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 #include <linux/average.h>
  30 #include <net/busy_poll.h>
  31
  32 static int napi_weight = NAPI_POLL_WEIGHT;
  33 module_param(napi_weight, int, 0444);
  34
  35 static bool csum = true, gso = true;
  36 module_param(csum, bool, 0444);
  37 module_param(gso, bool, 0444);
  38
  39 /* FIXME: MTU in config. */
  40 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  41 #define GOOD_COPY_LEN   128
  42
  43 /* RX packet size EWMA. The average packet size is used to determine the packet
  44  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  45  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  46  * term, transient changes in packet size.
  47  */
  48 DECLARE_EWMA(pkt_len, 1, 64)
  49
  50 /* With mergeable buffers we align buffer address and use the low bits to
  51  * encode its true size. Buffer size is up to 1 page so we need to align to
  52  * square root of page size to ensure we reserve enough bits to encode the true
  53  * size.
  54  */
  55 #define MERGEABLE_BUFFER_MIN_ALIGN_SHIFT ((PAGE_SHIFT + 1) / 2)
  56
  57 /* Minimum alignment for mergeable packet buffers. */
  58 #define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, \
  59                                    1 << MERGEABLE_BUFFER_MIN_ALIGN_SHIFT)
  60
  61 #define VIRTNET_DRIVER_VERSION "1.0.0"
  62
  63 struct virtnet_stats {
  64         struct u64_stats_sync tx_syncp;
  65         struct u64_stats_sync rx_syncp;
  66         u64 tx_bytes;
  67         u64 tx_packets;
  68
  69         u64 rx_bytes;
  70         u64 rx_packets;
  71 };
  72
  73 /* Internal representation of a send virtqueue */
  74 struct send_queue {
  75         /* Virtqueue associated with this send _queue */
  76         struct virtqueue *vq;
  77
  78         /* TX: fragments + linear part + virtio header */
  79         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  80
  81         /* Name of the send queue: output.$index */
  82         char name[40];
  83 };
  84
  85 /* Internal representation of a receive virtqueue */
  86 struct receive_queue {
  87         /* Virtqueue associated with this receive_queue */
  88         struct virtqueue *vq;
  89
  90         struct napi_struct napi;
  91
  92         /* Chain pages by the private ptr. */
  93         struct page *pages;
  94
  95         /* Average packet length for mergeable receive buffers. */
  96         struct ewma_pkt_len mrg_avg_pkt_len;
  97
  98         /* Page frag for packet buffer allocation. */
  99         struct page_frag alloc_frag;
 100
 101         /* RX: fragments + linear part + virtio header */
 102         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 103
 104         /* Name of this receive queue: input.$index */
 105         char name[40];
 106 };
 107
 108 struct virtnet_info {
 109         struct virtio_device *vdev;
 110         struct virtqueue *cvq;
 111         struct net_device *dev;
 112         struct send_queue *sq;
 113         struct receive_queue *rq;
 114         unsigned int status;
 115
 116         /* Max # of queue pairs supported by the device */
 117         u16 max_queue_pairs;
 118
 119         /* # of queue pairs currently used by the driver */
 120         u16 curr_queue_pairs;
 121
 122         /* I like... big packets and I cannot lie! */
 123         bool big_packets;
 124
 125         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 126         bool mergeable_rx_bufs;
 127
 128         /* Has control virtqueue */
 129         bool has_cvq;
 130
 131         /* Host can handle any s/g split between our header and packet data */
 132         bool any_header_sg;
 133
 134         /* Packet virtio header size */
 135         u8 hdr_len;
 136
 137         /* Active statistics */
 138         struct virtnet_stats __percpu *stats;
 139
 140         /* Work struct for refilling if we run low on memory. */
 141         struct delayed_work refill;
 142
 143         /* Work struct for config space updates */
 144         struct work_struct config_work;
 145
 146         /* Does the affinity hint is set for virtqueues? */
 147         bool affinity_hint_set;
 148
 149         /* CPU hotplug instances for online & dead */
 150         struct hlist_node node;
 151         struct hlist_node node_dead;
 152
 153         /* Control VQ buffers: protected by the rtnl lock */
 154         struct virtio_net_ctrl_hdr ctrl_hdr;
 155         virtio_net_ctrl_ack ctrl_status;
 156         struct virtio_net_ctrl_mq ctrl_mq;
 157         u8 ctrl_promisc;
 158         u8 ctrl_allmulti;
 159         u16 ctrl_vid;
 160
 161         /* Ethtool settings */
 162         u8 duplex;
 163         u32 speed;
 164 };
 165
 166 struct padded_vnet_hdr {
 167         struct virtio_net_hdr_mrg_rxbuf hdr;
 168         /*
 169          * hdr is in a separate sg buffer, and data sg buffer shares same page
 170          * with this header sg. This padding makes next sg 16 byte aligned
 171          * after the header.
 172          */
 173         char padding[4];
 174 };
 175
 176 /* Converting between virtqueue no. and kernel tx/rx queue no.
 177  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 178  */
 179 static int vq2txq(struct virtqueue *vq)
 180 {
 181         return (vq->index - 1) / 2;
 182 }
 183
 184 static int txq2vq(int txq)
 185 {
 186         return txq * 2 + 1;
 187 }
 188
 189 static int vq2rxq(struct virtqueue *vq)
 190 {
 191         return vq->index / 2;
 192 }
 193
 194 static int rxq2vq(int rxq)
 195 {
 196         return rxq * 2;
 197 }
 198
 199 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 200 {
 201         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 202 }
 203
 204 /*
 205  * private is used to chain pages for big packets, put the whole
 206  * most recent used list in the beginning for reuse
 207  */
 208 static void give_pages(struct receive_queue *rq, struct page *page)
 209 {
 210         struct page *end;
 211
 212         /* Find end of list, sew whole thing into vi->rq.pages. */
 213         for (end = page; end->private; end = (struct page *)end->private);
 214         end->private = (unsigned long)rq->pages;
 215         rq->pages = page;
 216 }
 217
 218 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 219 {
 220         struct page *p = rq->pages;
 221
 222         if (p) {
 223                 rq->pages = (struct page *)p->private;
 224                 /* clear private here, it is used to chain pages */
 225                 p->private = 0;
 226         } else
 227                 p = alloc_page(gfp_mask);
 228         return p;
 229 }
 230
 231 static void skb_xmit_done(struct virtqueue *vq)
 232 {
 233         struct virtnet_info *vi = vq->vdev->priv;
 234
 235         /* Suppress further interrupts. */
 236         virtqueue_disable_cb(vq);
 237
 238         /* We were probably waiting for more output buffers. */
 239         netif_wake_subqueue(vi->dev, vq2txq(vq));
 240 }
 241
 242 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
 243 {
 244         unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
 245         return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
 246 }
 247
 248 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
 249 {
 250         return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
 251
 252 }
 253
 254 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
 255 {
 256         unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
 257         return (unsigned long)buf | (size - 1);
 258 }
 259
 260 /* Called from bottom half context */
 261 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 262                                    struct receive_queue *rq,
 263                                    struct page *page, unsigned int offset,
 264                                    unsigned int len, unsigned int truesize)
 265 {
 266         struct sk_buff *skb;
 267         struct virtio_net_hdr_mrg_rxbuf *hdr;
 268         unsigned int copy, hdr_len, hdr_padded_len;
 269         char *p;
 270
 271         p = page_address(page) + offset;
 272
 273         /* copy small packet so we can reuse these pages for small data */
 274         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 275         if (unlikely(!skb))
 276                 return NULL;
 277
 278         hdr = skb_vnet_hdr(skb);
 279
 280         hdr_len = vi->hdr_len;
 281         if (vi->mergeable_rx_bufs)
 282                 hdr_padded_len = sizeof *hdr;
 283         else
 284                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 285
 286         memcpy(hdr, p, hdr_len);
 287
 288         len -= hdr_len;
 289         offset += hdr_padded_len;
 290         p += hdr_padded_len;
 291
 292         copy = len;
 293         if (copy > skb_tailroom(skb))
 294                 copy = skb_tailroom(skb);
 295         memcpy(skb_put(skb, copy), p, copy);
 296
 297         len -= copy;
 298         offset += copy;
 299
 300         if (vi->mergeable_rx_bufs) {
 301                 if (len)
 302                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 303                 else
 304                         put_page(page);
 305                 return skb;
 306         }
 307
 308         /*
 309          * Verify that we can indeed put this data into a skb.
 310          * This is here to handle cases when the device erroneously
 311          * tries to receive more than is possible. This is usually
 312          * the case of a broken device.
 313          */
 314         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 315                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 316                 dev_kfree_skb(skb);
 317                 return NULL;
 318         }
 319         BUG_ON(offset >= PAGE_SIZE);
 320         while (len) {
 321                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 322                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 323                                 frag_size, truesize);
 324                 len -= frag_size;
 325                 page = (struct page *)page->private;
 326                 offset = 0;
 327         }
 328
 329         if (page)
 330                 give_pages(rq, page);
 331
 332         return skb;
 333 }
 334
 335 static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
 336 {
 337         struct sk_buff * skb = buf;
 338
 339         len -= vi->hdr_len;
 340         skb_trim(skb, len);
 341
 342         return skb;
 343 }
 344
 345 static struct sk_buff *receive_big(struct net_device *dev,
 346                                    struct virtnet_info *vi,
 347                                    struct receive_queue *rq,
 348                                    void *buf,
 349                                    unsigned int len)
 350 {
 351         struct page *page = buf;
 352         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 353
 354         if (unlikely(!skb))
 355                 goto err;
 356
 357         return skb;
 358
 359 err:
 360         dev->stats.rx_dropped++;
 361         give_pages(rq, page);
 362         return NULL;
 363 }
 364
 365 static struct sk_buff *receive_mergeable(struct net_device *dev,
 366                                          struct virtnet_info *vi,
 367                                          struct receive_queue *rq,
 368                                          unsigned long ctx,
 369                                          unsigned int len)
 370 {
 371         void *buf = mergeable_ctx_to_buf_address(ctx);
 372         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 373         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 374         struct page *page = virt_to_head_page(buf);
 375         int offset = buf - page_address(page);
 376         unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 377
 378         struct sk_buff *head_skb = page_to_skb(vi, rq, page, offset, len,
 379                                                truesize);
 380         struct sk_buff *curr_skb = head_skb;
 381
 382         if (unlikely(!curr_skb))
 383                 goto err_skb;
 384         while (--num_buf) {
 385                 int num_skb_frags;
 386
 387                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 388                 if (unlikely(!ctx)) {
 389                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 390                                  dev->name, num_buf,
 391                                  virtio16_to_cpu(vi->vdev,
 392                                                  hdr->num_buffers));
 393                         dev->stats.rx_length_errors++;
 394                         goto err_buf;
 395                 }
 396
 397                 buf = mergeable_ctx_to_buf_address(ctx);
 398                 page = virt_to_head_page(buf);
 399
 400                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 401                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 402                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 403
 404                         if (unlikely(!nskb))
 405                                 goto err_skb;
 406                         if (curr_skb == head_skb)
 407                                 skb_shinfo(curr_skb)->frag_list = nskb;
 408                         else
 409                                 curr_skb->next = nskb;
 410                         curr_skb = nskb;
 411                         head_skb->truesize += nskb->truesize;
 412                         num_skb_frags = 0;
 413                 }
 414                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 415                 if (curr_skb != head_skb) {
 416                         head_skb->data_len += len;
 417                         head_skb->len += len;
 418                         head_skb->truesize += truesize;
 419                 }
 420                 offset = buf - page_address(page);
 421                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 422                         put_page(page);
 423                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 424                                              len, truesize);
 425                 } else {
 426                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 427                                         offset, len, truesize);
 428                 }
 429         }
 430
 431         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 432         return head_skb;
 433
 434 err_skb:
 435         put_page(page);
 436         while (--num_buf) {
 437                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 438                 if (unlikely(!ctx)) {
 439                         pr_debug("%s: rx error: %d buffers missing\n",
 440                                  dev->name, num_buf);
 441                         dev->stats.rx_length_errors++;
 442                         break;
 443                 }
 444                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 445                 put_page(page);
 446         }
 447 err_buf:
 448         dev->stats.rx_dropped++;
 449         dev_kfree_skb(head_skb);
 450         return NULL;
 451 }
 452
 453 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 454                         void *buf, unsigned int len)
 455 {
 456         struct net_device *dev = vi->dev;
 457         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 458         struct sk_buff *skb;
 459         struct virtio_net_hdr_mrg_rxbuf *hdr;
 460
 461         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 462                 pr_debug("%s: short packet %i\n", dev->name, len);
 463                 dev->stats.rx_length_errors++;
 464                 if (vi->mergeable_rx_bufs) {
 465                         unsigned long ctx = (unsigned long)buf;
 466                         void *base = mergeable_ctx_to_buf_address(ctx);
 467                         put_page(virt_to_head_page(base));
 468                 } else if (vi->big_packets) {
 469                         give_pages(rq, buf);
 470                 } else {
 471                         dev_kfree_skb(buf);
 472                 }
 473                 return;
 474         }
 475
 476         if (vi->mergeable_rx_bufs)
 477                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
 478         else if (vi->big_packets)
 479                 skb = receive_big(dev, vi, rq, buf, len);
 480         else
 481                 skb = receive_small(vi, buf, len);
 482
 483         if (unlikely(!skb))
 484                 return;
 485
 486         hdr = skb_vnet_hdr(skb);
 487
 488         u64_stats_update_begin(&stats->rx_syncp);
 489         stats->rx_bytes += skb->len;
 490         stats->rx_packets++;
 491         u64_stats_update_end(&stats->rx_syncp);
 492
 493         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 494                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 495
 496         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 497                                   virtio_is_little_endian(vi->vdev))) {
 498                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 499                                      dev->name, hdr->hdr.gso_type,
 500                                      hdr->hdr.gso_size);
 501                 goto frame_err;
 502         }
 503
 504         skb->protocol = eth_type_trans(skb, dev);
 505         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 506                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 507
 508         napi_gro_receive(&rq->napi, skb);
 509         return;
 510
 511 frame_err:
 512         dev->stats.rx_frame_errors++;
 513         dev_kfree_skb(skb);
 514 }
 515
 516 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 517                              gfp_t gfp)
 518 {
 519         struct sk_buff *skb;
 520         struct virtio_net_hdr_mrg_rxbuf *hdr;
 521         int err;
 522
 523         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 524         if (unlikely(!skb))
 525                 return -ENOMEM;
 526
 527         skb_put(skb, GOOD_PACKET_LEN);
 528
 529         hdr = skb_vnet_hdr(skb);
 530         sg_init_table(rq->sg, 2);
 531         sg_set_buf(rq->sg, hdr, vi->hdr_len);
 532         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 533
 534         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 535         if (err < 0)
 536                 dev_kfree_skb(skb);
 537
 538         return err;
 539 }
 540
 541 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 542                            gfp_t gfp)
 543 {
 544         struct page *first, *list = NULL;
 545         char *p;
 546         int i, err, offset;
 547
 548         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 549
 550         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 551         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 552                 first = get_a_page(rq, gfp);
 553                 if (!first) {
 554                         if (list)
 555                                 give_pages(rq, list);
 556                         return -ENOMEM;
 557                 }
 558                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 559
 560                 /* chain new page in list head to match sg */
 561                 first->private = (unsigned long)list;
 562                 list = first;
 563         }
 564
 565         first = get_a_page(rq, gfp);
 566         if (!first) {
 567                 give_pages(rq, list);
 568                 return -ENOMEM;
 569         }
 570         p = page_address(first);
 571
 572         /* rq->sg[0], rq->sg[1] share the same page */
 573         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 574         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 575
 576         /* rq->sg[1] for data packet, from offset */
 577         offset = sizeof(struct padded_vnet_hdr);
 578         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 579
 580         /* chain first in list head */
 581         first->private = (unsigned long)list;
 582         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 583                                   first, gfp);
 584         if (err < 0)
 585                 give_pages(rq, first);
 586
 587         return err;
 588 }
 589
 590 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
 591 {
 592         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 593         unsigned int len;
 594
 595         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
 596                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 597         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 598 }
 599
 600 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 601 {
 602         struct page_frag *alloc_frag = &rq->alloc_frag;
 603         char *buf;
 604         unsigned long ctx;
 605         int err;
 606         unsigned int len, hole;
 607
 608         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 609         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 610                 return -ENOMEM;
 611
 612         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 613         ctx = mergeable_buf_to_ctx(buf, len);
 614         get_page(alloc_frag->page);
 615         alloc_frag->offset += len;
 616         hole = alloc_frag->size - alloc_frag->offset;
 617         if (hole < len) {
 618                 /* To avoid internal fragmentation, if there is very likely not
 619                  * enough space for another buffer, add the remaining space to
 620                  * the current buffer. This extra space is not included in
 621                  * the truesize stored in ctx.
 622                  */
 623                 len += hole;
 624                 alloc_frag->offset += hole;
 625         }
 626
 627         sg_init_one(rq->sg, buf, len);
 628         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 629         if (err < 0)
 630                 put_page(virt_to_head_page(buf));
 631
 632         return err;
 633 }
 634
 635 /*
 636  * Returns false if we couldn't fill entirely (OOM).
 637  *
 638  * Normally run in the receive path, but can also be run from ndo_open
 639  * before we're receiving packets, or from refill_work which is
 640  * careful to disable receiving (using napi_disable).
 641  */
 642 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 643                           gfp_t gfp)
 644 {
 645         int err;
 646         bool oom;
 647
 648         gfp |= __GFP_COLD;
 649         do {
 650                 if (vi->mergeable_rx_bufs)
 651                         err = add_recvbuf_mergeable(rq, gfp);
 652                 else if (vi->big_packets)
 653                         err = add_recvbuf_big(vi, rq, gfp);
 654                 else
 655                         err = add_recvbuf_small(vi, rq, gfp);
 656
 657                 oom = err == -ENOMEM;
 658                 if (err)
 659                         break;
 660         } while (rq->vq->num_free);
 661         virtqueue_kick(rq->vq);
 662         return !oom;
 663 }
 664
 665 static void skb_recv_done(struct virtqueue *rvq)
 666 {
 667         struct virtnet_info *vi = rvq->vdev->priv;
 668         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 669
 670         /* Schedule NAPI, Suppress further interrupts if successful. */
 671         if (napi_schedule_prep(&rq->napi)) {
 672                 virtqueue_disable_cb(rvq);
 673                 __napi_schedule(&rq->napi);
 674         }
 675 }
 676
 677 static void virtnet_napi_enable(struct receive_queue *rq)
 678 {
 679         napi_enable(&rq->napi);
 680
 681         /* If all buffers were filled by other side before we napi_enabled, we
 682          * won't get another interrupt, so process any outstanding packets
 683          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 684          * We synchronize against interrupts via NAPI_STATE_SCHED */
 685         if (napi_schedule_prep(&rq->napi)) {
 686                 virtqueue_disable_cb(rq->vq);
 687                 local_bh_disable();
 688                 __napi_schedule(&rq->napi);
 689                 local_bh_enable();
 690         }
 691 }
 692
 693 static void refill_work(struct work_struct *work)
 694 {
 695         struct virtnet_info *vi =
 696                 container_of(work, struct virtnet_info, refill.work);
 697         bool still_empty;
 698         int i;
 699
 700         for (i = 0; i < vi->curr_queue_pairs; i++) {
 701                 struct receive_queue *rq = &vi->rq[i];
 702
 703                 napi_disable(&rq->napi);
 704                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
 705                 virtnet_napi_enable(rq);
 706
 707                 /* In theory, this can happen: if we don't get any buffers in
 708                  * we will *never* try to fill again.
 709                  */
 710                 if (still_empty)
 711                         schedule_delayed_work(&vi->refill, HZ/2);
 712         }
 713 }
 714
 715 static int virtnet_receive(struct receive_queue *rq, int budget)
 716 {
 717         struct virtnet_info *vi = rq->vq->vdev->priv;
 718         unsigned int len, received = 0;
 719         void *buf;
 720
 721         while (received < budget &&
 722                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 723                 receive_buf(vi, rq, buf, len);
 724                 received++;
 725         }
 726
 727         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 728                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 729                         schedule_delayed_work(&vi->refill, 0);
 730         }
 731
 732         return received;
 733 }
 734
 735 static int virtnet_poll(struct napi_struct *napi, int budget)
 736 {
 737         struct receive_queue *rq =
 738                 container_of(napi, struct receive_queue, napi);
 739         unsigned int r, received;
 740
 741         received = virtnet_receive(rq, budget);
 742
 743         /* Out of packets? */
 744         if (received < budget) {
 745                 r = virtqueue_enable_cb_prepare(rq->vq);
 746                 napi_complete_done(napi, received);
 747                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 748                     napi_schedule_prep(napi)) {
 749                         virtqueue_disable_cb(rq->vq);
 750                         __napi_schedule(napi);
 751                 }
 752         }
 753
 754         return received;
 755 }
 756
 757 #ifdef CONFIG_NET_RX_BUSY_POLL
 758 /* must be called with local_bh_disable()d */
 759 static int virtnet_busy_poll(struct napi_struct *napi)
 760 {
 761         struct receive_queue *rq =
 762                 container_of(napi, struct receive_queue, napi);
 763         struct virtnet_info *vi = rq->vq->vdev->priv;
 764         int r, received = 0, budget = 4;
 765
 766         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
 767                 return LL_FLUSH_FAILED;
 768
 769         if (!napi_schedule_prep(napi))
 770                 return LL_FLUSH_BUSY;
 771
 772         virtqueue_disable_cb(rq->vq);
 773
 774 again:
 775         received += virtnet_receive(rq, budget);
 776
 777         r = virtqueue_enable_cb_prepare(rq->vq);
 778         clear_bit(NAPI_STATE_SCHED, &napi->state);
 779         if (unlikely(virtqueue_poll(rq->vq, r)) &&
 780             napi_schedule_prep(napi)) {
 781                 virtqueue_disable_cb(rq->vq);
 782                 if (received < budget) {
 783                         budget -= received;
 784                         goto again;
 785                 } else {
 786                         __napi_schedule(napi);
 787                 }
 788         }
 789
 790         return received;
 791 }
 792 #endif  /* CONFIG_NET_RX_BUSY_POLL */
 793
 794 static int virtnet_open(struct net_device *dev)
 795 {
 796         struct virtnet_info *vi = netdev_priv(dev);
 797         int i;
 798
 799         for (i = 0; i < vi->max_queue_pairs; i++) {
 800                 if (i < vi->curr_queue_pairs)
 801                         /* Make sure we have some buffers: if oom use wq. */
 802                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
 803                                 schedule_delayed_work(&vi->refill, 0);
 804                 virtnet_napi_enable(&vi->rq[i]);
 805         }
 806
 807         return 0;
 808 }
 809
 810 static void free_old_xmit_skbs(struct send_queue *sq)
 811 {
 812         struct sk_buff *skb;
 813         unsigned int len;
 814         struct virtnet_info *vi = sq->vq->vdev->priv;
 815         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 816
 817         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 818                 pr_debug("Sent skb %p\n", skb);
 819
 820                 u64_stats_update_begin(&stats->tx_syncp);
 821                 stats->tx_bytes += skb->len;
 822                 stats->tx_packets++;
 823                 u64_stats_update_end(&stats->tx_syncp);
 824
 825                 dev_kfree_skb_any(skb);
 826         }
 827 }
 828
 829 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 830 {
 831         struct virtio_net_hdr_mrg_rxbuf *hdr;
 832         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 833         struct virtnet_info *vi = sq->vq->vdev->priv;
 834         unsigned num_sg;
 835         unsigned hdr_len = vi->hdr_len;
 836         bool can_push;
 837
 838         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 839
 840         can_push = vi->any_header_sg &&
 841                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 842                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 843         /* Even if we can, don't push here yet as this would skew
 844          * csum_start offset below. */
 845         if (can_push)
 846                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
 847         else
 848                 hdr = skb_vnet_hdr(skb);
 849
 850         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
 851                                     virtio_is_little_endian(vi->vdev), false))
 852                 BUG();
 853
 854         if (vi->mergeable_rx_bufs)
 855                 hdr->num_buffers = 0;
 856
 857         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
 858         if (can_push) {
 859                 __skb_push(skb, hdr_len);
 860                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
 861                 /* Pull header back to avoid skew in tx bytes calculations. */
 862                 __skb_pull(skb, hdr_len);
 863         } else {
 864                 sg_set_buf(sq->sg, hdr, hdr_len);
 865                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
 866         }
 867         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 868 }
 869
 870 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 871 {
 872         struct virtnet_info *vi = netdev_priv(dev);
 873         int qnum = skb_get_queue_mapping(skb);
 874         struct send_queue *sq = &vi->sq[qnum];
 875         int err;
 876         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
 877         bool kick = !skb->xmit_more;
 878
 879         /* Free up any pending old buffers before queueing new ones. */
 880         free_old_xmit_skbs(sq);
 881
 882         /* timestamp packet in software */
 883         skb_tx_timestamp(skb);
 884
 885         /* Try to transmit */
 886         err = xmit_skb(sq, skb);
 887
 888         /* This should not happen! */
 889         if (unlikely(err)) {
 890                 dev->stats.tx_fifo_errors++;
 891                 if (net_ratelimit())
 892                         dev_warn(&dev->dev,
 893                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
 894                 dev->stats.tx_dropped++;
 895                 dev_kfree_skb_any(skb);
 896                 return NETDEV_TX_OK;
 897         }
 898
 899         /* Don't wait up for transmitted skbs to be freed. */
 900         skb_orphan(skb);
 901         nf_reset(skb);
 902
 903         /* If running out of space, stop queue to avoid getting packets that we
 904          * are then unable to transmit.
 905          * An alternative would be to force queuing layer to requeue the skb by
 906          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
 907          * returned in a normal path of operation: it means that driver is not
 908          * maintaining the TX queue stop/start state properly, and causes
 909          * the stack to do a non-trivial amount of useless work.
 910          * Since most packets only take 1 or 2 ring slots, stopping the queue
 911          * early means 16 slots are typically wasted.
 912          */
 913         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 914                 netif_stop_subqueue(dev, qnum);
 915                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 916                         /* More just got used, free them then recheck. */
 917                         free_old_xmit_skbs(sq);
 918                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 919                                 netif_start_subqueue(dev, qnum);
 920                                 virtqueue_disable_cb(sq->vq);
 921                         }
 922                 }
 923         }
 924
 925         if (kick || netif_xmit_stopped(txq))
 926                 virtqueue_kick(sq->vq);
 927
 928         return NETDEV_TX_OK;
 929 }
 930
 931 /*
 932  * Send command via the control virtqueue and check status.  Commands
 933  * supported by the hypervisor, as indicated by feature bits, should
 934  * never fail unless improperly formatted.
 935  */
 936 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
 937                                  struct scatterlist *out)
 938 {
 939         struct scatterlist *sgs[4], hdr, stat;
 940         unsigned out_num = 0, tmp;
 941
 942         /* Caller should know better */
 943         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
 944
 945         vi->ctrl_status = ~0;
 946         vi->ctrl_hdr.class = class;
 947         vi->ctrl_hdr.cmd = cmd;
 948         /* Add header */
 949         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
 950         sgs[out_num++] = &hdr;
 951
 952         if (out)
 953                 sgs[out_num++] = out;
 954
 955         /* Add return status. */
 956         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
 957         sgs[out_num] = &stat;
 958
 959         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
 960         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
 961
 962         if (unlikely(!virtqueue_kick(vi->cvq)))
 963                 return vi->ctrl_status == VIRTIO_NET_OK;
 964
 965         /* Spin for a response, the kick causes an ioport write, trapping
 966          * into the hypervisor, so the request should be handled immediately.
 967          */
 968         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
 969                !virtqueue_is_broken(vi->cvq))
 970                 cpu_relax();
 971
 972         return vi->ctrl_status == VIRTIO_NET_OK;
 973 }
 974
 975 static int virtnet_set_mac_address(struct net_device *dev, void *p)
 976 {
 977         struct virtnet_info *vi = netdev_priv(dev);
 978         struct virtio_device *vdev = vi->vdev;
 979         int ret;
 980         struct sockaddr *addr;
 981         struct scatterlist sg;
 982
 983         addr = kmalloc(sizeof(*addr), GFP_KERNEL);
 984         if (!addr)
 985                 return -ENOMEM;
 986         memcpy(addr, p, sizeof(*addr));
 987
 988         ret = eth_prepare_mac_addr_change(dev, addr);
 989         if (ret)
 990                 goto out;
 991
 992         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
 993                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
 994                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
 995                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
 996                         dev_warn(&vdev->dev,
 997                                  "Failed to set mac address by vq command.\n");
 998                         ret = -EINVAL;
 999                         goto out;
1000                 }
1001         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1002                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1003                 unsigned int i;
1004
1005                 /* Naturally, this has an atomicity problem. */
1006                 for (i = 0; i < dev->addr_len; i++)
1007                         virtio_cwrite8(vdev,
1008                                        offsetof(struct virtio_net_config, mac) +
1009                                        i, addr->sa_data[i]);
1010         }
1011
1012         eth_commit_mac_addr_change(dev, p);
1013         ret = 0;
1014
1015 out:
1016         kfree(addr);
1017         return ret;
1018 }
1019
1020 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1021                                                struct rtnl_link_stats64 *tot)
1022 {
1023         struct virtnet_info *vi = netdev_priv(dev);
1024         int cpu;
1025         unsigned int start;
1026
1027         for_each_possible_cpu(cpu) {
1028                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1029                 u64 tpackets, tbytes, rpackets, rbytes;
1030
1031                 do {
1032                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1033                         tpackets = stats->tx_packets;
1034                         tbytes   = stats->tx_bytes;
1035                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1036
1037                 do {
1038                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1039                         rpackets = stats->rx_packets;
1040                         rbytes   = stats->rx_bytes;
1041                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1042
1043                 tot->rx_packets += rpackets;
1044                 tot->tx_packets += tpackets;
1045                 tot->rx_bytes   += rbytes;
1046                 tot->tx_bytes   += tbytes;
1047         }
1048
1049         tot->tx_dropped = dev->stats.tx_dropped;
1050         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1051         tot->rx_dropped = dev->stats.rx_dropped;
1052         tot->rx_length_errors = dev->stats.rx_length_errors;
1053         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1054
1055         return tot;
1056 }
1057
1058 #ifdef CONFIG_NET_POLL_CONTROLLER
1059 static void virtnet_netpoll(struct net_device *dev)
1060 {
1061         struct virtnet_info *vi = netdev_priv(dev);
1062         int i;
1063
1064         for (i = 0; i < vi->curr_queue_pairs; i++)
1065                 napi_schedule(&vi->rq[i].napi);
1066 }
1067 #endif
1068
1069 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1070 {
1071         rtnl_lock();
1072         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1073                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1074                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1075         rtnl_unlock();
1076 }
1077
1078 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1079 {
1080         struct scatterlist sg;
1081         struct net_device *dev = vi->dev;
1082
1083         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1084                 return 0;
1085
1086         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1087         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1088
1089         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1090                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1091                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1092                          queue_pairs);
1093                 return -EINVAL;
1094         } else {
1095                 vi->curr_queue_pairs = queue_pairs;
1096                 /* virtnet_open() will refill when device is going to up. */
1097                 if (dev->flags & IFF_UP)
1098                         schedule_delayed_work(&vi->refill, 0);
1099         }
1100
1101         return 0;
1102 }
1103
1104 static int virtnet_close(struct net_device *dev)
1105 {
1106         struct virtnet_info *vi = netdev_priv(dev);
1107         int i;
1108
1109         /* Make sure refill_work doesn't re-enable napi! */
1110         cancel_delayed_work_sync(&vi->refill);
1111
1112         for (i = 0; i < vi->max_queue_pairs; i++)
1113                 napi_disable(&vi->rq[i].napi);
1114
1115         return 0;
1116 }
1117
1118 static void virtnet_set_rx_mode(struct net_device *dev)
1119 {
1120         struct virtnet_info *vi = netdev_priv(dev);
1121         struct scatterlist sg[2];
1122         struct virtio_net_ctrl_mac *mac_data;
1123         struct netdev_hw_addr *ha;
1124         int uc_count;
1125         int mc_count;
1126         void *buf;
1127         int i;
1128
1129         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1130         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1131                 return;
1132
1133         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1134         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1135
1136         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1137
1138         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1139                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1140                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1141                          vi->ctrl_promisc ? "en" : "dis");
1142
1143         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1144
1145         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1146                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1147                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1148                          vi->ctrl_allmulti ? "en" : "dis");
1149
1150         uc_count = netdev_uc_count(dev);
1151         mc_count = netdev_mc_count(dev);
1152         /* MAC filter - use one buffer for both lists */
1153         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1154                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1155         mac_data = buf;
1156         if (!buf)
1157                 return;
1158
1159         sg_init_table(sg, 2);
1160
1161         /* Store the unicast list and count in the front of the buffer */
1162         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1163         i = 0;
1164         netdev_for_each_uc_addr(ha, dev)
1165                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1166
1167         sg_set_buf(&sg[0], mac_data,
1168                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1169
1170         /* multicast list and count fill the end */
1171         mac_data = (void *)&mac_data->macs[uc_count][0];
1172
1173         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1174         i = 0;
1175         netdev_for_each_mc_addr(ha, dev)
1176                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1177
1178         sg_set_buf(&sg[1], mac_data,
1179                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1180
1181         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1182                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1183                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1184
1185         kfree(buf);
1186 }
1187
1188 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1189                                    __be16 proto, u16 vid)
1190 {
1191         struct virtnet_info *vi = netdev_priv(dev);
1192         struct scatterlist sg;
1193
1194         vi->ctrl_vid = vid;
1195         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1196
1197         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1198                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1199                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1200         return 0;
1201 }
1202
1203 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1204                                     __be16 proto, u16 vid)
1205 {
1206         struct virtnet_info *vi = netdev_priv(dev);
1207         struct scatterlist sg;
1208
1209         vi->ctrl_vid = vid;
1210         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1211
1212         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1213                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1214                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1215         return 0;
1216 }
1217
1218 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1219 {
1220         int i;
1221
1222         if (vi->affinity_hint_set) {
1223                 for (i = 0; i < vi->max_queue_pairs; i++) {
1224                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1225                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1226                 }
1227
1228                 vi->affinity_hint_set = false;
1229         }
1230 }
1231
1232 static void virtnet_set_affinity(struct virtnet_info *vi)
1233 {
1234         int i;
1235         int cpu;
1236
1237         /* In multiqueue mode, when the number of cpu is equal to the number of
1238          * queue pairs, we let the queue pairs to be private to one cpu by
1239          * setting the affinity hint to eliminate the contention.
1240          */
1241         if (vi->curr_queue_pairs == 1 ||
1242             vi->max_queue_pairs != num_online_cpus()) {
1243                 virtnet_clean_affinity(vi, -1);
1244                 return;
1245         }
1246
1247         i = 0;
1248         for_each_online_cpu(cpu) {
1249                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1250                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1251                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1252                 i++;
1253         }
1254
1255         vi->affinity_hint_set = true;
1256 }
1257
1258 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1259 {
1260         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1261                                                    node);
1262         virtnet_set_affinity(vi);
1263         return 0;
1264 }
1265
1266 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1267 {
1268         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1269                                                    node_dead);
1270         virtnet_set_affinity(vi);
1271         return 0;
1272 }
1273
1274 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1275 {
1276         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1277                                                    node);
1278
1279         virtnet_clean_affinity(vi, cpu);
1280         return 0;
1281 }
1282
1283 static enum cpuhp_state virtionet_online;
1284
1285 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1286 {
1287         int ret;
1288
1289         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1290         if (ret)
1291                 return ret;
1292         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1293                                                &vi->node_dead);
1294         if (!ret)
1295                 return ret;
1296         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1297         return ret;
1298 }
1299
1300 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1301 {
1302         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1303         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1304                                             &vi->node_dead);
1305 }
1306
1307 static void virtnet_get_ringparam(struct net_device *dev,
1308                                 struct ethtool_ringparam *ring)
1309 {
1310         struct virtnet_info *vi = netdev_priv(dev);
1311
1312         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1313         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1314         ring->rx_pending = ring->rx_max_pending;
1315         ring->tx_pending = ring->tx_max_pending;
1316 }
1317
1318
1319 static void virtnet_get_drvinfo(struct net_device *dev,
1320                                 struct ethtool_drvinfo *info)
1321 {
1322         struct virtnet_info *vi = netdev_priv(dev);
1323         struct virtio_device *vdev = vi->vdev;
1324
1325         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1326         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1327         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1328
1329 }
1330
1331 /* TODO: Eliminate OOO packets during switching */
1332 static int virtnet_set_channels(struct net_device *dev,
1333                                 struct ethtool_channels *channels)
1334 {
1335         struct virtnet_info *vi = netdev_priv(dev);
1336         u16 queue_pairs = channels->combined_count;
1337         int err;
1338
1339         /* We don't support separate rx/tx channels.
1340          * We don't allow setting 'other' channels.
1341          */
1342         if (channels->rx_count || channels->tx_count || channels->other_count)
1343                 return -EINVAL;
1344
1345         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1346                 return -EINVAL;
1347
1348         get_online_cpus();
1349         err = virtnet_set_queues(vi, queue_pairs);
1350         if (!err) {
1351                 netif_set_real_num_tx_queues(dev, queue_pairs);
1352                 netif_set_real_num_rx_queues(dev, queue_pairs);
1353
1354                 virtnet_set_affinity(vi);
1355         }
1356         put_online_cpus();
1357
1358         return err;
1359 }
1360
1361 static void virtnet_get_channels(struct net_device *dev,
1362                                  struct ethtool_channels *channels)
1363 {
1364         struct virtnet_info *vi = netdev_priv(dev);
1365
1366         channels->combined_count = vi->curr_queue_pairs;
1367         channels->max_combined = vi->max_queue_pairs;
1368         channels->max_other = 0;
1369         channels->rx_count = 0;
1370         channels->tx_count = 0;
1371         channels->other_count = 0;
1372 }
1373
1374 /* Check if the user is trying to change anything besides speed/duplex */
1375 static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1376 {
1377         struct ethtool_cmd diff1 = *cmd;
1378         struct ethtool_cmd diff2 = {};
1379
1380         /* cmd is always set so we need to clear it, validate the port type
1381          * and also without autonegotiation we can ignore advertising
1382          */
1383         ethtool_cmd_speed_set(&diff1, 0);
1384         diff2.port = PORT_OTHER;
1385         diff1.advertising = 0;
1386         diff1.duplex = 0;
1387         diff1.cmd = 0;
1388
1389         return !memcmp(&diff1, &diff2, sizeof(diff1));
1390 }
1391
1392 static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1393 {
1394         struct virtnet_info *vi = netdev_priv(dev);
1395         u32 speed;
1396
1397         speed = ethtool_cmd_speed(cmd);
1398         /* don't allow custom speed and duplex */
1399         if (!ethtool_validate_speed(speed) ||
1400             !ethtool_validate_duplex(cmd->duplex) ||
1401             !virtnet_validate_ethtool_cmd(cmd))
1402                 return -EINVAL;
1403         vi->speed = speed;
1404         vi->duplex = cmd->duplex;
1405
1406         return 0;
1407 }
1408
1409 static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1410 {
1411         struct virtnet_info *vi = netdev_priv(dev);
1412
1413         ethtool_cmd_speed_set(cmd, vi->speed);
1414         cmd->duplex = vi->duplex;
1415         cmd->port = PORT_OTHER;
1416
1417         return 0;
1418 }
1419
1420 static void virtnet_init_settings(struct net_device *dev)
1421 {
1422         struct virtnet_info *vi = netdev_priv(dev);
1423
1424         vi->speed = SPEED_UNKNOWN;
1425         vi->duplex = DUPLEX_UNKNOWN;
1426 }
1427
1428 static const struct ethtool_ops virtnet_ethtool_ops = {
1429         .get_drvinfo = virtnet_get_drvinfo,
1430         .get_link = ethtool_op_get_link,
1431         .get_ringparam = virtnet_get_ringparam,
1432         .set_channels = virtnet_set_channels,
1433         .get_channels = virtnet_get_channels,
1434         .get_ts_info = ethtool_op_get_ts_info,
1435         .get_settings = virtnet_get_settings,
1436         .set_settings = virtnet_set_settings,
1437 };
1438
1439 #define MIN_MTU 68
1440 #define MAX_MTU 65535
1441
1442 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1443 {
1444         if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1445                 return -EINVAL;
1446         dev->mtu = new_mtu;
1447         return 0;
1448 }
1449
1450 static const struct net_device_ops virtnet_netdev = {
1451         .ndo_open            = virtnet_open,
1452         .ndo_stop            = virtnet_close,
1453         .ndo_start_xmit      = start_xmit,
1454         .ndo_validate_addr   = eth_validate_addr,
1455         .ndo_set_mac_address = virtnet_set_mac_address,
1456         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1457         .ndo_change_mtu      = virtnet_change_mtu,
1458         .ndo_get_stats64     = virtnet_stats,
1459         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1460         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1461 #ifdef CONFIG_NET_POLL_CONTROLLER
1462         .ndo_poll_controller = virtnet_netpoll,
1463 #endif
1464 #ifdef CONFIG_NET_RX_BUSY_POLL
1465         .ndo_busy_poll          = virtnet_busy_poll,
1466 #endif
1467         .ndo_features_check     = passthru_features_check,
1468 };
1469
1470 static void virtnet_config_changed_work(struct work_struct *work)
1471 {
1472         struct virtnet_info *vi =
1473                 container_of(work, struct virtnet_info, config_work);
1474         u16 v;
1475
1476         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1477                                  struct virtio_net_config, status, &v) < 0)
1478                 return;
1479
1480         if (v & VIRTIO_NET_S_ANNOUNCE) {
1481                 netdev_notify_peers(vi->dev);
1482                 virtnet_ack_link_announce(vi);
1483         }
1484
1485         /* Ignore unknown (future) status bits */
1486         v &= VIRTIO_NET_S_LINK_UP;
1487
1488         if (vi->status == v)
1489                 return;
1490
1491         vi->status = v;
1492
1493         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1494                 netif_carrier_on(vi->dev);
1495                 netif_tx_wake_all_queues(vi->dev);
1496         } else {
1497                 netif_carrier_off(vi->dev);
1498                 netif_tx_stop_all_queues(vi->dev);
1499         }
1500 }
1501
1502 static void virtnet_config_changed(struct virtio_device *vdev)
1503 {
1504         struct virtnet_info *vi = vdev->priv;
1505
1506         schedule_work(&vi->config_work);
1507 }
1508
1509 static void virtnet_free_queues(struct virtnet_info *vi)
1510 {
1511         int i;
1512
1513         for (i = 0; i < vi->max_queue_pairs; i++) {
1514                 napi_hash_del(&vi->rq[i].napi);
1515                 netif_napi_del(&vi->rq[i].napi);
1516         }
1517
1518         /* We called napi_hash_del() before netif_napi_del(),
1519          * we need to respect an RCU grace period before freeing vi->rq
1520          */
1521         synchronize_net();
1522
1523         kfree(vi->rq);
1524         kfree(vi->sq);
1525 }
1526
1527 static void free_receive_bufs(struct virtnet_info *vi)
1528 {
1529         int i;
1530
1531         for (i = 0; i < vi->max_queue_pairs; i++) {
1532                 while (vi->rq[i].pages)
1533                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1534         }
1535 }
1536
1537 static void free_receive_page_frags(struct virtnet_info *vi)
1538 {
1539         int i;
1540         for (i = 0; i < vi->max_queue_pairs; i++)
1541                 if (vi->rq[i].alloc_frag.page)
1542                         put_page(vi->rq[i].alloc_frag.page);
1543 }
1544
1545 static void free_unused_bufs(struct virtnet_info *vi)
1546 {
1547         void *buf;
1548         int i;
1549
1550         for (i = 0; i < vi->max_queue_pairs; i++) {
1551                 struct virtqueue *vq = vi->sq[i].vq;
1552                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1553                         dev_kfree_skb(buf);
1554         }
1555
1556         for (i = 0; i < vi->max_queue_pairs; i++) {
1557                 struct virtqueue *vq = vi->rq[i].vq;
1558
1559                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1560                         if (vi->mergeable_rx_bufs) {
1561                                 unsigned long ctx = (unsigned long)buf;
1562                                 void *base = mergeable_ctx_to_buf_address(ctx);
1563                                 put_page(virt_to_head_page(base));
1564                         } else if (vi->big_packets) {
1565                                 give_pages(&vi->rq[i], buf);
1566                         } else {
1567                                 dev_kfree_skb(buf);
1568                         }
1569                 }
1570         }
1571 }
1572
1573 static void virtnet_del_vqs(struct virtnet_info *vi)
1574 {
1575         struct virtio_device *vdev = vi->vdev;
1576
1577         virtnet_clean_affinity(vi, -1);
1578
1579         vdev->config->del_vqs(vdev);
1580
1581         virtnet_free_queues(vi);
1582 }
1583
1584 static int virtnet_find_vqs(struct virtnet_info *vi)
1585 {
1586         vq_callback_t **callbacks;
1587         struct virtqueue **vqs;
1588         int ret = -ENOMEM;
1589         int i, total_vqs;
1590         const char **names;
1591
1592         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1593          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1594          * possible control vq.
1595          */
1596         total_vqs = vi->max_queue_pairs * 2 +
1597                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1598
1599         /* Allocate space for find_vqs parameters */
1600         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1601         if (!vqs)
1602                 goto err_vq;
1603         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1604         if (!callbacks)
1605                 goto err_callback;
1606         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1607         if (!names)
1608                 goto err_names;
1609
1610         /* Parameters for control virtqueue, if any */
1611         if (vi->has_cvq) {
1612                 callbacks[total_vqs - 1] = NULL;
1613                 names[total_vqs - 1] = "control";
1614         }
1615
1616         /* Allocate/initialize parameters for send/receive virtqueues */
1617         for (i = 0; i < vi->max_queue_pairs; i++) {
1618                 callbacks[rxq2vq(i)] = skb_recv_done;
1619                 callbacks[txq2vq(i)] = skb_xmit_done;
1620                 sprintf(vi->rq[i].name, "input.%d", i);
1621                 sprintf(vi->sq[i].name, "output.%d", i);
1622                 names[rxq2vq(i)] = vi->rq[i].name;
1623                 names[txq2vq(i)] = vi->sq[i].name;
1624         }
1625
1626         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1627                                          names);
1628         if (ret)
1629                 goto err_find;
1630
1631         if (vi->has_cvq) {
1632                 vi->cvq = vqs[total_vqs - 1];
1633                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1634                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1635         }
1636
1637         for (i = 0; i < vi->max_queue_pairs; i++) {
1638                 vi->rq[i].vq = vqs[rxq2vq(i)];
1639                 vi->sq[i].vq = vqs[txq2vq(i)];
1640         }
1641
1642         kfree(names);
1643         kfree(callbacks);
1644         kfree(vqs);
1645
1646         return 0;
1647
1648 err_find:
1649         kfree(names);
1650 err_names:
1651         kfree(callbacks);
1652 err_callback:
1653         kfree(vqs);
1654 err_vq:
1655         return ret;
1656 }
1657
1658 static int virtnet_alloc_queues(struct virtnet_info *vi)
1659 {
1660         int i;
1661
1662         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1663         if (!vi->sq)
1664                 goto err_sq;
1665         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1666         if (!vi->rq)
1667                 goto err_rq;
1668
1669         INIT_DELAYED_WORK(&vi->refill, refill_work);
1670         for (i = 0; i < vi->max_queue_pairs; i++) {
1671                 vi->rq[i].pages = NULL;
1672                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1673                                napi_weight);
1674
1675                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1676                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
1677                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1678         }
1679
1680         return 0;
1681
1682 err_rq:
1683         kfree(vi->sq);
1684 err_sq:
1685         return -ENOMEM;
1686 }
1687
1688 static int init_vqs(struct virtnet_info *vi)
1689 {
1690         int ret;
1691
1692         /* Allocate send & receive queues */
1693         ret = virtnet_alloc_queues(vi);
1694         if (ret)
1695                 goto err;
1696
1697         ret = virtnet_find_vqs(vi);
1698         if (ret)
1699                 goto err_free;
1700
1701         get_online_cpus();
1702         virtnet_set_affinity(vi);
1703         put_online_cpus();
1704
1705         return 0;
1706
1707 err_free:
1708         virtnet_free_queues(vi);
1709 err:
1710         return ret;
1711 }
1712
1713 #ifdef CONFIG_SYSFS
1714 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1715                 struct rx_queue_attribute *attribute, char *buf)
1716 {
1717         struct virtnet_info *vi = netdev_priv(queue->dev);
1718         unsigned int queue_index = get_netdev_rx_queue_index(queue);
1719         struct ewma_pkt_len *avg;
1720
1721         BUG_ON(queue_index >= vi->max_queue_pairs);
1722         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1723         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1724 }
1725
1726 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1727         __ATTR_RO(mergeable_rx_buffer_size);
1728
1729 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1730         &mergeable_rx_buffer_size_attribute.attr,
1731         NULL
1732 };
1733
1734 static const struct attribute_group virtio_net_mrg_rx_group = {
1735         .name = "virtio_net",
1736         .attrs = virtio_net_mrg_rx_attrs
1737 };
1738 #endif
1739
1740 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1741                                     unsigned int fbit,
1742                                     const char *fname, const char *dname)
1743 {
1744         if (!virtio_has_feature(vdev, fbit))
1745                 return false;
1746
1747         dev_err(&vdev->dev, "device advertises feature %s but not %s",
1748                 fname, dname);
1749
1750         return true;
1751 }
1752
1753 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
1754         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1755
1756 static bool virtnet_validate_features(struct virtio_device *vdev)
1757 {
1758         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1759             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1760                              "VIRTIO_NET_F_CTRL_VQ") ||
1761              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1762                              "VIRTIO_NET_F_CTRL_VQ") ||
1763              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1764                              "VIRTIO_NET_F_CTRL_VQ") ||
1765              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1766              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1767                              "VIRTIO_NET_F_CTRL_VQ"))) {
1768                 return false;
1769         }
1770
1771         return true;
1772 }
1773
1774 static int virtnet_probe(struct virtio_device *vdev)
1775 {
1776         int i, err;
1777         struct net_device *dev;
1778         struct virtnet_info *vi;
1779         u16 max_queue_pairs;
1780         int mtu;
1781
1782         if (!vdev->config->get) {
1783                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
1784                         __func__);
1785                 return -EINVAL;
1786         }
1787
1788         if (!virtnet_validate_features(vdev))
1789                 return -EINVAL;
1790
1791         /* Find if host supports multiqueue virtio_net device */
1792         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1793                                    struct virtio_net_config,
1794                                    max_virtqueue_pairs, &max_queue_pairs);
1795
1796         /* We need at least 2 queue's */
1797         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1798             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1799             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1800                 max_queue_pairs = 1;
1801
1802         /* Allocate ourselves a network device with room for our info */
1803         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1804         if (!dev)
1805                 return -ENOMEM;
1806
1807         /* Set up network device as normal. */
1808         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1809         dev->netdev_ops = &virtnet_netdev;
1810         dev->features = NETIF_F_HIGHDMA;
1811
1812         dev->ethtool_ops = &virtnet_ethtool_ops;
1813         SET_NETDEV_DEV(dev, &vdev->dev);
1814
1815         /* Do we support "hardware" checksums? */
1816         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1817                 /* This opens up the world of extra features. */
1818                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1819                 if (csum)
1820                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1821
1822                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1823                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1824                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1825                 }
1826                 /* Individual feature bits: what can host handle? */
1827                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1828                         dev->hw_features |= NETIF_F_TSO;
1829                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1830                         dev->hw_features |= NETIF_F_TSO6;
1831                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1832                         dev->hw_features |= NETIF_F_TSO_ECN;
1833                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1834                         dev->hw_features |= NETIF_F_UFO;
1835
1836                 dev->features |= NETIF_F_GSO_ROBUST;
1837
1838                 if (gso)
1839                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1840                 /* (!csum && gso) case will be fixed by register_netdev() */
1841         }
1842         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1843                 dev->features |= NETIF_F_RXCSUM;
1844
1845         dev->vlan_features = dev->features;
1846
1847         /* Configuration may specify what MAC to use.  Otherwise random. */
1848         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1849                 virtio_cread_bytes(vdev,
1850                                    offsetof(struct virtio_net_config, mac),
1851                                    dev->dev_addr, dev->addr_len);
1852         else
1853                 eth_hw_addr_random(dev);
1854
1855         /* Set up our device-specific information */
1856         vi = netdev_priv(dev);
1857         vi->dev = dev;
1858         vi->vdev = vdev;
1859         vdev->priv = vi;
1860         vi->stats = alloc_percpu(struct virtnet_stats);
1861         err = -ENOMEM;
1862         if (vi->stats == NULL)
1863                 goto free;
1864
1865         for_each_possible_cpu(i) {
1866                 struct virtnet_stats *virtnet_stats;
1867                 virtnet_stats = per_cpu_ptr(vi->stats, i);
1868                 u64_stats_init(&virtnet_stats->tx_syncp);
1869                 u64_stats_init(&virtnet_stats->rx_syncp);
1870         }
1871
1872         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1873
1874         /* If we can receive ANY GSO packets, we must allocate large ones. */
1875         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1876             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1877             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1878             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1879                 vi->big_packets = true;
1880
1881         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1882                 vi->mergeable_rx_bufs = true;
1883
1884         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
1885             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1886                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1887         else
1888                 vi->hdr_len = sizeof(struct virtio_net_hdr);
1889
1890         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
1891             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1892                 vi->any_header_sg = true;
1893
1894         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1895                 vi->has_cvq = true;
1896
1897         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
1898                 mtu = virtio_cread16(vdev,
1899                                      offsetof(struct virtio_net_config,
1900                                               mtu));
1901                 if (virtnet_change_mtu(dev, mtu))
1902                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
1903         }
1904
1905         if (vi->any_header_sg)
1906                 dev->needed_headroom = vi->hdr_len;
1907
1908         /* Use single tx/rx queue pair as default */
1909         vi->curr_queue_pairs = 1;
1910         vi->max_queue_pairs = max_queue_pairs;
1911
1912         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1913         err = init_vqs(vi);
1914         if (err)
1915                 goto free_stats;
1916
1917 #ifdef CONFIG_SYSFS
1918         if (vi->mergeable_rx_bufs)
1919                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1920 #endif
1921         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1922         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1923
1924         virtnet_init_settings(dev);
1925
1926         err = register_netdev(dev);
1927         if (err) {
1928                 pr_debug("virtio_net: registering device failed\n");
1929                 goto free_vqs;
1930         }
1931
1932         virtio_device_ready(vdev);
1933
1934         err = virtnet_cpu_notif_add(vi);
1935         if (err) {
1936                 pr_debug("virtio_net: registering cpu notifier failed\n");
1937                 goto free_unregister_netdev;
1938         }
1939
1940         /* Assume link up if device can't report link status,
1941            otherwise get link status from config. */
1942         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1943                 netif_carrier_off(dev);
1944                 schedule_work(&vi->config_work);
1945         } else {
1946                 vi->status = VIRTIO_NET_S_LINK_UP;
1947                 netif_carrier_on(dev);
1948         }
1949
1950         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1951                  dev->name, max_queue_pairs);
1952
1953         return 0;
1954
1955 free_unregister_netdev:
1956         vi->vdev->config->reset(vdev);
1957
1958         unregister_netdev(dev);
1959 free_vqs:
1960         cancel_delayed_work_sync(&vi->refill);
1961         free_receive_page_frags(vi);
1962         virtnet_del_vqs(vi);
1963 free_stats:
1964         free_percpu(vi->stats);
1965 free:
1966         free_netdev(dev);
1967         return err;
1968 }
1969
1970 static void remove_vq_common(struct virtnet_info *vi)
1971 {
1972         vi->vdev->config->reset(vi->vdev);
1973
1974         /* Free unused buffers in both send and recv, if any. */
1975         free_unused_bufs(vi);
1976
1977         free_receive_bufs(vi);
1978
1979         free_receive_page_frags(vi);
1980
1981         virtnet_del_vqs(vi);
1982 }
1983
1984 static void virtnet_remove(struct virtio_device *vdev)
1985 {
1986         struct virtnet_info *vi = vdev->priv;
1987
1988         virtnet_cpu_notif_remove(vi);
1989
1990         /* Make sure no work handler is accessing the device. */
1991         flush_work(&vi->config_work);
1992
1993         unregister_netdev(vi->dev);
1994
1995         remove_vq_common(vi);
1996
1997         free_percpu(vi->stats);
1998         free_netdev(vi->dev);
1999 }
2000
2001 #ifdef CONFIG_PM_SLEEP
2002 static int virtnet_freeze(struct virtio_device *vdev)
2003 {
2004         struct virtnet_info *vi = vdev->priv;
2005         int i;
2006
2007         virtnet_cpu_notif_remove(vi);
2008
2009         /* Make sure no work handler is accessing the device */
2010         flush_work(&vi->config_work);
2011
2012         netif_device_detach(vi->dev);
2013         cancel_delayed_work_sync(&vi->refill);
2014
2015         if (netif_running(vi->dev)) {
2016                 for (i = 0; i < vi->max_queue_pairs; i++)
2017                         napi_disable(&vi->rq[i].napi);
2018         }
2019
2020         remove_vq_common(vi);
2021
2022         return 0;
2023 }
2024
2025 static int virtnet_restore(struct virtio_device *vdev)
2026 {
2027         struct virtnet_info *vi = vdev->priv;
2028         int err, i;
2029
2030         err = init_vqs(vi);
2031         if (err)
2032                 return err;
2033
2034         virtio_device_ready(vdev);
2035
2036         if (netif_running(vi->dev)) {
2037                 for (i = 0; i < vi->curr_queue_pairs; i++)
2038                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2039                                 schedule_delayed_work(&vi->refill, 0);
2040
2041                 for (i = 0; i < vi->max_queue_pairs; i++)
2042                         virtnet_napi_enable(&vi->rq[i]);
2043         }
2044
2045         netif_device_attach(vi->dev);
2046
2047         rtnl_lock();
2048         virtnet_set_queues(vi, vi->curr_queue_pairs);
2049         rtnl_unlock();
2050
2051         err = virtnet_cpu_notif_add(vi);
2052         if (err)
2053                 return err;
2054
2055         return 0;
2056 }
2057 #endif
2058
2059 static struct virtio_device_id id_table[] = {
2060         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2061         { 0 },
2062 };
2063
2064 #define VIRTNET_FEATURES \
2065         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2066         VIRTIO_NET_F_MAC, \
2067         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2068         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2069         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2070         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2071         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2072         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2073         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2074         VIRTIO_NET_F_MTU
2075
2076 static unsigned int features[] = {
2077         VIRTNET_FEATURES,
2078 };
2079
2080 static unsigned int features_legacy[] = {
2081         VIRTNET_FEATURES,
2082         VIRTIO_NET_F_GSO,
2083         VIRTIO_F_ANY_LAYOUT,
2084 };
2085
2086 static struct virtio_driver virtio_net_driver = {
2087         .feature_table = features,
2088         .feature_table_size = ARRAY_SIZE(features),
2089         .feature_table_legacy = features_legacy,
2090         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2091         .driver.name =  KBUILD_MODNAME,
2092         .driver.owner = THIS_MODULE,
2093         .id_table =     id_table,
2094         .probe =        virtnet_probe,
2095         .remove =       virtnet_remove,
2096         .config_changed = virtnet_config_changed,
2097 #ifdef CONFIG_PM_SLEEP
2098         .freeze =       virtnet_freeze,
2099         .restore =      virtnet_restore,
2100 #endif
2101 };
2102
2103 static __init int virtio_net_driver_init(void)
2104 {
2105         int ret;
2106
2107         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "AP_VIRT_NET_ONLINE",
2108                                       virtnet_cpu_online,
2109                                       virtnet_cpu_down_prep);
2110         if (ret < 0)
2111                 goto out;
2112         virtionet_online = ret;
2113         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "VIRT_NET_DEAD",
2114                                       NULL, virtnet_cpu_dead);
2115         if (ret)
2116                 goto err_dead;
2117
2118         ret = register_virtio_driver(&virtio_net_driver);
2119         if (ret)
2120                 goto err_virtio;
2121         return 0;
2122 err_virtio:
2123         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2124 err_dead:
2125         cpuhp_remove_multi_state(virtionet_online);
2126 out:
2127         return ret;
2128 }
2129 module_init(virtio_net_driver_init);
2130
2131 static __exit void virtio_net_driver_exit(void)
2132 {
2133         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2134         cpuhp_remove_multi_state(virtionet_online);
2135         unregister_virtio_driver(&virtio_net_driver);
2136 }
2137 module_exit(virtio_net_driver_exit);
2138
2139 MODULE_DEVICE_TABLE(virtio, id_table);
2140 MODULE_DESCRIPTION("Virtio network driver");
2141 MODULE_LICENSE("GPL");