464         int i, err;
1465         struct net_device *dev;
1466         struct virtnet_info *vi;
1467         u16 max_queue_pairs;
1468
1469         /* Find if host supports multiqueue virtio_net device */
1470         err = virtio_config_val(vdev, VIRTIO_NET_F_MQ,
1471                                 offsetof(struct virtio_net_config,
1472                                 max_virtqueue_pairs), &max_queue_pairs);
1473
1474         /* We need at least 2 queue's */
1475         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1476             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1477             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1478                 max_queue_pairs = 1;
1479
1480         /* Allocate ourselves a network device with room for our info */
1481         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1482         if (!dev)
1483                 return -ENOMEM;
1484
1485         /* Set up network device as normal. */
1486         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1487         dev->netdev_ops = &virtnet_netdev;
1488         dev->features = NETIF_F_HIGHDMA;
1489
1490         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1491         SET_NETDEV_DEV(dev, &vdev->dev);
1492
1493         /* Do we support "hardware" checksums? */
1494         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1495                 /* This opens up the world of extra features. */
1496                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1497                 if (csum)
1498                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1499
1500                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1501                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1502                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1503                 }
1504                 /* Individual feature bits: what can host handle? */
1505                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1506                         dev->hw_features |= NETIF_F_TSO;
1507                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1508                         dev->hw_features |= NETIF_F_TSO6;
1509                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1510                         dev->hw_features |= NETIF_F_TSO_ECN;
1511                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1512                         dev->hw_features |= NETIF_F_UFO;
1513
1514                 if (gso)
1515                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1516                 /* (!csum && gso) case will be fixed by register_netdev() */
1517         }
1518
1519         /* Configuration may specify what MAC to use.  Otherwise random. */
1520         if (virtio_config_val_len(vdev, VIRTIO_NET_F_MAC,
1521                                   offsetof(struct virtio_net_config, mac),
1522                                   dev->dev_addr, dev->addr_len) < 0)
1523                 eth_hw_addr_random(dev);
1524
1525         /* Set up our device-specific information */
1526         vi = netdev_priv(dev);
1527         vi->dev = dev;
1528         vi->vdev = vdev;
1529         vdev->priv = vi;
1530         vi->stats = alloc_percpu(struct virtnet_stats);
1531         err = -ENOMEM;
1532         if (vi->stats == NULL)
1533                 goto free;
1534
1535         vi->vq_index = alloc_percpu(int);
1536         if (vi->vq_index == NULL)
1537                 goto free_stats;
1538
1539         mutex_init(&vi->config_lock);
1540         vi->config_enable = true;
1541         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1542
1543         /* If we can receive ANY GSO packets, we must allocate large ones. */
1544         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1545             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1546             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1547                 vi->big_packets = true;
1548
1549         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1550                 vi->mergeable_rx_bufs = true;
1551
1552         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1553                 vi->has_cvq = true;
1554
1555         /* Use single tx/rx queue pair as default */
1556         vi->curr_queue_pairs = 1;
1557         vi->max_queue_pairs = max_queue_pairs;
1558
1559         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1560         err = init_vqs(vi);
1561         if (err)
1562                 goto free_index;
1563
1564         netif_set_real_num_tx_queues(dev, 1);
1565         netif_set_real_num_rx_queues(dev, 1);
1566
1567         err = register_netdev(dev);
1568         if (err) {
1569                 pr_debug("virtio_net: registering device failed\n");
1570                 goto free_vqs;
1571         }
1572
1573         /* Last of all, set up some receive buffers. */
1574         for (i = 0; i < vi->max_queue_pairs; i++) {
1575                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1576
1577                 /* If we didn't even get one input buffer, we're useless. */
1578                 if (vi->rq[i].num == 0) {
1579                         free_unused_bufs(vi);
1580                         err = -ENOMEM;
1581                         goto free_recv_bufs;
1582                 }
1583         }
1584
1585         vi->nb.notifier_call = &virtnet_cpu_callback;
1586         err = register_hotcpu_notifier(&vi->nb);
1587         if (err) {
1588                 pr_debug("virtio_net: registering cpu notifier failed\n");
1589                 goto free_recv_bufs;
1590         }
1591
1592         /* Assume link up if device can't report link status,
1593            otherwise get link status from config. */
1594         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1595                 netif_carrier_off(dev);
1596                 schedule_work(&vi->config_work);
1597         } else {
1598                 vi->status = VIRTIO_NET_S_LINK_UP;
1599                 netif_carrier_on(dev);
1600         }
1601
1602         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1603                  dev->name, max_queue_pairs);
1604
1605         return 0;
1606
1607 free_recv_bufs:
1608         free_receive_bufs(vi);
1609         unregister_netdev(dev);
1610 free_vqs:
1611         cancel_delayed_work_sync(&vi->refill);
1612         virtnet_del_vqs(vi);
1613 free_index:
1614         free_percpu(vi->vq_index);
1615 free_stats:
1616         free_percpu(vi->stats);
1617 free:
1618         free_netdev(dev);
1619         return err;
1620 }
1621
1622 static void remove_vq_common(struct virtnet_info *vi)
1623 {
1624         vi->vdev->config->reset(vi->vdev);
1625
1626         /* Free unused buffers in both send and recv, if any. */
1627         free_unused_bufs(vi);
1628
1629         free_receive_bufs(vi);
1630
1631         virtnet_del_vqs(vi);
1632 }
1633
1634 static void virtnet_remove(struct virtio_device *vdev)
1635 {
1636         struct virtnet_info *vi = vdev->priv;
1637
1638         unregister_hotcpu_notifier(&vi->nb);
1639
1640         /* Prevent config work handler from accessing the device. */
1641         mutex_lock(&vi->config_lock);
1642         vi->config_enable = false;
1643         mutex_unlock(&vi->config_lock);
1644
1645         unregister_netdev(vi->dev);
1646
1647         remove_vq_common(vi);
1648
1649         flush_work(&vi->config_work);
1650
1651         free_percpu(vi->vq_index);
1652         free_percpu(vi->stats);
1653         free_netdev(vi->dev);
1654 }
1655
1656 #ifdef CONFIG_PM
1657 static int virtnet_freeze(struct virtio_device *vdev)
1658 {
1659         struct virtnet_info *vi = vdev->priv;
1660         int i;
1661
1662         /* Prevent config work handler from accessing the device */
1663         mutex_lock(&vi->config_lock);
1664         vi->config_enable = false;
1665         mutex_unlock(&vi->config_lock);
1666
1667         netif_device_detach(vi->dev);
1668         cancel_delayed_work_sync(&vi->refill);
1669
1670         if (netif_running(vi->dev))
1671                 for (i = 0; i < vi->max_queue_pairs; i++) {
1672                         napi_disable(&vi->rq[i].napi);
1673                         netif_napi_del(&vi->rq[i].napi);
1674                 }
1675
1676         remove_vq_common(vi);
1677
1678         flush_work(&vi->config_work);
1679
1680         return 0;
1681 }
1682
1683 static int virtnet_restore(struct virtio_device *vdev)
1684 {
1685         struct virtnet_info *vi = vdev->priv;
1686         int err, i;
1687
1688         err = init_vqs(vi);
1689         if (err)
1690                 return err;
1691
1692         if (netif_running(vi->dev))
1693                 for (i = 0; i < vi->max_queue_pairs; i++)
1694                         virtnet_napi_enable(&vi->rq[i]);
1695
1696         netif_device_attach(vi->dev);
1697
1698         for (i = 0; i < vi->max_queue_pairs; i++)
1699                 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1700                         schedule_delayed_work(&vi->refill, 0);
1701
1702         mutex_lock(&vi->config_lock);
1703         vi->config_enable = true;
1704         mutex_unlock(&vi->config_lock);
1705
1706         virtnet_set_queues(vi, vi->curr_queue_pairs);
1707
1708         return 0;
1709 }
1710 #endif
1711
1712 static struct virtio_device_id id_table[] = {
1713         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1714         { 0 },
1715 };
1716
1717 static unsigned int features[] = {
1718         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1719         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1720         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1721         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1722         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1723         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1724         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1725         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1726         VIRTIO_NET_F_CTRL_MAC_ADDR,
1727 };
1728
1729 static struct virtio_driver virtio_net_driver = {
1730         .feature_table = features,
1731         .feature_table_size = ARRAY_SIZE(features),
1732         .driver.name =  KBUILD_MODNAME,
1733         .driver.owner = THIS_MODULE,
1734         .id_table =     id_table,
1735         .probe =        virtnet_probe,
1736         .remove =       virtnet_remove,
1737         .config_changed = virtnet_config_changed,
1738 #ifdef CONFIG_PM
1739         .freeze =       virtnet_freeze,
1740         .restore =      virtnet_restore,
1741 #endif
1742 };
1743
1744 static int __init init(void)
1745 {
1746         return register_virtio_driver(&virtio_net_driver);
1747 }
1748
1749 static void __exit fini(void)
1750 {
1751         unregister_virtio_driver(&virtio_net_driver);
1752 }
1753 module_init(init);
1754 module_exit(fini);
1755
1756 MODULE_DEVICE_TABLE(virtio, id_table);
1757 MODULE_DESCRIPTION("Virtio network driver");
1758 MODULE_LICENSE("GPL");