Pileus Git - ~andy/linux/blob - 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, see <http://www.gnu.org/licenses/>.
  17  */
  18 //#define DEBUG
  19 #include <linux/netdevice.h>
  20 #include <linux/etherdevice.h>
  21 #include <linux/ethtool.h>
  22 #include <linux/module.h>
  23 #include <linux/virtio.h>
  24 #include <linux/virtio_net.h>
  25 #include <linux/scatterlist.h>
  26 #include <linux/if_vlan.h>
  27 #include <linux/slab.h>
  28 #include <linux/cpu.h>
  29
  30 static int napi_weight = NAPI_POLL_WEIGHT;
  31 module_param(napi_weight, int, 0444);
  32
  33 static bool csum = true, gso = true;
  34 module_param(csum, bool, 0444);
  35 module_param(gso, bool, 0444);
  36
  37 /* FIXME: MTU in config. */
  38 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  39 #define MERGE_BUFFER_LEN (ALIGN(GOOD_PACKET_LEN + \
  40                                 sizeof(struct virtio_net_hdr_mrg_rxbuf), \
  41                                 L1_CACHE_BYTES))
  42 #define GOOD_COPY_LEN   128
  43
  44 #define VIRTNET_DRIVER_VERSION "1.0.0"
  45
  46 struct virtnet_stats {
  47         struct u64_stats_sync tx_syncp;
  48         struct u64_stats_sync rx_syncp;
  49         u64 tx_bytes;
  50         u64 tx_packets;
  51
  52         u64 rx_bytes;
  53         u64 rx_packets;
  54 };
  55
  56 /* Internal representation of a send virtqueue */
  57 struct send_queue {
  58         /* Virtqueue associated with this send _queue */
  59         struct virtqueue *vq;
  60
  61         /* TX: fragments + linear part + virtio header */
  62         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  63
  64         /* Name of the send queue: output.$index */
  65         char name[40];
  66 };
  67
  68 /* Internal representation of a receive virtqueue */
  69 struct receive_queue {
  70         /* Virtqueue associated with this receive_queue */
  71         struct virtqueue *vq;
  72
  73         struct napi_struct napi;
  74
  75         /* Chain pages by the private ptr. */
  76         struct page *pages;
  77
  78         /* RX: fragments + linear part + virtio header */
  79         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  80
  81         /* Name of this receive queue: input.$index */
  82         char name[40];
  83 };
  84
  85 struct virtnet_info {
  86         struct virtio_device *vdev;
  87         struct virtqueue *cvq;
  88         struct net_device *dev;
  89         struct send_queue *sq;
  90         struct receive_queue *rq;
  91         unsigned int status;
  92
  93         /* Max # of queue pairs supported by the device */
  94         u16 max_queue_pairs;
  95
  96         /* # of queue pairs currently used by the driver */
  97         u16 curr_queue_pairs;
  98
  99         /* I like... big packets and I cannot lie! */
 100         bool big_packets;
 101
 102         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 103         bool mergeable_rx_bufs;
 104
 105         /* Has control virtqueue */
 106         bool has_cvq;
 107
 108         /* Host can handle any s/g split between our header and packet data */
 109         bool any_header_sg;
 110
 111         /* enable config space updates */
 112         bool config_enable;
 113
 114         /* Active statistics */
 115         struct virtnet_stats __percpu *stats;
 116
 117         /* Work struct for refilling if we run low on memory. */
 118         struct delayed_work refill;
 119
 120         /* Work struct for config space updates */
 121         struct work_struct config_work;
 122
 123         /* Lock for config space updates */
 124         struct mutex config_lock;
 125
 126         /* Page_frag for GFP_KERNEL packet buffer allocation when we run
 127          * low on memory.
 128          */
 129         struct page_frag alloc_frag;
 130
 131         /* Does the affinity hint is set for virtqueues? */
 132         bool affinity_hint_set;
 133
 134         /* CPU hot plug notifier */
 135         struct notifier_block nb;
 136 };
 137
 138 struct skb_vnet_hdr {
 139         union {
 140                 struct virtio_net_hdr hdr;
 141                 struct virtio_net_hdr_mrg_rxbuf mhdr;
 142         };
 143 };
 144
 145 struct padded_vnet_hdr {
 146         struct virtio_net_hdr hdr;
 147         /*
 148          * virtio_net_hdr should be in a separated sg buffer because of a
 149          * QEMU bug, and data sg buffer shares same page with this header sg.
 150          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
 151          */
 152         char padding[6];
 153 };
 154
 155 /* Converting between virtqueue no. and kernel tx/rx queue no.
 156  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 157  */
 158 static int vq2txq(struct virtqueue *vq)
 159 {
 160         return (vq->index - 1) / 2;
 161 }
 162
 163 static int txq2vq(int txq)
 164 {
 165         return txq * 2 + 1;
 166 }
 167
 168 static int vq2rxq(struct virtqueue *vq)
 169 {
 170         return vq->index / 2;
 171 }
 172
 173 static int rxq2vq(int rxq)
 174 {
 175         return rxq * 2;
 176 }
 177
 178 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 179 {
 180         return (struct skb_vnet_hdr *)skb->cb;
 181 }
 182
 183 /*
 184  * private is used to chain pages for big packets, put the whole
 185  * most recent used list in the beginning for reuse
 186  */
 187 static void give_pages(struct receive_queue *rq, struct page *page)
 188 {
 189         struct page *end;
 190
 191         /* Find end of list, sew whole thing into vi->rq.pages. */
 192         for (end = page; end->private; end = (struct page *)end->private);
 193         end->private = (unsigned long)rq->pages;
 194         rq->pages = page;
 195 }
 196
 197 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 198 {
 199         struct page *p = rq->pages;
 200
 201         if (p) {
 202                 rq->pages = (struct page *)p->private;
 203                 /* clear private here, it is used to chain pages */
 204                 p->private = 0;
 205         } else
 206                 p = alloc_page(gfp_mask);
 207         return p;
 208 }
 209
 210 static void skb_xmit_done(struct virtqueue *vq)
 211 {
 212         struct virtnet_info *vi = vq->vdev->priv;
 213
 214         /* Suppress further interrupts. */
 215         virtqueue_disable_cb(vq);
 216
 217         /* We were probably waiting for more output buffers. */
 218         netif_wake_subqueue(vi->dev, vq2txq(vq));
 219 }
 220
 221 /* Called from bottom half context */
 222 static struct sk_buff *page_to_skb(struct receive_queue *rq,
 223                                    struct page *page, unsigned int offset,
 224                                    unsigned int len, unsigned int truesize)
 225 {
 226         struct virtnet_info *vi = rq->vq->vdev->priv;
 227         struct sk_buff *skb;
 228         struct skb_vnet_hdr *hdr;
 229         unsigned int copy, hdr_len, hdr_padded_len;
 230         char *p;
 231
 232         p = page_address(page) + offset;
 233
 234         /* copy small packet so we can reuse these pages for small data */
 235         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 236         if (unlikely(!skb))
 237                 return NULL;
 238
 239         hdr = skb_vnet_hdr(skb);
 240
 241         if (vi->mergeable_rx_bufs) {
 242                 hdr_len = sizeof hdr->mhdr;
 243                 hdr_padded_len = sizeof hdr->mhdr;
 244         } else {
 245                 hdr_len = sizeof hdr->hdr;
 246                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 247         }
 248
 249         memcpy(hdr, p, hdr_len);
 250
 251         len -= hdr_len;
 252         offset += hdr_padded_len;
 253         p += hdr_padded_len;
 254
 255         copy = len;
 256         if (copy > skb_tailroom(skb))
 257                 copy = skb_tailroom(skb);
 258         memcpy(skb_put(skb, copy), p, copy);
 259
 260         len -= copy;
 261         offset += copy;
 262
 263         if (vi->mergeable_rx_bufs) {
 264                 if (len)
 265                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 266                 else
 267                         put_page(page);
 268                 return skb;
 269         }
 270
 271         /*
 272          * Verify that we can indeed put this data into a skb.
 273          * This is here to handle cases when the device erroneously
 274          * tries to receive more than is possible. This is usually
 275          * the case of a broken device.
 276          */
 277         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 278                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 279                 dev_kfree_skb(skb);
 280                 return NULL;
 281         }
 282         BUG_ON(offset >= PAGE_SIZE);
 283         while (len) {
 284                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 285                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 286                                 frag_size, truesize);
 287                 len -= frag_size;
 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 struct sk_buff *receive_small(void *buf, unsigned int len)
 299 {
 300         struct sk_buff * skb = buf;
 301
 302         len -= sizeof(struct virtio_net_hdr);
 303         skb_trim(skb, len);
 304
 305         return skb;
 306 }
 307
 308 static struct sk_buff *receive_big(struct net_device *dev,
 309                                    struct receive_queue *rq,
 310                                    void *buf,
 311                                    unsigned int len)
 312 {
 313         struct page *page = buf;
 314         struct sk_buff *skb = page_to_skb(rq, page, 0, len, PAGE_SIZE);
 315
 316         if (unlikely(!skb))
 317                 goto err;
 318
 319         return skb;
 320
 321 err:
 322         dev->stats.rx_dropped++;
 323         give_pages(rq, page);
 324         return NULL;
 325 }
 326
 327 static struct sk_buff *receive_mergeable(struct net_device *dev,
 328                                          struct receive_queue *rq,
 329                                          void *buf,
 330                                          unsigned int len)
 331 {
 332         struct skb_vnet_hdr *hdr = buf;
 333         int num_buf = hdr->mhdr.num_buffers;
 334         struct page *page = virt_to_head_page(buf);
 335         int offset = buf - page_address(page);
 336         struct sk_buff *head_skb = page_to_skb(rq, page, offset, len,
 337                                                MERGE_BUFFER_LEN);
 338         struct sk_buff *curr_skb = head_skb;
 339
 340         if (unlikely(!curr_skb))
 341                 goto err_skb;
 342
 343         while (--num_buf) {
 344                 int num_skb_frags;
 345
 346                 buf = virtqueue_get_buf(rq->vq, &len);
 347                 if (unlikely(!buf)) {
 348                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 349                                  dev->name, num_buf, hdr->mhdr.num_buffers);
 350                         dev->stats.rx_length_errors++;
 351                         goto err_buf;
 352                 }
 353                 if (unlikely(len > MERGE_BUFFER_LEN)) {
 354                         pr_debug("%s: rx error: merge buffer too long\n",
 355                                  dev->name);
 356                         len = MERGE_BUFFER_LEN;
 357                 }
 358
 359                 page = virt_to_head_page(buf);
 360
 361                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 362                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 363                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 364
 365                         if (unlikely(!nskb))
 366                                 goto err_skb;
 367                         if (curr_skb == head_skb)
 368                                 skb_shinfo(curr_skb)->frag_list = nskb;
 369                         else
 370                                 curr_skb->next = nskb;
 371                         curr_skb = nskb;
 372                         head_skb->truesize += nskb->truesize;
 373                         num_skb_frags = 0;
 374                 }
 375                 if (curr_skb != head_skb) {
 376                         head_skb->data_len += len;
 377                         head_skb->len += len;
 378                         head_skb->truesize += MERGE_BUFFER_LEN;
 379                 }
 380                 offset = buf - page_address(page);
 381                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 382                         put_page(page);
 383                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 384                                              len, MERGE_BUFFER_LEN);
 385                 } else {
 386                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 387                                         offset, len, MERGE_BUFFER_LEN);
 388                 }
 389         }
 390
 391         return head_skb;
 392
 393 err_skb:
 394         put_page(page);
 395         while (--num_buf) {
 396                 buf = virtqueue_get_buf(rq->vq, &len);
 397                 if (unlikely(!buf)) {
 398                         pr_debug("%s: rx error: %d buffers missing\n",
 399                                  dev->name, num_buf);
 400                         dev->stats.rx_length_errors++;
 401                         break;
 402                 }
 403                 page = virt_to_head_page(buf);
 404                 put_page(page);
 405         }
 406 err_buf:
 407         dev->stats.rx_dropped++;
 408         dev_kfree_skb(head_skb);
 409         return NULL;
 410 }
 411
 412 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
 413 {
 414         struct virtnet_info *vi = rq->vq->vdev->priv;
 415         struct net_device *dev = vi->dev;
 416         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 417         struct sk_buff *skb;
 418         struct skb_vnet_hdr *hdr;
 419
 420         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 421                 pr_debug("%s: short packet %i\n", dev->name, len);
 422                 dev->stats.rx_length_errors++;
 423                 if (vi->mergeable_rx_bufs)
 424                         put_page(virt_to_head_page(buf));
 425                 else if (vi->big_packets)
 426                         give_pages(rq, buf);
 427                 else
 428                         dev_kfree_skb(buf);
 429                 return;
 430         }
 431
 432         if (vi->mergeable_rx_bufs)
 433                 skb = receive_mergeable(dev, rq, buf, len);
 434         else if (vi->big_packets)
 435                 skb = receive_big(dev, rq, buf, len);
 436         else
 437                 skb = receive_small(buf, len);
 438
 439         if (unlikely(!skb))
 440                 return;
 441
 442         hdr = skb_vnet_hdr(skb);
 443
 444         u64_stats_update_begin(&stats->rx_syncp);
 445         stats->rx_bytes += skb->len;
 446         stats->rx_packets++;
 447         u64_stats_update_end(&stats->rx_syncp);
 448
 449         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 450                 pr_debug("Needs csum!\n");
 451                 if (!skb_partial_csum_set(skb,
 452                                           hdr->hdr.csum_start,
 453                                           hdr->hdr.csum_offset))
 454                         goto frame_err;
 455         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 456                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 457         }
 458
 459         skb->protocol = eth_type_trans(skb, dev);
 460         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 461                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 462
 463         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 464                 pr_debug("GSO!\n");
 465                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 466                 case VIRTIO_NET_HDR_GSO_TCPV4:
 467                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 468                         break;
 469                 case VIRTIO_NET_HDR_GSO_UDP:
 470                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 471                         break;
 472                 case VIRTIO_NET_HDR_GSO_TCPV6:
 473                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 474                         break;
 475                 default:
 476                         net_warn_ratelimited("%s: bad gso type %u.\n",
 477                                              dev->name, hdr->hdr.gso_type);
 478                         goto frame_err;
 479                 }
 480
 481                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 482                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 483
 484                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 485                 if (skb_shinfo(skb)->gso_size == 0) {
 486                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 487                         goto frame_err;
 488                 }
 489
 490                 /* Header must be checked, and gso_segs computed. */
 491                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 492                 skb_shinfo(skb)->gso_segs = 0;
 493         }
 494
 495         netif_receive_skb(skb);
 496         return;
 497
 498 frame_err:
 499         dev->stats.rx_frame_errors++;
 500         dev_kfree_skb(skb);
 501 }
 502
 503 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
 504 {
 505         struct virtnet_info *vi = rq->vq->vdev->priv;
 506         struct sk_buff *skb;
 507         struct skb_vnet_hdr *hdr;
 508         int err;
 509
 510         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 511         if (unlikely(!skb))
 512                 return -ENOMEM;
 513
 514         skb_put(skb, GOOD_PACKET_LEN);
 515
 516         hdr = skb_vnet_hdr(skb);
 517         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
 518
 519         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 520
 521         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 522         if (err < 0)
 523                 dev_kfree_skb(skb);
 524
 525         return err;
 526 }
 527
 528 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
 529 {
 530         struct page *first, *list = NULL;
 531         char *p;
 532         int i, err, offset;
 533
 534         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 535         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 536                 first = get_a_page(rq, gfp);
 537                 if (!first) {
 538                         if (list)
 539                                 give_pages(rq, list);
 540                         return -ENOMEM;
 541                 }
 542                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 543
 544                 /* chain new page in list head to match sg */
 545                 first->private = (unsigned long)list;
 546                 list = first;
 547         }
 548
 549         first = get_a_page(rq, gfp);
 550         if (!first) {
 551                 give_pages(rq, list);
 552                 return -ENOMEM;
 553         }
 554         p = page_address(first);
 555
 556         /* rq->sg[0], rq->sg[1] share the same page */
 557         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
 558         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
 559
 560         /* rq->sg[1] for data packet, from offset */
 561         offset = sizeof(struct padded_vnet_hdr);
 562         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 563
 564         /* chain first in list head */
 565         first->private = (unsigned long)list;
 566         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 567                                   first, gfp);
 568         if (err < 0)
 569                 give_pages(rq, first);
 570
 571         return err;
 572 }
 573
 574 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 575 {
 576         struct virtnet_info *vi = rq->vq->vdev->priv;
 577         char *buf = NULL;
 578         int err;
 579
 580         if (gfp & __GFP_WAIT) {
 581                 if (skb_page_frag_refill(MERGE_BUFFER_LEN, &vi->alloc_frag,
 582                                          gfp)) {
 583                         buf = (char *)page_address(vi->alloc_frag.page) +
 584                               vi->alloc_frag.offset;
 585                         get_page(vi->alloc_frag.page);
 586                         vi->alloc_frag.offset += MERGE_BUFFER_LEN;
 587                 }
 588         } else {
 589                 buf = netdev_alloc_frag(MERGE_BUFFER_LEN);
 590         }
 591         if (!buf)
 592                 return -ENOMEM;
 593
 594         sg_init_one(rq->sg, buf, MERGE_BUFFER_LEN);
 595         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
 596         if (err < 0)
 597                 put_page(virt_to_head_page(buf));
 598
 599         return err;
 600 }
 601
 602 /*
 603  * Returns false if we couldn't fill entirely (OOM).
 604  *
 605  * Normally run in the receive path, but can also be run from ndo_open
 606  * before we're receiving packets, or from refill_work which is
 607  * careful to disable receiving (using napi_disable).
 608  */
 609 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
 610 {
 611         struct virtnet_info *vi = rq->vq->vdev->priv;
 612         int err;
 613         bool oom;
 614
 615         do {
 616                 if (vi->mergeable_rx_bufs)
 617                         err = add_recvbuf_mergeable(rq, gfp);
 618                 else if (vi->big_packets)
 619                         err = add_recvbuf_big(rq, gfp);
 620                 else
 621                         err = add_recvbuf_small(rq, gfp);
 622
 623                 oom = err == -ENOMEM;
 624                 if (err)
 625                         break;
 626         } while (rq->vq->num_free);
 627         if (unlikely(!virtqueue_kick(rq->vq)))
 628                 return false;
 629         return !oom;
 630 }
 631
 632 static void skb_recv_done(struct virtqueue *rvq)
 633 {
 634         struct virtnet_info *vi = rvq->vdev->priv;
 635         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 636
 637         /* Schedule NAPI, Suppress further interrupts if successful. */
 638         if (napi_schedule_prep(&rq->napi)) {
 639                 virtqueue_disable_cb(rvq);
 640                 __napi_schedule(&rq->napi);
 641         }
 642 }
 643
 644 static void virtnet_napi_enable(struct receive_queue *rq)
 645 {
 646         napi_enable(&rq->napi);
 647
 648         /* If all buffers were filled by other side before we napi_enabled, we
 649          * won't get another interrupt, so process any outstanding packets
 650          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 651          * We synchronize against interrupts via NAPI_STATE_SCHED */
 652         if (napi_schedule_prep(&rq->napi)) {
 653                 virtqueue_disable_cb(rq->vq);
 654                 local_bh_disable();
 655                 __napi_schedule(&rq->napi);
 656                 local_bh_enable();
 657         }
 658 }
 659
 660 static void refill_work(struct work_struct *work)
 661 {
 662         struct virtnet_info *vi =
 663                 container_of(work, struct virtnet_info, refill.work);
 664         bool still_empty;
 665         int i;
 666
 667         for (i = 0; i < vi->curr_queue_pairs; i++) {
 668                 struct receive_queue *rq = &vi->rq[i];
 669
 670                 napi_disable(&rq->napi);
 671                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 672                 virtnet_napi_enable(rq);
 673
 674                 /* In theory, this can happen: if we don't get any buffers in
 675                  * we will *never* try to fill again.
 676                  */
 677                 if (still_empty)
 678                         schedule_delayed_work(&vi->refill, HZ/2);
 679         }
 680 }
 681
 682 static int virtnet_poll(struct napi_struct *napi, int budget)
 683 {
 684         struct receive_queue *rq =
 685                 container_of(napi, struct receive_queue, napi);
 686         struct virtnet_info *vi = rq->vq->vdev->priv;
 687         void *buf;
 688         unsigned int r, len, received = 0;
 689
 690 again:
 691         while (received < budget &&
 692                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 693                 receive_buf(rq, buf, len);
 694                 received++;
 695         }
 696
 697         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 698                 if (!try_fill_recv(rq, GFP_ATOMIC))
 699                         schedule_delayed_work(&vi->refill, 0);
 700         }
 701
 702         /* Out of packets? */
 703         if (received < budget) {
 704                 r = virtqueue_enable_cb_prepare(rq->vq);
 705                 napi_complete(napi);
 706                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 707                     napi_schedule_prep(napi)) {
 708                         virtqueue_disable_cb(rq->vq);
 709                         __napi_schedule(napi);
 710                         goto again;
 711                 }
 712         }
 713
 714         return received;
 715 }
 716
 717 static int virtnet_open(struct net_device *dev)
 718 {
 719         struct virtnet_info *vi = netdev_priv(dev);
 720         int i;
 721
 722         for (i = 0; i < vi->max_queue_pairs; i++) {
 723                 if (i < vi->curr_queue_pairs)
 724                         /* Make sure we have some buffers: if oom use wq. */
 725                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 726                                 schedule_delayed_work(&vi->refill, 0);
 727                 virtnet_napi_enable(&vi->rq[i]);
 728         }
 729
 730         return 0;
 731 }
 732
 733 static void free_old_xmit_skbs(struct send_queue *sq)
 734 {
 735         struct sk_buff *skb;
 736         unsigned int len;
 737         struct virtnet_info *vi = sq->vq->vdev->priv;
 738         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 739
 740         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 741                 pr_debug("Sent skb %p\n", skb);
 742
 743                 u64_stats_update_begin(&stats->tx_syncp);
 744                 stats->tx_bytes += skb->len;
 745                 stats->tx_packets++;
 746                 u64_stats_update_end(&stats->tx_syncp);
 747
 748                 dev_kfree_skb_any(skb);
 749         }
 750 }
 751
 752 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 753 {
 754         struct skb_vnet_hdr *hdr;
 755         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 756         struct virtnet_info *vi = sq->vq->vdev->priv;
 757         unsigned num_sg;
 758         unsigned hdr_len;
 759         bool can_push;
 760
 761         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 762         if (vi->mergeable_rx_bufs)
 763                 hdr_len = sizeof hdr->mhdr;
 764         else
 765                 hdr_len = sizeof hdr->hdr;
 766
 767         can_push = vi->any_header_sg &&
 768                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 769                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 770         /* Even if we can, don't push here yet as this would skew
 771          * csum_start offset below. */
 772         if (can_push)
 773                 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
 774         else
 775                 hdr = skb_vnet_hdr(skb);
 776
 777         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 778                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 779                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 780                 hdr->hdr.csum_offset = skb->csum_offset;
 781         } else {
 782                 hdr->hdr.flags = 0;
 783                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 784         }
 785
 786         if (skb_is_gso(skb)) {
 787                 hdr->hdr.hdr_len = skb_headlen(skb);
 788                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 789                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 790                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 791                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 792                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 793                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 794                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 795                 else
 796                         BUG();
 797                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 798                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 799         } else {
 800                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 801                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 802         }
 803
 804         if (vi->mergeable_rx_bufs)
 805                 hdr->mhdr.num_buffers = 0;
 806
 807         if (can_push) {
 808                 __skb_push(skb, hdr_len);
 809                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 810                 /* Pull header back to avoid skew in tx bytes calculations. */
 811                 __skb_pull(skb, hdr_len);
 812         } else {
 813                 sg_set_buf(sq->sg, hdr, hdr_len);
 814                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 815         }
 816         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 817 }
 818
 819 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 820 {
 821         struct virtnet_info *vi = netdev_priv(dev);
 822         int qnum = skb_get_queue_mapping(skb);
 823         struct send_queue *sq = &vi->sq[qnum];
 824         int err;
 825
 826         /* Free up any pending old buffers before queueing new ones. */
 827         free_old_xmit_skbs(sq);
 828
 829         /* Try to transmit */
 830         err = xmit_skb(sq, skb);
 831
 832         /* This should not happen! */
 833         if (unlikely(err) || unlikely(!virtqueue_kick(sq->vq))) {
 834                 dev->stats.tx_fifo_errors++;
 835                 if (net_ratelimit())
 836                         dev_warn(&dev->dev,
 837                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 838                 dev->stats.tx_dropped++;
 839                 kfree_skb(skb);
 840                 return NETDEV_TX_OK;
 841         }
 842
 843         /* Don't wait up for transmitted skbs to be freed. */
 844         skb_orphan(skb);
 845         nf_reset(skb);
 846
 847         /* Apparently nice girls don't return TX_BUSY; stop the queue
 848          * before it gets out of hand.  Naturally, this wastes entries. */
 849         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 850                 netif_stop_subqueue(dev, qnum);
 851                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 852                         /* More just got used, free them then recheck. */
 853                         free_old_xmit_skbs(sq);
 854                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 855                                 netif_start_subqueue(dev, qnum);
 856                                 virtqueue_disable_cb(sq->vq);
 857                         }
 858                 }
 859         }
 860
 861         return NETDEV_TX_OK;
 862 }
 863
 864 /*
 865  * Send command via the control virtqueue and check status.  Commands
 866  * supported by the hypervisor, as indicated by feature bits, should
 867  * never fail unless improperly formatted.
 868  */
 869 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 870                                  struct scatterlist *out)
 871 {
 872         struct scatterlist *sgs[4], hdr, stat;
 873         struct virtio_net_ctrl_hdr ctrl;
 874         virtio_net_ctrl_ack status = ~0;
 875         unsigned out_num = 0, tmp;
 876
 877         /* Caller should know better */
 878         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 879
 880         ctrl.class = class;
 881         ctrl.cmd = cmd;
 882         /* Add header */
 883         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 884         sgs[out_num++] = &hdr;
 885
 886         if (out)
 887                 sgs[out_num++] = out;
 888
 889         /* Add return status. */
 890         sg_init_one(&stat, &status, sizeof(status));
 891         sgs[out_num] = &stat;
 892
 893         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
 894         BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC) < 0);
 895
 896         if (unlikely(!virtqueue_kick(vi->cvq)))
 897                 return status == VIRTIO_NET_OK;
 898
 899         /* Spin for a response, the kick causes an ioport write, trapping
 900          * into the hypervisor, so the request should be handled immediately.
 901          */
 902         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
 903                !virtqueue_is_broken(vi->cvq))
 904                 cpu_relax();
 905
 906         return status == VIRTIO_NET_OK;
 907 }
 908
 909 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 910 {
 911         struct virtnet_info *vi = netdev_priv(dev);
 912         struct virtio_device *vdev = vi->vdev;
 913         int ret;
 914         struct sockaddr *addr = p;
 915         struct scatterlist sg;
 916
 917         ret = eth_prepare_mac_addr_change(dev, p);
 918         if (ret)
 919                 return ret;
 920
 921         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 922                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 923                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 924                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
 925                         dev_warn(&vdev->dev,
 926                                  "Failed to set mac address by vq command.\n");
 927                         return -EINVAL;
 928                 }
 929         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 930                 unsigned int i;
 931
 932                 /* Naturally, this has an atomicity problem. */
 933                 for (i = 0; i < dev->addr_len; i++)
 934                         virtio_cwrite8(vdev,
 935                                        offsetof(struct virtio_net_config, mac) +
 936                                        i, addr->sa_data[i]);
 937         }
 938
 939         eth_commit_mac_addr_change(dev, p);
 940
 941         return 0;
 942 }
 943
 944 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 945                                                struct rtnl_link_stats64 *tot)
 946 {
 947         struct virtnet_info *vi = netdev_priv(dev);
 948         int cpu;
 949         unsigned int start;
 950
 951         for_each_possible_cpu(cpu) {
 952                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 953                 u64 tpackets, tbytes, rpackets, rbytes;
 954
 955                 do {
 956                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 957                         tpackets = stats->tx_packets;
 958                         tbytes   = stats->tx_bytes;
 959                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 960
 961                 do {
 962                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 963                         rpackets = stats->rx_packets;
 964                         rbytes   = stats->rx_bytes;
 965                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 966
 967                 tot->rx_packets += rpackets;
 968                 tot->tx_packets += tpackets;
 969                 tot->rx_bytes   += rbytes;
 970                 tot->tx_bytes   += tbytes;
 971         }
 972
 973         tot->tx_dropped = dev->stats.tx_dropped;
 974         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 975         tot->rx_dropped = dev->stats.rx_dropped;
 976         tot->rx_length_errors = dev->stats.rx_length_errors;
 977         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 978
 979         return tot;
 980 }
 981
 982 #ifdef CONFIG_NET_POLL_CONTROLLER
 983 static void virtnet_netpoll(struct net_device *dev)
 984 {
 985         struct virtnet_info *vi = netdev_priv(dev);
 986         int i;
 987
 988         for (i = 0; i < vi->curr_queue_pairs; i++)
 989                 napi_schedule(&vi->rq[i].napi);
 990 }
 991 #endif
 992
 993 static void virtnet_ack_link_announce(struct virtnet_info *vi)
 994 {
 995         rtnl_lock();
 996         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
 997                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
 998                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
 999         rtnl_unlock();
1000 }
1001
1002 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1003 {
1004         struct scatterlist sg;
1005         struct virtio_net_ctrl_mq s;
1006         struct net_device *dev = vi->dev;
1007
1008         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1009                 return 0;
1010
1011         s.virtqueue_pairs = queue_pairs;
1012         sg_init_one(&sg, &s, sizeof(s));
1013
1014         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1015                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1016                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1017                          queue_pairs);
1018                 return -EINVAL;
1019         } else {
1020                 vi->curr_queue_pairs = queue_pairs;
1021                 /* virtnet_open() will refill when device is going to up. */
1022                 if (dev->flags & IFF_UP)
1023                         schedule_delayed_work(&vi->refill, 0);
1024         }
1025
1026         return 0;
1027 }
1028
1029 static int virtnet_close(struct net_device *dev)
1030 {
1031         struct virtnet_info *vi = netdev_priv(dev);
1032         int i;
1033
1034         /* Make sure refill_work doesn't re-enable napi! */
1035         cancel_delayed_work_sync(&vi->refill);
1036
1037         for (i = 0; i < vi->max_queue_pairs; i++)
1038                 napi_disable(&vi->rq[i].napi);
1039
1040         return 0;
1041 }
1042
1043 static void virtnet_set_rx_mode(struct net_device *dev)
1044 {
1045         struct virtnet_info *vi = netdev_priv(dev);
1046         struct scatterlist sg[2];
1047         u8 promisc, allmulti;
1048         struct virtio_net_ctrl_mac *mac_data;
1049         struct netdev_hw_addr *ha;
1050         int uc_count;
1051         int mc_count;
1052         void *buf;
1053         int i;
1054
1055         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1056         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1057                 return;
1058
1059         promisc = ((dev->flags & IFF_PROMISC) != 0);
1060         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1061
1062         sg_init_one(sg, &promisc, sizeof(promisc));
1063
1064         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1065                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1066                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1067                          promisc ? "en" : "dis");
1068
1069         sg_init_one(sg, &allmulti, sizeof(allmulti));
1070
1071         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1072                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1073                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1074                          allmulti ? "en" : "dis");
1075
1076         uc_count = netdev_uc_count(dev);
1077         mc_count = netdev_mc_count(dev);
1078         /* MAC filter - use one buffer for both lists */
1079         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1080                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1081         mac_data = buf;
1082         if (!buf)
1083                 return;
1084
1085         sg_init_table(sg, 2);
1086
1087         /* Store the unicast list and count in the front of the buffer */
1088         mac_data->entries = uc_count;
1089         i = 0;
1090         netdev_for_each_uc_addr(ha, dev)
1091                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1092
1093         sg_set_buf(&sg[0], mac_data,
1094                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1095
1096         /* multicast list and count fill the end */
1097         mac_data = (void *)&mac_data->macs[uc_count][0];
1098
1099         mac_data->entries = mc_count;
1100         i = 0;
1101         netdev_for_each_mc_addr(ha, dev)
1102                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1103
1104         sg_set_buf(&sg[1], mac_data,
1105                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1106
1107         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1108                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1109                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1110
1111         kfree(buf);
1112 }
1113
1114 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1115                                    __be16 proto, u16 vid)
1116 {
1117         struct virtnet_info *vi = netdev_priv(dev);
1118         struct scatterlist sg;
1119
1120         sg_init_one(&sg, &vid, sizeof(vid));
1121
1122         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1123                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1124                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1125         return 0;
1126 }
1127
1128 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1129                                     __be16 proto, u16 vid)
1130 {
1131         struct virtnet_info *vi = netdev_priv(dev);
1132         struct scatterlist sg;
1133
1134         sg_init_one(&sg, &vid, sizeof(vid));
1135
1136         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1137                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1138                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1139         return 0;
1140 }
1141
1142 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1143 {
1144         int i;
1145
1146         if (vi->affinity_hint_set) {
1147                 for (i = 0; i < vi->max_queue_pairs; i++) {
1148                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1149                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1150                 }
1151
1152                 vi->affinity_hint_set = false;
1153         }
1154 }
1155
1156 static void virtnet_set_affinity(struct virtnet_info *vi)
1157 {
1158         int i;
1159         int cpu;
1160
1161         /* In multiqueue mode, when the number of cpu is equal to the number of
1162          * queue pairs, we let the queue pairs to be private to one cpu by
1163          * setting the affinity hint to eliminate the contention.
1164          */
1165         if (vi->curr_queue_pairs == 1 ||
1166             vi->max_queue_pairs != num_online_cpus()) {
1167                 virtnet_clean_affinity(vi, -1);
1168                 return;
1169         }
1170
1171         i = 0;
1172         for_each_online_cpu(cpu) {
1173                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1174                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1175                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1176                 i++;
1177         }
1178
1179         vi->affinity_hint_set = true;
1180 }
1181
1182 static int virtnet_cpu_callback(struct notifier_block *nfb,
1183                                 unsigned long action, void *hcpu)
1184 {
1185         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1186
1187         switch(action & ~CPU_TASKS_FROZEN) {
1188         case CPU_ONLINE:
1189         case CPU_DOWN_FAILED:
1190         case CPU_DEAD:
1191                 virtnet_set_affinity(vi);
1192                 break;
1193         case CPU_DOWN_PREPARE:
1194                 virtnet_clean_affinity(vi, (long)hcpu);
1195                 break;
1196         default:
1197                 break;
1198         }
1199
1200         return NOTIFY_OK;
1201 }
1202
1203 static void virtnet_get_ringparam(struct net_device *dev,
1204                                 struct ethtool_ringparam *ring)
1205 {
1206         struct virtnet_info *vi = netdev_priv(dev);
1207
1208         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1209         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1210         ring->rx_pending = ring->rx_max_pending;
1211         ring->tx_pending = ring->tx_max_pending;
1212 }
1213
1214
1215 static void virtnet_get_drvinfo(struct net_device *dev,
1216                                 struct ethtool_drvinfo *info)
1217 {
1218         struct virtnet_info *vi = netdev_priv(dev);
1219         struct virtio_device *vdev = vi->vdev;
1220
1221         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1222         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1223         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1224
1225 }
1226
1227 /* TODO: Eliminate OOO packets during switching */
1228 static int virtnet_set_channels(struct net_device *dev,
1229                                 struct ethtool_channels *channels)
1230 {
1231         struct virtnet_info *vi = netdev_priv(dev);
1232         u16 queue_pairs = channels->combined_count;
1233         int err;
1234
1235         /* We don't support separate rx/tx channels.
1236          * We don't allow setting 'other' channels.
1237          */
1238         if (channels->rx_count || channels->tx_count || channels->other_count)
1239                 return -EINVAL;
1240
1241         if (queue_pairs > vi->max_queue_pairs)
1242                 return -EINVAL;
1243
1244         get_online_cpus();
1245         err = virtnet_set_queues(vi, queue_pairs);
1246         if (!err) {
1247                 netif_set_real_num_tx_queues(dev, queue_pairs);
1248                 netif_set_real_num_rx_queues(dev, queue_pairs);
1249
1250                 virtnet_set_affinity(vi);
1251         }
1252         put_online_cpus();
1253
1254         return err;
1255 }
1256
1257 static void virtnet_get_channels(struct net_device *dev,
1258                                  struct ethtool_channels *channels)
1259 {
1260         struct virtnet_info *vi = netdev_priv(dev);
1261
1262         channels->combined_count = vi->curr_queue_pairs;
1263         channels->max_combined = vi->max_queue_pairs;
1264         channels->max_other = 0;
1265         channels->rx_count = 0;
1266         channels->tx_count = 0;
1267         channels->other_count = 0;
1268 }
1269
1270 static const struct ethtool_ops virtnet_ethtool_ops = {
1271         .get_drvinfo = virtnet_get_drvinfo,
1272         .get_link = ethtool_op_get_link,
1273         .get_ringparam = virtnet_get_ringparam,
1274         .set_channels = virtnet_set_channels,
1275         .get_channels = virtnet_get_channels,
1276 };
1277
1278 #define MIN_MTU 68
1279 #define MAX_MTU 65535
1280
1281 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1282 {
1283         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1284                 return -EINVAL;
1285         dev->mtu = new_mtu;
1286         return 0;
1287 }
1288
1289 static const struct net_device_ops virtnet_netdev = {
1290         .ndo_open            = virtnet_open,
1291         .ndo_stop            = virtnet_close,
1292         .ndo_start_xmit      = start_xmit,
1293         .ndo_validate_addr   = eth_validate_addr,
1294         .ndo_set_mac_address = virtnet_set_mac_address,
1295         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1296         .ndo_change_mtu      = virtnet_change_mtu,
1297         .ndo_get_stats64     = virtnet_stats,
1298         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1299         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1300 #ifdef CONFIG_NET_POLL_CONTROLLER
1301         .ndo_poll_controller = virtnet_netpoll,
1302 #endif
1303 };
1304
1305 static void virtnet_config_changed_work(struct work_struct *work)
1306 {
1307         struct virtnet_info *vi =
1308                 container_of(work, struct virtnet_info, config_work);
1309         u16 v;
1310
1311         mutex_lock(&vi->config_lock);
1312         if (!vi->config_enable)
1313                 goto done;
1314
1315         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1316                                  struct virtio_net_config, status, &v) < 0)
1317                 goto done;
1318
1319         if (v & VIRTIO_NET_S_ANNOUNCE) {
1320                 netdev_notify_peers(vi->dev);
1321                 virtnet_ack_link_announce(vi);
1322         }
1323
1324         /* Ignore unknown (future) status bits */
1325         v &= VIRTIO_NET_S_LINK_UP;
1326
1327         if (vi->status == v)
1328                 goto done;
1329
1330         vi->status = v;
1331
1332         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1333                 netif_carrier_on(vi->dev);
1334                 netif_tx_wake_all_queues(vi->dev);
1335         } else {
1336                 netif_carrier_off(vi->dev);
1337                 netif_tx_stop_all_queues(vi->dev);
1338         }
1339 done:
1340         mutex_unlock(&vi->config_lock);
1341 }
1342
1343 static void virtnet_config_changed(struct virtio_device *vdev)
1344 {
1345         struct virtnet_info *vi = vdev->priv;
1346
1347         schedule_work(&vi->config_work);
1348 }
1349
1350 static void virtnet_free_queues(struct virtnet_info *vi)
1351 {
1352         int i;
1353
1354         for (i = 0; i < vi->max_queue_pairs; i++)
1355                 netif_napi_del(&vi->rq[i].napi);
1356
1357         kfree(vi->rq);
1358         kfree(vi->sq);
1359 }
1360
1361 static void free_receive_bufs(struct virtnet_info *vi)
1362 {
1363         int i;
1364
1365         for (i = 0; i < vi->max_queue_pairs; i++) {
1366                 while (vi->rq[i].pages)
1367                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1368         }
1369 }
1370
1371 static void free_unused_bufs(struct virtnet_info *vi)
1372 {
1373         void *buf;
1374         int i;
1375
1376         for (i = 0; i < vi->max_queue_pairs; i++) {
1377                 struct virtqueue *vq = vi->sq[i].vq;
1378                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1379                         dev_kfree_skb(buf);
1380         }
1381
1382         for (i = 0; i < vi->max_queue_pairs; i++) {
1383                 struct virtqueue *vq = vi->rq[i].vq;
1384
1385                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1386                         if (vi->mergeable_rx_bufs)
1387                                 put_page(virt_to_head_page(buf));
1388                         else if (vi->big_packets)
1389                                 give_pages(&vi->rq[i], buf);
1390                         else
1391                                 dev_kfree_skb(buf);
1392                 }
1393         }
1394 }
1395
1396 static void virtnet_del_vqs(struct virtnet_info *vi)
1397 {
1398         struct virtio_device *vdev = vi->vdev;
1399
1400         virtnet_clean_affinity(vi, -1);
1401
1402         vdev->config->del_vqs(vdev);
1403
1404         virtnet_free_queues(vi);
1405 }
1406
1407 static int virtnet_find_vqs(struct virtnet_info *vi)
1408 {
1409         vq_callback_t **callbacks;
1410         struct virtqueue **vqs;
1411         int ret = -ENOMEM;
1412         int i, total_vqs;
1413         const char **names;
1414
1415         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1416          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1417          * possible control vq.
1418          */
1419         total_vqs = vi->max_queue_pairs * 2 +
1420                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1421
1422         /* Allocate space for find_vqs parameters */
1423         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1424         if (!vqs)
1425                 goto err_vq;
1426         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1427         if (!callbacks)
1428                 goto err_callback;
1429         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1430         if (!names)
1431                 goto err_names;
1432
1433         /* Parameters for control virtqueue, if any */
1434         if (vi->has_cvq) {
1435                 callbacks[total_vqs - 1] = NULL;
1436                 names[total_vqs - 1] = "control";
1437         }
1438
1439         /* Allocate/initialize parameters for send/receive virtqueues */
1440         for (i = 0; i < vi->max_queue_pairs; i++) {
1441                 callbacks[rxq2vq(i)] = skb_recv_done;
1442                 callbacks[txq2vq(i)] = skb_xmit_done;
1443                 sprintf(vi->rq[i].name, "input.%d", i);
1444                 sprintf(vi->sq[i].name, "output.%d", i);
1445                 names[rxq2vq(i)] = vi->rq[i].name;
1446                 names[txq2vq(i)] = vi->sq[i].name;
1447         }
1448
1449         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1450                                          names);
1451         if (ret)
1452                 goto err_find;
1453
1454         if (vi->has_cvq) {
1455                 vi->cvq = vqs[total_vqs - 1];
1456                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1457                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1458         }
1459
1460         for (i = 0; i < vi->max_queue_pairs; i++) {
1461                 vi->rq[i].vq = vqs[rxq2vq(i)];
1462                 vi->sq[i].vq = vqs[txq2vq(i)];
1463         }
1464
1465         kfree(names);
1466         kfree(callbacks);
1467         kfree(vqs);
1468
1469         return 0;
1470
1471 err_find:
1472         kfree(names);
1473 err_names:
1474         kfree(callbacks);
1475 err_callback:
1476         kfree(vqs);
1477 err_vq:
1478         return ret;
1479 }
1480
1481 static int virtnet_alloc_queues(struct virtnet_info *vi)
1482 {
1483         int i;
1484
1485         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1486         if (!vi->sq)
1487                 goto err_sq;
1488         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1489         if (!vi->rq)
1490                 goto err_rq;
1491
1492         INIT_DELAYED_WORK(&vi->refill, refill_work);
1493         for (i = 0; i < vi->max_queue_pairs; i++) {
1494                 vi->rq[i].pages = NULL;
1495                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1496                                napi_weight);
1497
1498                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1499                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1500         }
1501
1502         return 0;
1503
1504 err_rq:
1505         kfree(vi->sq);
1506 err_sq:
1507         return -ENOMEM;
1508 }
1509
1510 static int init_vqs(struct virtnet_info *vi)
1511 {
1512         int ret;
1513
1514         /* Allocate send & receive queues */
1515         ret = virtnet_alloc_queues(vi);
1516         if (ret)
1517                 goto err;
1518
1519         ret = virtnet_find_vqs(vi);
1520         if (ret)
1521                 goto err_free;
1522
1523         get_online_cpus();
1524         virtnet_set_affinity(vi);
1525         put_online_cpus();
1526
1527         return 0;
1528
1529 err_free:
1530         virtnet_free_queues(vi);
1531 err:
1532         return ret;
1533 }
1534
1535 static int virtnet_probe(struct virtio_device *vdev)
1536 {
1537         int i, err;
1538         struct net_device *dev;
1539         struct virtnet_info *vi;
1540         u16 max_queue_pairs;
1541
1542         /* Find if host supports multiqueue virtio_net device */
1543         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1544                                    struct virtio_net_config,
1545                                    max_virtqueue_pairs, &max_queue_pairs);
1546
1547         /* We need at least 2 queue's */
1548         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1549             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1550             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1551                 max_queue_pairs = 1;
1552
1553         /* Allocate ourselves a network device with room for our info */
1554         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1555         if (!dev)
1556                 return -ENOMEM;
1557
1558         /* Set up network device as normal. */
1559         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1560         dev->netdev_ops = &virtnet_netdev;
1561         dev->features = NETIF_F_HIGHDMA;
1562
1563         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1564         SET_NETDEV_DEV(dev, &vdev->dev);
1565
1566         /* Do we support "hardware" checksums? */
1567         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1568                 /* This opens up the world of extra features. */
1569                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1570                 if (csum)
1571                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1572
1573                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1574                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1575                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1576                 }
1577                 /* Individual feature bits: what can host handle? */
1578                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1579                         dev->hw_features |= NETIF_F_TSO;
1580                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1581                         dev->hw_features |= NETIF_F_TSO6;
1582                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1583                         dev->hw_features |= NETIF_F_TSO_ECN;
1584                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1585                         dev->hw_features |= NETIF_F_UFO;
1586
1587                 if (gso)
1588                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1589                 /* (!csum && gso) case will be fixed by register_netdev() */
1590         }
1591         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1592                 dev->features |= NETIF_F_RXCSUM;
1593
1594         dev->vlan_features = dev->features;
1595
1596         /* Configuration may specify what MAC to use.  Otherwise random. */
1597         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1598                 virtio_cread_bytes(vdev,
1599                                    offsetof(struct virtio_net_config, mac),
1600                                    dev->dev_addr, dev->addr_len);
1601         else
1602                 eth_hw_addr_random(dev);
1603
1604         /* Set up our device-specific information */
1605         vi = netdev_priv(dev);
1606         vi->dev = dev;
1607         vi->vdev = vdev;
1608         vdev->priv = vi;
1609         vi->stats = alloc_percpu(struct virtnet_stats);
1610         err = -ENOMEM;
1611         if (vi->stats == NULL)
1612                 goto free;
1613
1614         for_each_possible_cpu(i) {
1615                 struct virtnet_stats *virtnet_stats;
1616                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1617                 u64_stats_init(&virtnet_stats->tx_syncp);
1618                 u64_stats_init(&virtnet_stats->rx_syncp);
1619         }
1620
1621         mutex_init(&vi->config_lock);
1622         vi->config_enable = true;
1623         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1624
1625         /* If we can receive ANY GSO packets, we must allocate large ones. */
1626         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1627             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1628             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1629                 vi->big_packets = true;
1630
1631         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1632                 vi->mergeable_rx_bufs = true;
1633
1634         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1635                 vi->any_header_sg = true;
1636
1637         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1638                 vi->has_cvq = true;
1639
1640         /* Use single tx/rx queue pair as default */
1641         vi->curr_queue_pairs = 1;
1642         vi->max_queue_pairs = max_queue_pairs;
1643
1644         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1645         err = init_vqs(vi);
1646         if (err)
1647                 goto free_stats;
1648
1649         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1650         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1651
1652         err = register_netdev(dev);
1653         if (err) {
1654                 pr_debug("virtio_net: registering device failed\n");
1655                 goto free_vqs;
1656         }
1657
1658         /* Last of all, set up some receive buffers. */
1659         for (i = 0; i < vi->curr_queue_pairs; i++) {
1660                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1661
1662                 /* If we didn't even get one input buffer, we're useless. */
1663                 if (vi->rq[i].vq->num_free ==
1664                     virtqueue_get_vring_size(vi->rq[i].vq)) {
1665                         free_unused_bufs(vi);
1666                         err = -ENOMEM;
1667                         goto free_recv_bufs;
1668                 }
1669         }
1670
1671         vi->nb.notifier_call = &virtnet_cpu_callback;
1672         err = register_hotcpu_notifier(&vi->nb);
1673         if (err) {
1674                 pr_debug("virtio_net: registering cpu notifier failed\n");
1675                 goto free_recv_bufs;
1676         }
1677
1678         /* Assume link up if device can't report link status,
1679            otherwise get link status from config. */
1680         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1681                 netif_carrier_off(dev);
1682                 schedule_work(&vi->config_work);
1683         } else {
1684                 vi->status = VIRTIO_NET_S_LINK_UP;
1685                 netif_carrier_on(dev);
1686         }
1687
1688         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1689                  dev->name, max_queue_pairs);
1690
1691         return 0;
1692
1693 free_recv_bufs:
1694         free_receive_bufs(vi);
1695         unregister_netdev(dev);
1696 free_vqs:
1697         cancel_delayed_work_sync(&vi->refill);
1698         virtnet_del_vqs(vi);
1699         if (vi->alloc_frag.page)
1700                 put_page(vi->alloc_frag.page);
1701 free_stats:
1702         free_percpu(vi->stats);
1703 free:
1704         free_netdev(dev);
1705         return err;
1706 }
1707
1708 static void remove_vq_common(struct virtnet_info *vi)
1709 {
1710         vi->vdev->config->reset(vi->vdev);
1711
1712         /* Free unused buffers in both send and recv, if any. */
1713         free_unused_bufs(vi);
1714
1715         free_receive_bufs(vi);
1716
1717         virtnet_del_vqs(vi);
1718 }
1719
1720 static void virtnet_remove(struct virtio_device *vdev)
1721 {
1722         struct virtnet_info *vi = vdev->priv;
1723
1724         unregister_hotcpu_notifier(&vi->nb);
1725
1726         /* Prevent config work handler from accessing the device. */
1727         mutex_lock(&vi->config_lock);
1728         vi->config_enable = false;
1729         mutex_unlock(&vi->config_lock);
1730
1731         unregister_netdev(vi->dev);
1732
1733         remove_vq_common(vi);
1734         if (vi->alloc_frag.page)
1735                 put_page(vi->alloc_frag.page);
1736
1737         flush_work(&vi->config_work);
1738
1739         free_percpu(vi->stats);
1740         free_netdev(vi->dev);
1741 }
1742
1743 #ifdef CONFIG_PM_SLEEP
1744 static int virtnet_freeze(struct virtio_device *vdev)
1745 {
1746         struct virtnet_info *vi = vdev->priv;
1747         int i;
1748
1749         unregister_hotcpu_notifier(&vi->nb);
1750
1751         /* Prevent config work handler from accessing the device */
1752         mutex_lock(&vi->config_lock);
1753         vi->config_enable = false;
1754         mutex_unlock(&vi->config_lock);
1755
1756         netif_device_detach(vi->dev);
1757         cancel_delayed_work_sync(&vi->refill);
1758
1759         if (netif_running(vi->dev))
1760                 for (i = 0; i < vi->max_queue_pairs; i++) {
1761                         napi_disable(&vi->rq[i].napi);
1762                         netif_napi_del(&vi->rq[i].napi);
1763                 }
1764
1765         remove_vq_common(vi);
1766
1767         flush_work(&vi->config_work);
1768
1769         return 0;
1770 }
1771
1772 static int virtnet_restore(struct virtio_device *vdev)
1773 {
1774         struct virtnet_info *vi = vdev->priv;
1775         int err, i;
1776
1777         err = init_vqs(vi);
1778         if (err)
1779                 return err;
1780
1781         if (netif_running(vi->dev)) {
1782                 for (i = 0; i < vi->curr_queue_pairs; i++)
1783                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1784                                 schedule_delayed_work(&vi->refill, 0);
1785
1786                 for (i = 0; i < vi->max_queue_pairs; i++)
1787                         virtnet_napi_enable(&vi->rq[i]);
1788         }
1789
1790         netif_device_attach(vi->dev);
1791
1792         mutex_lock(&vi->config_lock);
1793         vi->config_enable = true;
1794         mutex_unlock(&vi->config_lock);
1795
1796         rtnl_lock();
1797         virtnet_set_queues(vi, vi->curr_queue_pairs);
1798         rtnl_unlock();
1799
1800         err = register_hotcpu_notifier(&vi->nb);
1801         if (err)
1802                 return err;
1803
1804         return 0;
1805 }
1806 #endif
1807
1808 static struct virtio_device_id id_table[] = {
1809         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1810         { 0 },
1811 };
1812
1813 static unsigned int features[] = {
1814         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1815         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1816         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1817         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1818         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1819         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1820         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1821         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1822         VIRTIO_NET_F_CTRL_MAC_ADDR,
1823         VIRTIO_F_ANY_LAYOUT,
1824 };
1825
1826 static struct virtio_driver virtio_net_driver = {
1827         .feature_table = features,
1828         .feature_table_size = ARRAY_SIZE(features),
1829         .driver.name =  KBUILD_MODNAME,
1830         .driver.owner = THIS_MODULE,
1831         .id_table =     id_table,
1832         .probe =        virtnet_probe,
1833         .remove =       virtnet_remove,
1834         .config_changed = virtnet_config_changed,
1835 #ifdef CONFIG_PM_SLEEP
1836         .freeze =       virtnet_freeze,
1837         .restore =      virtnet_restore,
1838 #endif
1839 };
1840
1841 module_virtio_driver(virtio_net_driver);
1842
1843 MODULE_DEVICE_TABLE(virtio, id_table);
1844 MODULE_DESCRIPTION("Virtio network driver");
1845 MODULE_LICENSE("GPL");