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.$index */
  65         char name[40];
  66 };
  67
  68 /* Internal representation of a receive virtqueue */
  69 struct receive_queue {
  70         /* Virtqueue associated with this receive_queue */
  71         struct virtqueue *vq;
  72
  73         struct napi_struct napi;
  74
  75         /* Number of input buffers, and max we've ever had. */
  76         unsigned int num, max;
  77
  78         /* Chain pages by the private ptr. */
  79         struct page *pages;
  80
  81         /* RX: fragments + linear part + virtio header */
  82         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  83
  84         /* Name of this receive queue: input.$index */
  85         char name[40];
  86 };
  87
  88 struct virtnet_info {
  89         struct virtio_device *vdev;
  90         struct virtqueue *cvq;
  91         struct net_device *dev;
  92         struct send_queue *sq;
  93         struct receive_queue *rq;
  94         unsigned int status;
  95
  96         /* Max # of queue pairs supported by the device */
  97         u16 max_queue_pairs;
  98
  99         /* # of queue pairs currently used by the driver */
 100         u16 curr_queue_pairs;
 101
 102         /* I like... big packets and I cannot lie! */
 103         bool big_packets;
 104
 105         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 106         bool mergeable_rx_bufs;
 107
 108         /* Has control virtqueue */
 109         bool has_cvq;
 110
 111         /* Host can handle any s/g split between our header and packet data */
 112         bool any_header_sg;
 113
 114         /* enable config space updates */
 115         bool config_enable;
 116
 117         /* Active statistics */
 118         struct virtnet_stats __percpu *stats;
 119
 120         /* Work struct for refilling if we run low on memory. */
 121         struct delayed_work refill;
 122
 123         /* Work struct for config space updates */
 124         struct work_struct config_work;
 125
 126         /* Lock for config space updates */
 127         struct mutex config_lock;
 128
 129         /* Page_frag for GFP_KERNEL packet buffer allocation when we run
 130          * low on memory.
 131          */
 132         struct page_frag alloc_frag;
 133
 134         /* Does the affinity hint is set for virtqueues? */
 135         bool affinity_hint_set;
 136
 137         /* CPU hot plug notifier */
 138         struct notifier_block nb;
 139 };
 140
 141 struct skb_vnet_hdr {
 142         union {
 143                 struct virtio_net_hdr hdr;
 144                 struct virtio_net_hdr_mrg_rxbuf mhdr;
 145         };
 146 };
 147
 148 struct padded_vnet_hdr {
 149         struct virtio_net_hdr hdr;
 150         /*
 151          * virtio_net_hdr should be in a separated sg buffer because of a
 152          * QEMU bug, and data sg buffer shares same page with this header sg.
 153          * This padding makes next sg 16 byte aligned after virtio_net_hdr.
 154          */
 155         char padding[6];
 156 };
 157
 158 /* Converting between virtqueue no. and kernel tx/rx queue no.
 159  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 160  */
 161 static int vq2txq(struct virtqueue *vq)
 162 {
 163         return (vq->index - 1) / 2;
 164 }
 165
 166 static int txq2vq(int txq)
 167 {
 168         return txq * 2 + 1;
 169 }
 170
 171 static int vq2rxq(struct virtqueue *vq)
 172 {
 173         return vq->index / 2;
 174 }
 175
 176 static int rxq2vq(int rxq)
 177 {
 178         return rxq * 2;
 179 }
 180
 181 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
 182 {
 183         return (struct skb_vnet_hdr *)skb->cb;
 184 }
 185
 186 /*
 187  * private is used to chain pages for big packets, put the whole
 188  * most recent used list in the beginning for reuse
 189  */
 190 static void give_pages(struct receive_queue *rq, struct page *page)
 191 {
 192         struct page *end;
 193
 194         /* Find end of list, sew whole thing into vi->rq.pages. */
 195         for (end = page; end->private; end = (struct page *)end->private);
 196         end->private = (unsigned long)rq->pages;
 197         rq->pages = page;
 198 }
 199
 200 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 201 {
 202         struct page *p = rq->pages;
 203
 204         if (p) {
 205                 rq->pages = (struct page *)p->private;
 206                 /* clear private here, it is used to chain pages */
 207                 p->private = 0;
 208         } else
 209                 p = alloc_page(gfp_mask);
 210         return p;
 211 }
 212
 213 static void skb_xmit_done(struct virtqueue *vq)
 214 {
 215         struct virtnet_info *vi = vq->vdev->priv;
 216
 217         /* Suppress further interrupts. */
 218         virtqueue_disable_cb(vq);
 219
 220         /* We were probably waiting for more output buffers. */
 221         netif_wake_subqueue(vi->dev, vq2txq(vq));
 222 }
 223
 224 /* Called from bottom half context */
 225 static struct sk_buff *page_to_skb(struct receive_queue *rq,
 226                                    struct page *page, unsigned int offset,
 227                                    unsigned int len, unsigned int truesize)
 228 {
 229         struct virtnet_info *vi = rq->vq->vdev->priv;
 230         struct sk_buff *skb;
 231         struct skb_vnet_hdr *hdr;
 232         unsigned int copy, hdr_len, hdr_padded_len;
 233         char *p;
 234
 235         p = page_address(page) + offset;
 236
 237         /* copy small packet so we can reuse these pages for small data */
 238         skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
 239         if (unlikely(!skb))
 240                 return NULL;
 241
 242         hdr = skb_vnet_hdr(skb);
 243
 244         if (vi->mergeable_rx_bufs) {
 245                 hdr_len = sizeof hdr->mhdr;
 246                 hdr_padded_len = sizeof hdr->mhdr;
 247         } else {
 248                 hdr_len = sizeof hdr->hdr;
 249                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 250         }
 251
 252         memcpy(hdr, p, hdr_len);
 253
 254         len -= hdr_len;
 255         offset += hdr_padded_len;
 256         p += hdr_padded_len;
 257
 258         copy = len;
 259         if (copy > skb_tailroom(skb))
 260                 copy = skb_tailroom(skb);
 261         memcpy(skb_put(skb, copy), p, copy);
 262
 263         len -= copy;
 264         offset += copy;
 265
 266         if (vi->mergeable_rx_bufs) {
 267                 if (len)
 268                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 269                 else
 270                         put_page(page);
 271                 return skb;
 272         }
 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         BUG_ON(offset >= PAGE_SIZE);
 286         while (len) {
 287                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 288                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 289                                 frag_size, truesize);
 290                 len -= frag_size;
 291                 page = (struct page *)page->private;
 292                 offset = 0;
 293         }
 294
 295         if (page)
 296                 give_pages(rq, page);
 297
 298         return skb;
 299 }
 300
 301 static struct sk_buff *receive_small(void *buf, unsigned int len)
 302 {
 303         struct sk_buff * skb = buf;
 304
 305         len -= sizeof(struct virtio_net_hdr);
 306         skb_trim(skb, len);
 307
 308         return skb;
 309 }
 310
 311 static struct sk_buff *receive_big(struct net_device *dev,
 312                                    struct receive_queue *rq,
 313                                    void *buf,
 314                                    unsigned int len)
 315 {
 316         struct page *page = buf;
 317         struct sk_buff *skb = page_to_skb(rq, page, 0, len, PAGE_SIZE);
 318
 319         if (unlikely(!skb))
 320                 goto err;
 321
 322         return skb;
 323
 324 err:
 325         dev->stats.rx_dropped++;
 326         give_pages(rq, page);
 327         return NULL;
 328 }
 329
 330 static struct sk_buff *receive_mergeable(struct net_device *dev,
 331                                          struct receive_queue *rq,
 332                                          void *buf,
 333                                          unsigned int len)
 334 {
 335         struct skb_vnet_hdr *hdr = buf;
 336         int num_buf = hdr->mhdr.num_buffers;
 337         struct page *page = virt_to_head_page(buf);
 338         int offset = buf - page_address(page);
 339         struct sk_buff *head_skb = page_to_skb(rq, page, offset, len,
 340                                                MERGE_BUFFER_LEN);
 341         struct sk_buff *curr_skb = head_skb;
 342
 343         if (unlikely(!curr_skb))
 344                 goto err_skb;
 345
 346         while (--num_buf) {
 347                 int num_skb_frags;
 348
 349                 buf = virtqueue_get_buf(rq->vq, &len);
 350                 if (unlikely(!buf)) {
 351                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 352                                  dev->name, num_buf, hdr->mhdr.num_buffers);
 353                         dev->stats.rx_length_errors++;
 354                         goto err_buf;
 355                 }
 356                 if (unlikely(len > MERGE_BUFFER_LEN)) {
 357                         pr_debug("%s: rx error: merge buffer too long\n",
 358                                  dev->name);
 359                         len = MERGE_BUFFER_LEN;
 360                 }
 361
 362                 page = virt_to_head_page(buf);
 363                 --rq->num;
 364
 365                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 366                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 367                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 368
 369                         if (unlikely(!nskb))
 370                                 goto err_skb;
 371                         if (curr_skb == head_skb)
 372                                 skb_shinfo(curr_skb)->frag_list = nskb;
 373                         else
 374                                 curr_skb->next = nskb;
 375                         curr_skb = nskb;
 376                         head_skb->truesize += nskb->truesize;
 377                         num_skb_frags = 0;
 378                 }
 379                 if (curr_skb != head_skb) {
 380                         head_skb->data_len += len;
 381                         head_skb->len += len;
 382                         head_skb->truesize += MERGE_BUFFER_LEN;
 383                 }
 384                 offset = buf - page_address(page);
 385                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 386                         put_page(page);
 387                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 388                                              len, MERGE_BUFFER_LEN);
 389                 } else {
 390                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 391                                         offset, len, MERGE_BUFFER_LEN);
 392                 }
 393         }
 394
 395         return head_skb;
 396
 397 err_skb:
 398         put_page(page);
 399         while (--num_buf) {
 400                 buf = virtqueue_get_buf(rq->vq, &len);
 401                 if (unlikely(!buf)) {
 402                         pr_debug("%s: rx error: %d buffers missing\n",
 403                                  dev->name, num_buf);
 404                         dev->stats.rx_length_errors++;
 405                         break;
 406                 }
 407                 page = virt_to_head_page(buf);
 408                 put_page(page);
 409                 --rq->num;
 410         }
 411 err_buf:
 412         dev->stats.rx_dropped++;
 413         dev_kfree_skb(head_skb);
 414         return NULL;
 415 }
 416
 417 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
 418 {
 419         struct virtnet_info *vi = rq->vq->vdev->priv;
 420         struct net_device *dev = vi->dev;
 421         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 422         struct sk_buff *skb;
 423         struct skb_vnet_hdr *hdr;
 424
 425         if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
 426                 pr_debug("%s: short packet %i\n", dev->name, len);
 427                 dev->stats.rx_length_errors++;
 428                 if (vi->mergeable_rx_bufs)
 429                         put_page(virt_to_head_page(buf));
 430                 else if (vi->big_packets)
 431                         give_pages(rq, buf);
 432                 else
 433                         dev_kfree_skb(buf);
 434                 return;
 435         }
 436
 437         if (vi->mergeable_rx_bufs)
 438                 skb = receive_mergeable(dev, rq, buf, len);
 439         else if (vi->big_packets)
 440                 skb = receive_big(dev, rq, buf, len);
 441         else
 442                 skb = receive_small(buf, len);
 443
 444         if (unlikely(!skb))
 445                 return;
 446
 447         hdr = skb_vnet_hdr(skb);
 448
 449         u64_stats_update_begin(&stats->rx_syncp);
 450         stats->rx_bytes += skb->len;
 451         stats->rx_packets++;
 452         u64_stats_update_end(&stats->rx_syncp);
 453
 454         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 455                 pr_debug("Needs csum!\n");
 456                 if (!skb_partial_csum_set(skb,
 457                                           hdr->hdr.csum_start,
 458                                           hdr->hdr.csum_offset))
 459                         goto frame_err;
 460         } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
 461                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 462         }
 463
 464         skb->protocol = eth_type_trans(skb, dev);
 465         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 466                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 467
 468         if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 469                 pr_debug("GSO!\n");
 470                 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 471                 case VIRTIO_NET_HDR_GSO_TCPV4:
 472                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 473                         break;
 474                 case VIRTIO_NET_HDR_GSO_UDP:
 475                         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 476                         break;
 477                 case VIRTIO_NET_HDR_GSO_TCPV6:
 478                         skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 479                         break;
 480                 default:
 481                         net_warn_ratelimited("%s: bad gso type %u.\n",
 482                                              dev->name, hdr->hdr.gso_type);
 483                         goto frame_err;
 484                 }
 485
 486                 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
 487                         skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 488
 489                 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
 490                 if (skb_shinfo(skb)->gso_size == 0) {
 491                         net_warn_ratelimited("%s: zero gso size.\n", dev->name);
 492                         goto frame_err;
 493                 }
 494
 495                 /* Header must be checked, and gso_segs computed. */
 496                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
 497                 skb_shinfo(skb)->gso_segs = 0;
 498         }
 499
 500         netif_receive_skb(skb);
 501         return;
 502
 503 frame_err:
 504         dev->stats.rx_frame_errors++;
 505         dev_kfree_skb(skb);
 506 }
 507
 508 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
 509 {
 510         struct virtnet_info *vi = rq->vq->vdev->priv;
 511         struct sk_buff *skb;
 512         struct skb_vnet_hdr *hdr;
 513         int err;
 514
 515         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 516         if (unlikely(!skb))
 517                 return -ENOMEM;
 518
 519         skb_put(skb, GOOD_PACKET_LEN);
 520
 521         hdr = skb_vnet_hdr(skb);
 522         sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
 523
 524         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 525
 526         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 527         if (err < 0)
 528                 dev_kfree_skb(skb);
 529
 530         return err;
 531 }
 532
 533 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
 534 {
 535         struct page *first, *list = NULL;
 536         char *p;
 537         int i, err, offset;
 538
 539         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 540         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 541                 first = get_a_page(rq, gfp);
 542                 if (!first) {
 543                         if (list)
 544                                 give_pages(rq, list);
 545                         return -ENOMEM;
 546                 }
 547                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 548
 549                 /* chain new page in list head to match sg */
 550                 first->private = (unsigned long)list;
 551                 list = first;
 552         }
 553
 554         first = get_a_page(rq, gfp);
 555         if (!first) {
 556                 give_pages(rq, list);
 557                 return -ENOMEM;
 558         }
 559         p = page_address(first);
 560
 561         /* rq->sg[0], rq->sg[1] share the same page */
 562         /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
 563         sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
 564
 565         /* rq->sg[1] for data packet, from offset */
 566         offset = sizeof(struct padded_vnet_hdr);
 567         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 568
 569         /* chain first in list head */
 570         first->private = (unsigned long)list;
 571         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 572                                   first, gfp);
 573         if (err < 0)
 574                 give_pages(rq, first);
 575
 576         return err;
 577 }
 578
 579 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 580 {
 581         struct virtnet_info *vi = rq->vq->vdev->priv;
 582         char *buf = NULL;
 583         int err;
 584
 585         if (gfp & __GFP_WAIT) {
 586                 if (skb_page_frag_refill(MERGE_BUFFER_LEN, &vi->alloc_frag,
 587                                          gfp)) {
 588                         buf = (char *)page_address(vi->alloc_frag.page) +
 589                               vi->alloc_frag.offset;
 590                         get_page(vi->alloc_frag.page);
 591                         vi->alloc_frag.offset += MERGE_BUFFER_LEN;
 592                 }
 593         } else {
 594                 buf = netdev_alloc_frag(MERGE_BUFFER_LEN);
 595         }
 596         if (!buf)
 597                 return -ENOMEM;
 598
 599         sg_init_one(rq->sg, buf, MERGE_BUFFER_LEN);
 600         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
 601         if (err < 0)
 602                 put_page(virt_to_head_page(buf));
 603
 604         return err;
 605 }
 606
 607 /*
 608  * Returns false if we couldn't fill entirely (OOM).
 609  *
 610  * Normally run in the receive path, but can also be run from ndo_open
 611  * before we're receiving packets, or from refill_work which is
 612  * careful to disable receiving (using napi_disable).
 613  */
 614 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
 615 {
 616         struct virtnet_info *vi = rq->vq->vdev->priv;
 617         int err;
 618         bool oom;
 619
 620         do {
 621                 if (vi->mergeable_rx_bufs)
 622                         err = add_recvbuf_mergeable(rq, gfp);
 623                 else if (vi->big_packets)
 624                         err = add_recvbuf_big(rq, gfp);
 625                 else
 626                         err = add_recvbuf_small(rq, gfp);
 627
 628                 oom = err == -ENOMEM;
 629                 if (err)
 630                         break;
 631                 ++rq->num;
 632         } while (rq->vq->num_free);
 633         if (unlikely(rq->num > rq->max))
 634                 rq->max = rq->num;
 635         if (unlikely(!virtqueue_kick(rq->vq)))
 636                 return false;
 637         return !oom;
 638 }
 639
 640 static void skb_recv_done(struct virtqueue *rvq)
 641 {
 642         struct virtnet_info *vi = rvq->vdev->priv;
 643         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 644
 645         /* Schedule NAPI, Suppress further interrupts if successful. */
 646         if (napi_schedule_prep(&rq->napi)) {
 647                 virtqueue_disable_cb(rvq);
 648                 __napi_schedule(&rq->napi);
 649         }
 650 }
 651
 652 static void virtnet_napi_enable(struct receive_queue *rq)
 653 {
 654         napi_enable(&rq->napi);
 655
 656         /* If all buffers were filled by other side before we napi_enabled, we
 657          * won't get another interrupt, so process any outstanding packets
 658          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 659          * We synchronize against interrupts via NAPI_STATE_SCHED */
 660         if (napi_schedule_prep(&rq->napi)) {
 661                 virtqueue_disable_cb(rq->vq);
 662                 local_bh_disable();
 663                 __napi_schedule(&rq->napi);
 664                 local_bh_enable();
 665         }
 666 }
 667
 668 static void refill_work(struct work_struct *work)
 669 {
 670         struct virtnet_info *vi =
 671                 container_of(work, struct virtnet_info, refill.work);
 672         bool still_empty;
 673         int i;
 674
 675         for (i = 0; i < vi->curr_queue_pairs; i++) {
 676                 struct receive_queue *rq = &vi->rq[i];
 677
 678                 napi_disable(&rq->napi);
 679                 still_empty = !try_fill_recv(rq, GFP_KERNEL);
 680                 virtnet_napi_enable(rq);
 681
 682                 /* In theory, this can happen: if we don't get any buffers in
 683                  * we will *never* try to fill again.
 684                  */
 685                 if (still_empty)
 686                         schedule_delayed_work(&vi->refill, HZ/2);
 687         }
 688 }
 689
 690 static int virtnet_poll(struct napi_struct *napi, int budget)
 691 {
 692         struct receive_queue *rq =
 693                 container_of(napi, struct receive_queue, napi);
 694         struct virtnet_info *vi = rq->vq->vdev->priv;
 695         void *buf;
 696         unsigned int r, len, received = 0;
 697
 698 again:
 699         while (received < budget &&
 700                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 701                 receive_buf(rq, buf, len);
 702                 --rq->num;
 703                 received++;
 704         }
 705
 706         if (rq->num < rq->max / 2) {
 707                 if (!try_fill_recv(rq, GFP_ATOMIC))
 708                         schedule_delayed_work(&vi->refill, 0);
 709         }
 710
 711         /* Out of packets? */
 712         if (received < budget) {
 713                 r = virtqueue_enable_cb_prepare(rq->vq);
 714                 napi_complete(napi);
 715                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 716                     napi_schedule_prep(napi)) {
 717                         virtqueue_disable_cb(rq->vq);
 718                         __napi_schedule(napi);
 719                         goto again;
 720                 }
 721         }
 722
 723         return received;
 724 }
 725
 726 static int virtnet_open(struct net_device *dev)
 727 {
 728         struct virtnet_info *vi = netdev_priv(dev);
 729         int i;
 730
 731         for (i = 0; i < vi->max_queue_pairs; i++) {
 732                 if (i < vi->curr_queue_pairs)
 733                         /* Make sure we have some buffers: if oom use wq. */
 734                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
 735                                 schedule_delayed_work(&vi->refill, 0);
 736                 virtnet_napi_enable(&vi->rq[i]);
 737         }
 738
 739         return 0;
 740 }
 741
 742 static void free_old_xmit_skbs(struct send_queue *sq)
 743 {
 744         struct sk_buff *skb;
 745         unsigned int len;
 746         struct virtnet_info *vi = sq->vq->vdev->priv;
 747         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 748
 749         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 750                 pr_debug("Sent skb %p\n", skb);
 751
 752                 u64_stats_update_begin(&stats->tx_syncp);
 753                 stats->tx_bytes += skb->len;
 754                 stats->tx_packets++;
 755                 u64_stats_update_end(&stats->tx_syncp);
 756
 757                 dev_kfree_skb_any(skb);
 758         }
 759 }
 760
 761 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 762 {
 763         struct skb_vnet_hdr *hdr;
 764         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 765         struct virtnet_info *vi = sq->vq->vdev->priv;
 766         unsigned num_sg;
 767         unsigned hdr_len;
 768         bool can_push;
 769
 770         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 771         if (vi->mergeable_rx_bufs)
 772                 hdr_len = sizeof hdr->mhdr;
 773         else
 774                 hdr_len = sizeof hdr->hdr;
 775
 776         can_push = vi->any_header_sg &&
 777                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 778                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 779         /* Even if we can, don't push here yet as this would skew
 780          * csum_start offset below. */
 781         if (can_push)
 782                 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
 783         else
 784                 hdr = skb_vnet_hdr(skb);
 785
 786         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 787                 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 788                 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
 789                 hdr->hdr.csum_offset = skb->csum_offset;
 790         } else {
 791                 hdr->hdr.flags = 0;
 792                 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
 793         }
 794
 795         if (skb_is_gso(skb)) {
 796                 hdr->hdr.hdr_len = skb_headlen(skb);
 797                 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
 798                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 799                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 800                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
 801                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 802                 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
 803                         hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
 804                 else
 805                         BUG();
 806                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
 807                         hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
 808         } else {
 809                 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
 810                 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
 811         }
 812
 813         if (vi->mergeable_rx_bufs)
 814                 hdr->mhdr.num_buffers = 0;
 815
 816         if (can_push) {
 817                 __skb_push(skb, hdr_len);
 818                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 819                 /* Pull header back to avoid skew in tx bytes calculations. */
 820                 __skb_pull(skb, hdr_len);
 821         } else {
 822                 sg_set_buf(sq->sg, hdr, hdr_len);
 823                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 824         }
 825         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 826 }
 827
 828 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 829 {
 830         struct virtnet_info *vi = netdev_priv(dev);
 831         int qnum = skb_get_queue_mapping(skb);
 832         struct send_queue *sq = &vi->sq[qnum];
 833         int err;
 834
 835         /* Free up any pending old buffers before queueing new ones. */
 836         free_old_xmit_skbs(sq);
 837
 838         /* Try to transmit */
 839         err = xmit_skb(sq, skb);
 840
 841         /* This should not happen! */
 842         if (unlikely(err) || unlikely(!virtqueue_kick(sq->vq))) {
 843                 dev->stats.tx_fifo_errors++;
 844                 if (net_ratelimit())
 845                         dev_warn(&dev->dev,
 846                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 847                 dev->stats.tx_dropped++;
 848                 kfree_skb(skb);
 849                 return NETDEV_TX_OK;
 850         }
 851
 852         /* Don't wait up for transmitted skbs to be freed. */
 853         skb_orphan(skb);
 854         nf_reset(skb);
 855
 856         /* Apparently nice girls don't return TX_BUSY; stop the queue
 857          * before it gets out of hand.  Naturally, this wastes entries. */
 858         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 859                 netif_stop_subqueue(dev, qnum);
 860                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 861                         /* More just got used, free them then recheck. */
 862                         free_old_xmit_skbs(sq);
 863                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 864                                 netif_start_subqueue(dev, qnum);
 865                                 virtqueue_disable_cb(sq->vq);
 866                         }
 867                 }
 868         }
 869
 870         return NETDEV_TX_OK;
 871 }
 872
 873 /*
 874  * Send command via the control virtqueue and check status.  Commands
 875  * supported by the hypervisor, as indicated by feature bits, should
 876  * never fail unless improperly formatted.
 877  */
 878 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 879                                  struct scatterlist *out)
 880 {
 881         struct scatterlist *sgs[4], hdr, stat;
 882         struct virtio_net_ctrl_hdr ctrl;
 883         virtio_net_ctrl_ack status = ~0;
 884         unsigned out_num = 0, tmp;
 885
 886         /* Caller should know better */
 887         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 888
 889         ctrl.class = class;
 890         ctrl.cmd = cmd;
 891         /* Add header */
 892         sg_init_one(&hdr, &ctrl, sizeof(ctrl));
 893         sgs[out_num++] = &hdr;
 894
 895         if (out)
 896                 sgs[out_num++] = out;
 897
 898         /* Add return status. */
 899         sg_init_one(&stat, &status, sizeof(status));
 900         sgs[out_num] = &stat;
 901
 902         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
 903         BUG_ON(virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC) < 0);
 904
 905         if (unlikely(!virtqueue_kick(vi->cvq)))
 906                 return status == VIRTIO_NET_OK;
 907
 908         /* Spin for a response, the kick causes an ioport write, trapping
 909          * into the hypervisor, so the request should be handled immediately.
 910          */
 911         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
 912                !virtqueue_is_broken(vi->cvq))
 913                 cpu_relax();
 914
 915         return status == VIRTIO_NET_OK;
 916 }
 917
 918 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 919 {
 920         struct virtnet_info *vi = netdev_priv(dev);
 921         struct virtio_device *vdev = vi->vdev;
 922         int ret;
 923         struct sockaddr *addr = p;
 924         struct scatterlist sg;
 925
 926         ret = eth_prepare_mac_addr_change(dev, p);
 927         if (ret)
 928                 return ret;
 929
 930         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 931                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 932                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 933                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
 934                         dev_warn(&vdev->dev,
 935                                  "Failed to set mac address by vq command.\n");
 936                         return -EINVAL;
 937                 }
 938         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
 939                 unsigned int i;
 940
 941                 /* Naturally, this has an atomicity problem. */
 942                 for (i = 0; i < dev->addr_len; i++)
 943                         virtio_cwrite8(vdev,
 944                                        offsetof(struct virtio_net_config, mac) +
 945                                        i, addr->sa_data[i]);
 946         }
 947
 948         eth_commit_mac_addr_change(dev, p);
 949
 950         return 0;
 951 }
 952
 953 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
 954                                                struct rtnl_link_stats64 *tot)
 955 {
 956         struct virtnet_info *vi = netdev_priv(dev);
 957         int cpu;
 958         unsigned int start;
 959
 960         for_each_possible_cpu(cpu) {
 961                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
 962                 u64 tpackets, tbytes, rpackets, rbytes;
 963
 964                 do {
 965                         start = u64_stats_fetch_begin_bh(&stats->tx_syncp);
 966                         tpackets = stats->tx_packets;
 967                         tbytes   = stats->tx_bytes;
 968                 } while (u64_stats_fetch_retry_bh(&stats->tx_syncp, start));
 969
 970                 do {
 971                         start = u64_stats_fetch_begin_bh(&stats->rx_syncp);
 972                         rpackets = stats->rx_packets;
 973                         rbytes   = stats->rx_bytes;
 974                 } while (u64_stats_fetch_retry_bh(&stats->rx_syncp, start));
 975
 976                 tot->rx_packets += rpackets;
 977                 tot->tx_packets += tpackets;
 978                 tot->rx_bytes   += rbytes;
 979                 tot->tx_bytes   += tbytes;
 980         }
 981
 982         tot->tx_dropped = dev->stats.tx_dropped;
 983         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
 984         tot->rx_dropped = dev->stats.rx_dropped;
 985         tot->rx_length_errors = dev->stats.rx_length_errors;
 986         tot->rx_frame_errors = dev->stats.rx_frame_errors;
 987
 988         return tot;
 989 }
 990
 991 #ifdef CONFIG_NET_POLL_CONTROLLER
 992 static void virtnet_netpoll(struct net_device *dev)
 993 {
 994         struct virtnet_info *vi = netdev_priv(dev);
 995         int i;
 996
 997         for (i = 0; i < vi->curr_queue_pairs; i++)
 998                 napi_schedule(&vi->rq[i].napi);
 999 }
1000 #endif
1001
1002 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1003 {
1004         rtnl_lock();
1005         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1006                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1007                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1008         rtnl_unlock();
1009 }
1010
1011 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1012 {
1013         struct scatterlist sg;
1014         struct virtio_net_ctrl_mq s;
1015         struct net_device *dev = vi->dev;
1016
1017         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1018                 return 0;
1019
1020         s.virtqueue_pairs = queue_pairs;
1021         sg_init_one(&sg, &s, sizeof(s));
1022
1023         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1024                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1025                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1026                          queue_pairs);
1027                 return -EINVAL;
1028         } else {
1029                 vi->curr_queue_pairs = queue_pairs;
1030                 /* virtnet_open() will refill when device is going to up. */
1031                 if (dev->flags & IFF_UP)
1032                         schedule_delayed_work(&vi->refill, 0);
1033         }
1034
1035         return 0;
1036 }
1037
1038 static int virtnet_close(struct net_device *dev)
1039 {
1040         struct virtnet_info *vi = netdev_priv(dev);
1041         int i;
1042
1043         /* Make sure refill_work doesn't re-enable napi! */
1044         cancel_delayed_work_sync(&vi->refill);
1045
1046         for (i = 0; i < vi->max_queue_pairs; i++)
1047                 napi_disable(&vi->rq[i].napi);
1048
1049         return 0;
1050 }
1051
1052 static void virtnet_set_rx_mode(struct net_device *dev)
1053 {
1054         struct virtnet_info *vi = netdev_priv(dev);
1055         struct scatterlist sg[2];
1056         u8 promisc, allmulti;
1057         struct virtio_net_ctrl_mac *mac_data;
1058         struct netdev_hw_addr *ha;
1059         int uc_count;
1060         int mc_count;
1061         void *buf;
1062         int i;
1063
1064         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1065         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1066                 return;
1067
1068         promisc = ((dev->flags & IFF_PROMISC) != 0);
1069         allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1070
1071         sg_init_one(sg, &promisc, sizeof(promisc));
1072
1073         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1074                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1075                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1076                          promisc ? "en" : "dis");
1077
1078         sg_init_one(sg, &allmulti, sizeof(allmulti));
1079
1080         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1081                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1082                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1083                          allmulti ? "en" : "dis");
1084
1085         uc_count = netdev_uc_count(dev);
1086         mc_count = netdev_mc_count(dev);
1087         /* MAC filter - use one buffer for both lists */
1088         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1089                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1090         mac_data = buf;
1091         if (!buf)
1092                 return;
1093
1094         sg_init_table(sg, 2);
1095
1096         /* Store the unicast list and count in the front of the buffer */
1097         mac_data->entries = uc_count;
1098         i = 0;
1099         netdev_for_each_uc_addr(ha, dev)
1100                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1101
1102         sg_set_buf(&sg[0], mac_data,
1103                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1104
1105         /* multicast list and count fill the end */
1106         mac_data = (void *)&mac_data->macs[uc_count][0];
1107
1108         mac_data->entries = mc_count;
1109         i = 0;
1110         netdev_for_each_mc_addr(ha, dev)
1111                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1112
1113         sg_set_buf(&sg[1], mac_data,
1114                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1115
1116         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1117                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1118                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1119
1120         kfree(buf);
1121 }
1122
1123 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1124                                    __be16 proto, u16 vid)
1125 {
1126         struct virtnet_info *vi = netdev_priv(dev);
1127         struct scatterlist sg;
1128
1129         sg_init_one(&sg, &vid, sizeof(vid));
1130
1131         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1132                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1133                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1134         return 0;
1135 }
1136
1137 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1138                                     __be16 proto, u16 vid)
1139 {
1140         struct virtnet_info *vi = netdev_priv(dev);
1141         struct scatterlist sg;
1142
1143         sg_init_one(&sg, &vid, sizeof(vid));
1144
1145         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1146                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1147                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1148         return 0;
1149 }
1150
1151 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1152 {
1153         int i;
1154
1155         if (vi->affinity_hint_set) {
1156                 for (i = 0; i < vi->max_queue_pairs; i++) {
1157                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1158                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1159                 }
1160
1161                 vi->affinity_hint_set = false;
1162         }
1163 }
1164
1165 static void virtnet_set_affinity(struct virtnet_info *vi)
1166 {
1167         int i;
1168         int cpu;
1169
1170         /* In multiqueue mode, when the number of cpu is equal to the number of
1171          * queue pairs, we let the queue pairs to be private to one cpu by
1172          * setting the affinity hint to eliminate the contention.
1173          */
1174         if (vi->curr_queue_pairs == 1 ||
1175             vi->max_queue_pairs != num_online_cpus()) {
1176                 virtnet_clean_affinity(vi, -1);
1177                 return;
1178         }
1179
1180         i = 0;
1181         for_each_online_cpu(cpu) {
1182                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1183                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1184                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1185                 i++;
1186         }
1187
1188         vi->affinity_hint_set = true;
1189 }
1190
1191 static int virtnet_cpu_callback(struct notifier_block *nfb,
1192                                 unsigned long action, void *hcpu)
1193 {
1194         struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1195
1196         switch(action & ~CPU_TASKS_FROZEN) {
1197         case CPU_ONLINE:
1198         case CPU_DOWN_FAILED:
1199         case CPU_DEAD:
1200                 virtnet_set_affinity(vi);
1201                 break;
1202         case CPU_DOWN_PREPARE:
1203                 virtnet_clean_affinity(vi, (long)hcpu);
1204                 break;
1205         default:
1206                 break;
1207         }
1208
1209         return NOTIFY_OK;
1210 }
1211
1212 static void virtnet_get_ringparam(struct net_device *dev,
1213                                 struct ethtool_ringparam *ring)
1214 {
1215         struct virtnet_info *vi = netdev_priv(dev);
1216
1217         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1218         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1219         ring->rx_pending = ring->rx_max_pending;
1220         ring->tx_pending = ring->tx_max_pending;
1221 }
1222
1223
1224 static void virtnet_get_drvinfo(struct net_device *dev,
1225                                 struct ethtool_drvinfo *info)
1226 {
1227         struct virtnet_info *vi = netdev_priv(dev);
1228         struct virtio_device *vdev = vi->vdev;
1229
1230         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1231         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1232         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1233
1234 }
1235
1236 /* TODO: Eliminate OOO packets during switching */
1237 static int virtnet_set_channels(struct net_device *dev,
1238                                 struct ethtool_channels *channels)
1239 {
1240         struct virtnet_info *vi = netdev_priv(dev);
1241         u16 queue_pairs = channels->combined_count;
1242         int err;
1243
1244         /* We don't support separate rx/tx channels.
1245          * We don't allow setting 'other' channels.
1246          */
1247         if (channels->rx_count || channels->tx_count || channels->other_count)
1248                 return -EINVAL;
1249
1250         if (queue_pairs > vi->max_queue_pairs)
1251                 return -EINVAL;
1252
1253         get_online_cpus();
1254         err = virtnet_set_queues(vi, queue_pairs);
1255         if (!err) {
1256                 netif_set_real_num_tx_queues(dev, queue_pairs);
1257                 netif_set_real_num_rx_queues(dev, queue_pairs);
1258
1259                 virtnet_set_affinity(vi);
1260         }
1261         put_online_cpus();
1262
1263         return err;
1264 }
1265
1266 static void virtnet_get_channels(struct net_device *dev,
1267                                  struct ethtool_channels *channels)
1268 {
1269         struct virtnet_info *vi = netdev_priv(dev);
1270
1271         channels->combined_count = vi->curr_queue_pairs;
1272         channels->max_combined = vi->max_queue_pairs;
1273         channels->max_other = 0;
1274         channels->rx_count = 0;
1275         channels->tx_count = 0;
1276         channels->other_count = 0;
1277 }
1278
1279 static const struct ethtool_ops virtnet_ethtool_ops = {
1280         .get_drvinfo = virtnet_get_drvinfo,
1281         .get_link = ethtool_op_get_link,
1282         .get_ringparam = virtnet_get_ringparam,
1283         .set_channels = virtnet_set_channels,
1284         .get_channels = virtnet_get_channels,
1285 };
1286
1287 #define MIN_MTU 68
1288 #define MAX_MTU 65535
1289
1290 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1291 {
1292         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1293                 return -EINVAL;
1294         dev->mtu = new_mtu;
1295         return 0;
1296 }
1297
1298 static const struct net_device_ops virtnet_netdev = {
1299         .ndo_open            = virtnet_open,
1300         .ndo_stop            = virtnet_close,
1301         .ndo_start_xmit      = start_xmit,
1302         .ndo_validate_addr   = eth_validate_addr,
1303         .ndo_set_mac_address = virtnet_set_mac_address,
1304         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1305         .ndo_change_mtu      = virtnet_change_mtu,
1306         .ndo_get_stats64     = virtnet_stats,
1307         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1308         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1309 #ifdef CONFIG_NET_POLL_CONTROLLER
1310         .ndo_poll_controller = virtnet_netpoll,
1311 #endif
1312 };
1313
1314 static void virtnet_config_changed_work(struct work_struct *work)
1315 {
1316         struct virtnet_info *vi =
1317                 container_of(work, struct virtnet_info, config_work);
1318         u16 v;
1319
1320         mutex_lock(&vi->config_lock);
1321         if (!vi->config_enable)
1322                 goto done;
1323
1324         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1325                                  struct virtio_net_config, status, &v) < 0)
1326                 goto done;
1327
1328         if (v & VIRTIO_NET_S_ANNOUNCE) {
1329                 netdev_notify_peers(vi->dev);
1330                 virtnet_ack_link_announce(vi);
1331         }
1332
1333         /* Ignore unknown (future) status bits */
1334         v &= VIRTIO_NET_S_LINK_UP;
1335
1336         if (vi->status == v)
1337                 goto done;
1338
1339         vi->status = v;
1340
1341         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1342                 netif_carrier_on(vi->dev);
1343                 netif_tx_wake_all_queues(vi->dev);
1344         } else {
1345                 netif_carrier_off(vi->dev);
1346                 netif_tx_stop_all_queues(vi->dev);
1347         }
1348 done:
1349         mutex_unlock(&vi->config_lock);
1350 }
1351
1352 static void virtnet_config_changed(struct virtio_device *vdev)
1353 {
1354         struct virtnet_info *vi = vdev->priv;
1355
1356         schedule_work(&vi->config_work);
1357 }
1358
1359 static void virtnet_free_queues(struct virtnet_info *vi)
1360 {
1361         int i;
1362
1363         for (i = 0; i < vi->max_queue_pairs; i++)
1364                 netif_napi_del(&vi->rq[i].napi);
1365
1366         kfree(vi->rq);
1367         kfree(vi->sq);
1368 }
1369
1370 static void free_receive_bufs(struct virtnet_info *vi)
1371 {
1372         int i;
1373
1374         for (i = 0; i < vi->max_queue_pairs; i++) {
1375                 while (vi->rq[i].pages)
1376                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1377         }
1378 }
1379
1380 static void free_unused_bufs(struct virtnet_info *vi)
1381 {
1382         void *buf;
1383         int i;
1384
1385         for (i = 0; i < vi->max_queue_pairs; i++) {
1386                 struct virtqueue *vq = vi->sq[i].vq;
1387                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1388                         dev_kfree_skb(buf);
1389         }
1390
1391         for (i = 0; i < vi->max_queue_pairs; i++) {
1392                 struct virtqueue *vq = vi->rq[i].vq;
1393
1394                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1395                         if (vi->mergeable_rx_bufs)
1396                                 put_page(virt_to_head_page(buf));
1397                         else if (vi->big_packets)
1398                                 give_pages(&vi->rq[i], buf);
1399                         else
1400                                 dev_kfree_skb(buf);
1401                         --vi->rq[i].num;
1402                 }
1403                 BUG_ON(vi->rq[i].num != 0);
1404         }
1405 }
1406
1407 static void virtnet_del_vqs(struct virtnet_info *vi)
1408 {
1409         struct virtio_device *vdev = vi->vdev;
1410
1411         virtnet_clean_affinity(vi, -1);
1412
1413         vdev->config->del_vqs(vdev);
1414
1415         virtnet_free_queues(vi);
1416 }
1417
1418 static int virtnet_find_vqs(struct virtnet_info *vi)
1419 {
1420         vq_callback_t **callbacks;
1421         struct virtqueue **vqs;
1422         int ret = -ENOMEM;
1423         int i, total_vqs;
1424         const char **names;
1425
1426         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1427          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1428          * possible control vq.
1429          */
1430         total_vqs = vi->max_queue_pairs * 2 +
1431                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1432
1433         /* Allocate space for find_vqs parameters */
1434         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1435         if (!vqs)
1436                 goto err_vq;
1437         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1438         if (!callbacks)
1439                 goto err_callback;
1440         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1441         if (!names)
1442                 goto err_names;
1443
1444         /* Parameters for control virtqueue, if any */
1445         if (vi->has_cvq) {
1446                 callbacks[total_vqs - 1] = NULL;
1447                 names[total_vqs - 1] = "control";
1448         }
1449
1450         /* Allocate/initialize parameters for send/receive virtqueues */
1451         for (i = 0; i < vi->max_queue_pairs; i++) {
1452                 callbacks[rxq2vq(i)] = skb_recv_done;
1453                 callbacks[txq2vq(i)] = skb_xmit_done;
1454                 sprintf(vi->rq[i].name, "input.%d", i);
1455                 sprintf(vi->sq[i].name, "output.%d", i);
1456                 names[rxq2vq(i)] = vi->rq[i].name;
1457                 names[txq2vq(i)] = vi->sq[i].name;
1458         }
1459
1460         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1461                                          names);
1462         if (ret)
1463                 goto err_find;
1464
1465         if (vi->has_cvq) {
1466                 vi->cvq = vqs[total_vqs - 1];
1467                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1468                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1469         }
1470
1471         for (i = 0; i < vi->max_queue_pairs; i++) {
1472                 vi->rq[i].vq = vqs[rxq2vq(i)];
1473                 vi->sq[i].vq = vqs[txq2vq(i)];
1474         }
1475
1476         kfree(names);
1477         kfree(callbacks);
1478         kfree(vqs);
1479
1480         return 0;
1481
1482 err_find:
1483         kfree(names);
1484 err_names:
1485         kfree(callbacks);
1486 err_callback:
1487         kfree(vqs);
1488 err_vq:
1489         return ret;
1490 }
1491
1492 static int virtnet_alloc_queues(struct virtnet_info *vi)
1493 {
1494         int i;
1495
1496         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1497         if (!vi->sq)
1498                 goto err_sq;
1499         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1500         if (!vi->rq)
1501                 goto err_rq;
1502
1503         INIT_DELAYED_WORK(&vi->refill, refill_work);
1504         for (i = 0; i < vi->max_queue_pairs; i++) {
1505                 vi->rq[i].pages = NULL;
1506                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1507                                napi_weight);
1508
1509                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1510                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1511         }
1512
1513         return 0;
1514
1515 err_rq:
1516         kfree(vi->sq);
1517 err_sq:
1518         return -ENOMEM;
1519 }
1520
1521 static int init_vqs(struct virtnet_info *vi)
1522 {
1523         int ret;
1524
1525         /* Allocate send & receive queues */
1526         ret = virtnet_alloc_queues(vi);
1527         if (ret)
1528                 goto err;
1529
1530         ret = virtnet_find_vqs(vi);
1531         if (ret)
1532                 goto err_free;
1533
1534         get_online_cpus();
1535         virtnet_set_affinity(vi);
1536         put_online_cpus();
1537
1538         return 0;
1539
1540 err_free:
1541         virtnet_free_queues(vi);
1542 err:
1543         return ret;
1544 }
1545
1546 static int virtnet_probe(struct virtio_device *vdev)
1547 {
1548         int i, err;
1549         struct net_device *dev;
1550         struct virtnet_info *vi;
1551         u16 max_queue_pairs;
1552
1553         /* Find if host supports multiqueue virtio_net device */
1554         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1555                                    struct virtio_net_config,
1556                                    max_virtqueue_pairs, &max_queue_pairs);
1557
1558         /* We need at least 2 queue's */
1559         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1560             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1561             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1562                 max_queue_pairs = 1;
1563
1564         /* Allocate ourselves a network device with room for our info */
1565         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1566         if (!dev)
1567                 return -ENOMEM;
1568
1569         /* Set up network device as normal. */
1570         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1571         dev->netdev_ops = &virtnet_netdev;
1572         dev->features = NETIF_F_HIGHDMA;
1573
1574         SET_ETHTOOL_OPS(dev, &virtnet_ethtool_ops);
1575         SET_NETDEV_DEV(dev, &vdev->dev);
1576
1577         /* Do we support "hardware" checksums? */
1578         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1579                 /* This opens up the world of extra features. */
1580                 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1581                 if (csum)
1582                         dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1583
1584                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1585                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1586                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1587                 }
1588                 /* Individual feature bits: what can host handle? */
1589                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1590                         dev->hw_features |= NETIF_F_TSO;
1591                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1592                         dev->hw_features |= NETIF_F_TSO6;
1593                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1594                         dev->hw_features |= NETIF_F_TSO_ECN;
1595                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1596                         dev->hw_features |= NETIF_F_UFO;
1597
1598                 if (gso)
1599                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1600                 /* (!csum && gso) case will be fixed by register_netdev() */
1601         }
1602         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1603                 dev->features |= NETIF_F_RXCSUM;
1604
1605         dev->vlan_features = dev->features;
1606
1607         /* Configuration may specify what MAC to use.  Otherwise random. */
1608         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1609                 virtio_cread_bytes(vdev,
1610                                    offsetof(struct virtio_net_config, mac),
1611                                    dev->dev_addr, dev->addr_len);
1612         else
1613                 eth_hw_addr_random(dev);
1614
1615         /* Set up our device-specific information */
1616         vi = netdev_priv(dev);
1617         vi->dev = dev;
1618         vi->vdev = vdev;
1619         vdev->priv = vi;
1620         vi->stats = alloc_percpu(struct virtnet_stats);
1621         err = -ENOMEM;
1622         if (vi->stats == NULL)
1623                 goto free;
1624
1625         for_each_possible_cpu(i) {
1626                 struct virtnet_stats *virtnet_stats;
1627                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1628                 u64_stats_init(&virtnet_stats->tx_syncp);
1629                 u64_stats_init(&virtnet_stats->rx_syncp);
1630         }
1631
1632         mutex_init(&vi->config_lock);
1633         vi->config_enable = true;
1634         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1635
1636         /* If we can receive ANY GSO packets, we must allocate large ones. */
1637         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1638             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1639             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1640                 vi->big_packets = true;
1641
1642         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1643                 vi->mergeable_rx_bufs = true;
1644
1645         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1646                 vi->any_header_sg = true;
1647
1648         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1649                 vi->has_cvq = true;
1650
1651         /* Use single tx/rx queue pair as default */
1652         vi->curr_queue_pairs = 1;
1653         vi->max_queue_pairs = max_queue_pairs;
1654
1655         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1656         err = init_vqs(vi);
1657         if (err)
1658                 goto free_stats;
1659
1660         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1661         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1662
1663         err = register_netdev(dev);
1664         if (err) {
1665                 pr_debug("virtio_net: registering device failed\n");
1666                 goto free_vqs;
1667         }
1668
1669         /* Last of all, set up some receive buffers. */
1670         for (i = 0; i < vi->curr_queue_pairs; i++) {
1671                 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1672
1673                 /* If we didn't even get one input buffer, we're useless. */
1674                 if (vi->rq[i].num == 0) {
1675                         free_unused_bufs(vi);
1676                         err = -ENOMEM;
1677                         goto free_recv_bufs;
1678                 }
1679         }
1680
1681         vi->nb.notifier_call = &virtnet_cpu_callback;
1682         err = register_hotcpu_notifier(&vi->nb);
1683         if (err) {
1684                 pr_debug("virtio_net: registering cpu notifier failed\n");
1685                 goto free_recv_bufs;
1686         }
1687
1688         /* Assume link up if device can't report link status,
1689            otherwise get link status from config. */
1690         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1691                 netif_carrier_off(dev);
1692                 schedule_work(&vi->config_work);
1693         } else {
1694                 vi->status = VIRTIO_NET_S_LINK_UP;
1695                 netif_carrier_on(dev);
1696         }
1697
1698         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1699                  dev->name, max_queue_pairs);
1700
1701         return 0;
1702
1703 free_recv_bufs:
1704         free_receive_bufs(vi);
1705         unregister_netdev(dev);
1706 free_vqs:
1707         cancel_delayed_work_sync(&vi->refill);
1708         virtnet_del_vqs(vi);
1709         if (vi->alloc_frag.page)
1710                 put_page(vi->alloc_frag.page);
1711 free_stats:
1712         free_percpu(vi->stats);
1713 free:
1714         free_netdev(dev);
1715         return err;
1716 }
1717
1718 static void remove_vq_common(struct virtnet_info *vi)
1719 {
1720         vi->vdev->config->reset(vi->vdev);
1721
1722         /* Free unused buffers in both send and recv, if any. */
1723         free_unused_bufs(vi);
1724
1725         free_receive_bufs(vi);
1726
1727         virtnet_del_vqs(vi);
1728 }
1729
1730 static void virtnet_remove(struct virtio_device *vdev)
1731 {
1732         struct virtnet_info *vi = vdev->priv;
1733
1734         unregister_hotcpu_notifier(&vi->nb);
1735
1736         /* Prevent config work handler from accessing the device. */
1737         mutex_lock(&vi->config_lock);
1738         vi->config_enable = false;
1739         mutex_unlock(&vi->config_lock);
1740
1741         unregister_netdev(vi->dev);
1742
1743         remove_vq_common(vi);
1744         if (vi->alloc_frag.page)
1745                 put_page(vi->alloc_frag.page);
1746
1747         flush_work(&vi->config_work);
1748
1749         free_percpu(vi->stats);
1750         free_netdev(vi->dev);
1751 }
1752
1753 #ifdef CONFIG_PM_SLEEP
1754 static int virtnet_freeze(struct virtio_device *vdev)
1755 {
1756         struct virtnet_info *vi = vdev->priv;
1757         int i;
1758
1759         unregister_hotcpu_notifier(&vi->nb);
1760
1761         /* Prevent config work handler from accessing the device */
1762         mutex_lock(&vi->config_lock);
1763         vi->config_enable = false;
1764         mutex_unlock(&vi->config_lock);
1765
1766         netif_device_detach(vi->dev);
1767         cancel_delayed_work_sync(&vi->refill);
1768
1769         if (netif_running(vi->dev))
1770                 for (i = 0; i < vi->max_queue_pairs; i++) {
1771                         napi_disable(&vi->rq[i].napi);
1772                         netif_napi_del(&vi->rq[i].napi);
1773                 }
1774
1775         remove_vq_common(vi);
1776
1777         flush_work(&vi->config_work);
1778
1779         return 0;
1780 }
1781
1782 static int virtnet_restore(struct virtio_device *vdev)
1783 {
1784         struct virtnet_info *vi = vdev->priv;
1785         int err, i;
1786
1787         err = init_vqs(vi);
1788         if (err)
1789                 return err;
1790
1791         if (netif_running(vi->dev)) {
1792                 for (i = 0; i < vi->curr_queue_pairs; i++)
1793                         if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1794                                 schedule_delayed_work(&vi->refill, 0);
1795
1796                 for (i = 0; i < vi->max_queue_pairs; i++)
1797                         virtnet_napi_enable(&vi->rq[i]);
1798         }
1799
1800         netif_device_attach(vi->dev);
1801
1802         mutex_lock(&vi->config_lock);
1803         vi->config_enable = true;
1804         mutex_unlock(&vi->config_lock);
1805
1806         rtnl_lock();
1807         virtnet_set_queues(vi, vi->curr_queue_pairs);
1808         rtnl_unlock();
1809
1810         err = register_hotcpu_notifier(&vi->nb);
1811         if (err)
1812                 return err;
1813
1814         return 0;
1815 }
1816 #endif
1817
1818 static struct virtio_device_id id_table[] = {
1819         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1820         { 0 },
1821 };
1822
1823 static unsigned int features[] = {
1824         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1825         VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1826         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1827         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1828         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1829         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1830         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1831         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1832         VIRTIO_NET_F_CTRL_MAC_ADDR,
1833         VIRTIO_F_ANY_LAYOUT,
1834 };
1835
1836 static struct virtio_driver virtio_net_driver = {
1837         .feature_table = features,
1838         .feature_table_size = ARRAY_SIZE(features),
1839         .driver.name =  KBUILD_MODNAME,
1840         .driver.owner = THIS_MODULE,
1841         .id_table =     id_table,
1842         .probe =        virtnet_probe,
1843         .remove =       virtnet_remove,
1844         .config_changed = virtnet_config_changed,
1845 #ifdef CONFIG_PM_SLEEP
1846         .freeze =       virtnet_freeze,
1847         .restore =      virtnet_restore,
1848 #endif
1849 };
1850
1851 module_virtio_driver(virtio_net_driver);
1852
1853 MODULE_DEVICE_TABLE(virtio, id_table);
1854 MODULE_DESCRIPTION("Virtio network driver");
1855 MODULE_LICENSE("GPL");