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/bpf.h>
  26 #include <linux/scatterlist.h>
  27 #include <linux/if_vlan.h>
  28 #include <linux/slab.h>
  29 #include <linux/cpu.h>
  30 #include <linux/average.h>
  31 #include <net/busy_poll.h>
  32
  33 static int napi_weight = NAPI_POLL_WEIGHT;
  34 module_param(napi_weight, int, 0444);
  35
  36 static bool csum = true, gso = true;
  37 module_param(csum, bool, 0444);
  38 module_param(gso, bool, 0444);
  39
  40 /* FIXME: MTU in config. */
  41 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  42 #define GOOD_COPY_LEN   128
  43
  44 /* RX packet size EWMA. The average packet size is used to determine the packet
  45  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  46  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  47  * term, transient changes in packet size.
  48  */
  49 DECLARE_EWMA(pkt_len, 1, 64)
  50
  51 /* Minimum alignment for mergeable packet buffers. */
  52 #define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, 256)
  53
  54 #define VIRTNET_DRIVER_VERSION "1.0.0"
  55
  56 struct virtnet_stats {
  57         struct u64_stats_sync tx_syncp;
  58         struct u64_stats_sync rx_syncp;
  59         u64 tx_bytes;
  60         u64 tx_packets;
  61
  62         u64 rx_bytes;
  63         u64 rx_packets;
  64 };
  65
  66 /* Internal representation of a send virtqueue */
  67 struct send_queue {
  68         /* Virtqueue associated with this send _queue */
  69         struct virtqueue *vq;
  70
  71         /* TX: fragments + linear part + virtio header */
  72         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  73
  74         /* Name of the send queue: output.$index */
  75         char name[40];
  76 };
  77
  78 /* Internal representation of a receive virtqueue */
  79 struct receive_queue {
  80         /* Virtqueue associated with this receive_queue */
  81         struct virtqueue *vq;
  82
  83         struct napi_struct napi;
  84
  85         struct bpf_prog __rcu *xdp_prog;
  86
  87         /* Chain pages by the private ptr. */
  88         struct page *pages;
  89
  90         /* Average packet length for mergeable receive buffers. */
  91         struct ewma_pkt_len mrg_avg_pkt_len;
  92
  93         /* Page frag for packet buffer allocation. */
  94         struct page_frag alloc_frag;
  95
  96         /* RX: fragments + linear part + virtio header */
  97         struct scatterlist sg[MAX_SKB_FRAGS + 2];
  98
  99         /* Name of this receive queue: input.$index */
 100         char name[40];
 101 };
 102
 103 struct virtnet_info {
 104         struct virtio_device *vdev;
 105         struct virtqueue *cvq;
 106         struct net_device *dev;
 107         struct send_queue *sq;
 108         struct receive_queue *rq;
 109         unsigned int status;
 110
 111         /* Max # of queue pairs supported by the device */
 112         u16 max_queue_pairs;
 113
 114         /* # of queue pairs currently used by the driver */
 115         u16 curr_queue_pairs;
 116
 117         /* # of XDP queue pairs currently used by the driver */
 118         u16 xdp_queue_pairs;
 119
 120         /* I like... big packets and I cannot lie! */
 121         bool big_packets;
 122
 123         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 124         bool mergeable_rx_bufs;
 125
 126         /* Has control virtqueue */
 127         bool has_cvq;
 128
 129         /* Host can handle any s/g split between our header and packet data */
 130         bool any_header_sg;
 131
 132         /* Packet virtio header size */
 133         u8 hdr_len;
 134
 135         /* Active statistics */
 136         struct virtnet_stats __percpu *stats;
 137
 138         /* Work struct for refilling if we run low on memory. */
 139         struct delayed_work refill;
 140
 141         /* Work struct for config space updates */
 142         struct work_struct config_work;
 143
 144         /* Does the affinity hint is set for virtqueues? */
 145         bool affinity_hint_set;
 146
 147         /* CPU hotplug instances for online & dead */
 148         struct hlist_node node;
 149         struct hlist_node node_dead;
 150
 151         /* Control VQ buffers: protected by the rtnl lock */
 152         struct virtio_net_ctrl_hdr ctrl_hdr;
 153         virtio_net_ctrl_ack ctrl_status;
 154         struct virtio_net_ctrl_mq ctrl_mq;
 155         u8 ctrl_promisc;
 156         u8 ctrl_allmulti;
 157         u16 ctrl_vid;
 158
 159         /* Ethtool settings */
 160         u8 duplex;
 161         u32 speed;
 162 };
 163
 164 struct padded_vnet_hdr {
 165         struct virtio_net_hdr_mrg_rxbuf hdr;
 166         /*
 167          * hdr is in a separate sg buffer, and data sg buffer shares same page
 168          * with this header sg. This padding makes next sg 16 byte aligned
 169          * after the header.
 170          */
 171         char padding[4];
 172 };
 173
 174 /* Converting between virtqueue no. and kernel tx/rx queue no.
 175  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 176  */
 177 static int vq2txq(struct virtqueue *vq)
 178 {
 179         return (vq->index - 1) / 2;
 180 }
 181
 182 static int txq2vq(int txq)
 183 {
 184         return txq * 2 + 1;
 185 }
 186
 187 static int vq2rxq(struct virtqueue *vq)
 188 {
 189         return vq->index / 2;
 190 }
 191
 192 static int rxq2vq(int rxq)
 193 {
 194         return rxq * 2;
 195 }
 196
 197 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 198 {
 199         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 200 }
 201
 202 /*
 203  * private is used to chain pages for big packets, put the whole
 204  * most recent used list in the beginning for reuse
 205  */
 206 static void give_pages(struct receive_queue *rq, struct page *page)
 207 {
 208         struct page *end;
 209
 210         /* Find end of list, sew whole thing into vi->rq.pages. */
 211         for (end = page; end->private; end = (struct page *)end->private);
 212         end->private = (unsigned long)rq->pages;
 213         rq->pages = page;
 214 }
 215
 216 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 217 {
 218         struct page *p = rq->pages;
 219
 220         if (p) {
 221                 rq->pages = (struct page *)p->private;
 222                 /* clear private here, it is used to chain pages */
 223                 p->private = 0;
 224         } else
 225                 p = alloc_page(gfp_mask);
 226         return p;
 227 }
 228
 229 static void skb_xmit_done(struct virtqueue *vq)
 230 {
 231         struct virtnet_info *vi = vq->vdev->priv;
 232
 233         /* Suppress further interrupts. */
 234         virtqueue_disable_cb(vq);
 235
 236         /* We were probably waiting for more output buffers. */
 237         netif_wake_subqueue(vi->dev, vq2txq(vq));
 238 }
 239
 240 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
 241 {
 242         unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
 243         return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
 244 }
 245
 246 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
 247 {
 248         return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
 249
 250 }
 251
 252 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
 253 {
 254         unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
 255         return (unsigned long)buf | (size - 1);
 256 }
 257
 258 /* Called from bottom half context */
 259 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 260                                    struct receive_queue *rq,
 261                                    struct page *page, unsigned int offset,
 262                                    unsigned int len, unsigned int truesize)
 263 {
 264         struct sk_buff *skb;
 265         struct virtio_net_hdr_mrg_rxbuf *hdr;
 266         unsigned int copy, hdr_len, hdr_padded_len;
 267         char *p;
 268
 269         p = page_address(page) + offset;
 270
 271         /* copy small packet so we can reuse these pages for small data */
 272         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 273         if (unlikely(!skb))
 274                 return NULL;
 275
 276         hdr = skb_vnet_hdr(skb);
 277
 278         hdr_len = vi->hdr_len;
 279         if (vi->mergeable_rx_bufs)
 280                 hdr_padded_len = sizeof *hdr;
 281         else
 282                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 283
 284         memcpy(hdr, p, hdr_len);
 285
 286         len -= hdr_len;
 287         offset += hdr_padded_len;
 288         p += hdr_padded_len;
 289
 290         copy = len;
 291         if (copy > skb_tailroom(skb))
 292                 copy = skb_tailroom(skb);
 293         memcpy(skb_put(skb, copy), p, copy);
 294
 295         len -= copy;
 296         offset += copy;
 297
 298         if (vi->mergeable_rx_bufs) {
 299                 if (len)
 300                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 301                 else
 302                         put_page(page);
 303                 return skb;
 304         }
 305
 306         /*
 307          * Verify that we can indeed put this data into a skb.
 308          * This is here to handle cases when the device erroneously
 309          * tries to receive more than is possible. This is usually
 310          * the case of a broken device.
 311          */
 312         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 313                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 314                 dev_kfree_skb(skb);
 315                 return NULL;
 316         }
 317         BUG_ON(offset >= PAGE_SIZE);
 318         while (len) {
 319                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 320                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 321                                 frag_size, truesize);
 322                 len -= frag_size;
 323                 page = (struct page *)page->private;
 324                 offset = 0;
 325         }
 326
 327         if (page)
 328                 give_pages(rq, page);
 329
 330         return skb;
 331 }
 332
 333 static void virtnet_xdp_xmit(struct virtnet_info *vi,
 334                              struct receive_queue *rq,
 335                              struct send_queue *sq,
 336                              struct xdp_buff *xdp)
 337 {
 338         struct page *page = virt_to_head_page(xdp->data);
 339         struct virtio_net_hdr_mrg_rxbuf *hdr;
 340         unsigned int num_sg, len;
 341         void *xdp_sent;
 342         int err;
 343
 344         /* Free up any pending old buffers before queueing new ones. */
 345         while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 346                 struct page *sent_page = virt_to_head_page(xdp_sent);
 347
 348                 if (vi->mergeable_rx_bufs)
 349                         put_page(sent_page);
 350                 else
 351                         give_pages(rq, sent_page);
 352         }
 353
 354         /* Zero header and leave csum up to XDP layers */
 355         hdr = xdp->data;
 356         memset(hdr, 0, vi->hdr_len);
 357
 358         num_sg = 1;
 359         sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
 360         err = virtqueue_add_outbuf(sq->vq, sq->sg, num_sg,
 361                                    xdp->data, GFP_ATOMIC);
 362         if (unlikely(err)) {
 363                 if (vi->mergeable_rx_bufs)
 364                         put_page(page);
 365                 else
 366                         give_pages(rq, page);
 367                 return; // On error abort to avoid unnecessary kick
 368         } else if (!vi->mergeable_rx_bufs) {
 369                 /* If not mergeable bufs must be big packets so cleanup pages */
 370                 give_pages(rq, (struct page *)page->private);
 371                 page->private = 0;
 372         }
 373
 374         virtqueue_kick(sq->vq);
 375 }
 376
 377 static u32 do_xdp_prog(struct virtnet_info *vi,
 378                        struct receive_queue *rq,
 379                        struct bpf_prog *xdp_prog,
 380                        struct page *page, int offset, int len)
 381 {
 382         int hdr_padded_len;
 383         struct xdp_buff xdp;
 384         unsigned int qp;
 385         u32 act;
 386         u8 *buf;
 387
 388         buf = page_address(page) + offset;
 389
 390         if (vi->mergeable_rx_bufs)
 391                 hdr_padded_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 392         else
 393                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 394
 395         xdp.data = buf + hdr_padded_len;
 396         xdp.data_end = xdp.data + (len - vi->hdr_len);
 397
 398         act = bpf_prog_run_xdp(xdp_prog, &xdp);
 399         switch (act) {
 400         case XDP_PASS:
 401                 return XDP_PASS;
 402         case XDP_TX:
 403                 qp = vi->curr_queue_pairs -
 404                         vi->xdp_queue_pairs +
 405                         smp_processor_id();
 406                 xdp.data = buf + (vi->mergeable_rx_bufs ? 0 : 4);
 407                 virtnet_xdp_xmit(vi, rq, &vi->sq[qp], &xdp);
 408                 return XDP_TX;
 409         default:
 410                 bpf_warn_invalid_xdp_action(act);
 411         case XDP_ABORTED:
 412         case XDP_DROP:
 413                 return XDP_DROP;
 414         }
 415 }
 416
 417 static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
 418 {
 419         struct sk_buff * skb = buf;
 420
 421         len -= vi->hdr_len;
 422         skb_trim(skb, len);
 423
 424         return skb;
 425 }
 426
 427 static struct sk_buff *receive_big(struct net_device *dev,
 428                                    struct virtnet_info *vi,
 429                                    struct receive_queue *rq,
 430                                    void *buf,
 431                                    unsigned int len)
 432 {
 433         struct bpf_prog *xdp_prog;
 434         struct page *page = buf;
 435         struct sk_buff *skb;
 436
 437         rcu_read_lock();
 438         xdp_prog = rcu_dereference(rq->xdp_prog);
 439         if (xdp_prog) {
 440                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 441                 u32 act;
 442
 443                 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
 444                         goto err_xdp;
 445                 act = do_xdp_prog(vi, rq, xdp_prog, page, 0, len);
 446                 switch (act) {
 447                 case XDP_PASS:
 448                         break;
 449                 case XDP_TX:
 450                         rcu_read_unlock();
 451                         goto xdp_xmit;
 452                 case XDP_DROP:
 453                 default:
 454                         goto err_xdp;
 455                 }
 456         }
 457         rcu_read_unlock();
 458
 459         skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 460         if (unlikely(!skb))
 461                 goto err;
 462
 463         return skb;
 464
 465 err_xdp:
 466         rcu_read_unlock();
 467 err:
 468         dev->stats.rx_dropped++;
 469         give_pages(rq, page);
 470 xdp_xmit:
 471         return NULL;
 472 }
 473
 474 static struct sk_buff *receive_mergeable(struct net_device *dev,
 475                                          struct virtnet_info *vi,
 476                                          struct receive_queue *rq,
 477                                          unsigned long ctx,
 478                                          unsigned int len)
 479 {
 480         void *buf = mergeable_ctx_to_buf_address(ctx);
 481         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 482         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 483         struct page *page = virt_to_head_page(buf);
 484         int offset = buf - page_address(page);
 485         struct sk_buff *head_skb, *curr_skb;
 486         struct bpf_prog *xdp_prog;
 487         unsigned int truesize;
 488
 489         head_skb = NULL;
 490
 491         rcu_read_lock();
 492         xdp_prog = rcu_dereference(rq->xdp_prog);
 493         if (xdp_prog) {
 494                 u32 act;
 495
 496                 /* No known backend devices should send packets with
 497                  * more than a single buffer when XDP conditions are
 498                  * met. However it is not strictly illegal so the case
 499                  * is handled as an exception and a warning is thrown.
 500                  */
 501                 if (unlikely(num_buf > 1)) {
 502                         bpf_warn_invalid_xdp_buffer();
 503                         goto err_xdp;
 504                 }
 505
 506                 /* Transient failure which in theory could occur if
 507                  * in-flight packets from before XDP was enabled reach
 508                  * the receive path after XDP is loaded. In practice I
 509                  * was not able to create this condition.
 510                  */
 511                 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
 512                         goto err_xdp;
 513
 514                 act = do_xdp_prog(vi, rq, xdp_prog, page, offset, len);
 515                 switch (act) {
 516                 case XDP_PASS:
 517                         break;
 518                 case XDP_TX:
 519                         rcu_read_unlock();
 520                         goto xdp_xmit;
 521                 case XDP_DROP:
 522                 default:
 523                         goto err_xdp;
 524                 }
 525         }
 526         rcu_read_unlock();
 527
 528         truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 529         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
 530         curr_skb = head_skb;
 531
 532         if (unlikely(!curr_skb))
 533                 goto err_skb;
 534         while (--num_buf) {
 535                 int num_skb_frags;
 536
 537                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 538                 if (unlikely(!ctx)) {
 539                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 540                                  dev->name, num_buf,
 541                                  virtio16_to_cpu(vi->vdev,
 542                                                  hdr->num_buffers));
 543                         dev->stats.rx_length_errors++;
 544                         goto err_buf;
 545                 }
 546
 547                 buf = mergeable_ctx_to_buf_address(ctx);
 548                 page = virt_to_head_page(buf);
 549
 550                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 551                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 552                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 553
 554                         if (unlikely(!nskb))
 555                                 goto err_skb;
 556                         if (curr_skb == head_skb)
 557                                 skb_shinfo(curr_skb)->frag_list = nskb;
 558                         else
 559                                 curr_skb->next = nskb;
 560                         curr_skb = nskb;
 561                         head_skb->truesize += nskb->truesize;
 562                         num_skb_frags = 0;
 563                 }
 564                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 565                 if (curr_skb != head_skb) {
 566                         head_skb->data_len += len;
 567                         head_skb->len += len;
 568                         head_skb->truesize += truesize;
 569                 }
 570                 offset = buf - page_address(page);
 571                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 572                         put_page(page);
 573                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 574                                              len, truesize);
 575                 } else {
 576                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 577                                         offset, len, truesize);
 578                 }
 579         }
 580
 581         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 582         return head_skb;
 583
 584 err_xdp:
 585         rcu_read_unlock();
 586 err_skb:
 587         put_page(page);
 588         while (--num_buf) {
 589                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 590                 if (unlikely(!ctx)) {
 591                         pr_debug("%s: rx error: %d buffers missing\n",
 592                                  dev->name, num_buf);
 593                         dev->stats.rx_length_errors++;
 594                         break;
 595                 }
 596                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 597                 put_page(page);
 598         }
 599 err_buf:
 600         dev->stats.rx_dropped++;
 601         dev_kfree_skb(head_skb);
 602 xdp_xmit:
 603         return NULL;
 604 }
 605
 606 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 607                         void *buf, unsigned int len)
 608 {
 609         struct net_device *dev = vi->dev;
 610         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 611         struct sk_buff *skb;
 612         struct virtio_net_hdr_mrg_rxbuf *hdr;
 613
 614         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 615                 pr_debug("%s: short packet %i\n", dev->name, len);
 616                 dev->stats.rx_length_errors++;
 617                 if (vi->mergeable_rx_bufs) {
 618                         unsigned long ctx = (unsigned long)buf;
 619                         void *base = mergeable_ctx_to_buf_address(ctx);
 620                         put_page(virt_to_head_page(base));
 621                 } else if (vi->big_packets) {
 622                         give_pages(rq, buf);
 623                 } else {
 624                         dev_kfree_skb(buf);
 625                 }
 626                 return;
 627         }
 628
 629         if (vi->mergeable_rx_bufs)
 630                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
 631         else if (vi->big_packets)
 632                 skb = receive_big(dev, vi, rq, buf, len);
 633         else
 634                 skb = receive_small(vi, buf, len);
 635
 636         if (unlikely(!skb))
 637                 return;
 638
 639         hdr = skb_vnet_hdr(skb);
 640
 641         u64_stats_update_begin(&stats->rx_syncp);
 642         stats->rx_bytes += skb->len;
 643         stats->rx_packets++;
 644         u64_stats_update_end(&stats->rx_syncp);
 645
 646         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 647                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 648
 649         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 650                                   virtio_is_little_endian(vi->vdev))) {
 651                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 652                                      dev->name, hdr->hdr.gso_type,
 653                                      hdr->hdr.gso_size);
 654                 goto frame_err;
 655         }
 656
 657         skb->protocol = eth_type_trans(skb, dev);
 658         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 659                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 660
 661         napi_gro_receive(&rq->napi, skb);
 662         return;
 663
 664 frame_err:
 665         dev->stats.rx_frame_errors++;
 666         dev_kfree_skb(skb);
 667 }
 668
 669 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 670                              gfp_t gfp)
 671 {
 672         struct sk_buff *skb;
 673         struct virtio_net_hdr_mrg_rxbuf *hdr;
 674         int err;
 675
 676         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 677         if (unlikely(!skb))
 678                 return -ENOMEM;
 679
 680         skb_put(skb, GOOD_PACKET_LEN);
 681
 682         hdr = skb_vnet_hdr(skb);
 683         sg_init_table(rq->sg, 2);
 684         sg_set_buf(rq->sg, hdr, vi->hdr_len);
 685         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 686
 687         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 688         if (err < 0)
 689                 dev_kfree_skb(skb);
 690
 691         return err;
 692 }
 693
 694 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 695                            gfp_t gfp)
 696 {
 697         struct page *first, *list = NULL;
 698         char *p;
 699         int i, err, offset;
 700
 701         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 702
 703         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 704         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 705                 first = get_a_page(rq, gfp);
 706                 if (!first) {
 707                         if (list)
 708                                 give_pages(rq, list);
 709                         return -ENOMEM;
 710                 }
 711                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 712
 713                 /* chain new page in list head to match sg */
 714                 first->private = (unsigned long)list;
 715                 list = first;
 716         }
 717
 718         first = get_a_page(rq, gfp);
 719         if (!first) {
 720                 give_pages(rq, list);
 721                 return -ENOMEM;
 722         }
 723         p = page_address(first);
 724
 725         /* rq->sg[0], rq->sg[1] share the same page */
 726         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 727         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 728
 729         /* rq->sg[1] for data packet, from offset */
 730         offset = sizeof(struct padded_vnet_hdr);
 731         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 732
 733         /* chain first in list head */
 734         first->private = (unsigned long)list;
 735         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 736                                   first, gfp);
 737         if (err < 0)
 738                 give_pages(rq, first);
 739
 740         return err;
 741 }
 742
 743 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
 744 {
 745         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 746         unsigned int len;
 747
 748         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
 749                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 750         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 751 }
 752
 753 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 754 {
 755         struct page_frag *alloc_frag = &rq->alloc_frag;
 756         char *buf;
 757         unsigned long ctx;
 758         int err;
 759         unsigned int len, hole;
 760
 761         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 762         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 763                 return -ENOMEM;
 764
 765         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 766         ctx = mergeable_buf_to_ctx(buf, len);
 767         get_page(alloc_frag->page);
 768         alloc_frag->offset += len;
 769         hole = alloc_frag->size - alloc_frag->offset;
 770         if (hole < len) {
 771                 /* To avoid internal fragmentation, if there is very likely not
 772                  * enough space for another buffer, add the remaining space to
 773                  * the current buffer. This extra space is not included in
 774                  * the truesize stored in ctx.
 775                  */
 776                 len += hole;
 777                 alloc_frag->offset += hole;
 778         }
 779
 780         sg_init_one(rq->sg, buf, len);
 781         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 782         if (err < 0)
 783                 put_page(virt_to_head_page(buf));
 784
 785         return err;
 786 }
 787
 788 /*
 789  * Returns false if we couldn't fill entirely (OOM).
 790  *
 791  * Normally run in the receive path, but can also be run from ndo_open
 792  * before we're receiving packets, or from refill_work which is
 793  * careful to disable receiving (using napi_disable).
 794  */
 795 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 796                           gfp_t gfp)
 797 {
 798         int err;
 799         bool oom;
 800
 801         gfp |= __GFP_COLD;
 802         do {
 803                 if (vi->mergeable_rx_bufs)
 804                         err = add_recvbuf_mergeable(rq, gfp);
 805                 else if (vi->big_packets)
 806                         err = add_recvbuf_big(vi, rq, gfp);
 807                 else
 808                         err = add_recvbuf_small(vi, rq, gfp);
 809
 810                 oom = err == -ENOMEM;
 811                 if (err)
 812                         break;
 813         } while (rq->vq->num_free);
 814         virtqueue_kick(rq->vq);
 815         return !oom;
 816 }
 817
 818 static void skb_recv_done(struct virtqueue *rvq)
 819 {
 820         struct virtnet_info *vi = rvq->vdev->priv;
 821         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 822
 823         /* Schedule NAPI, Suppress further interrupts if successful. */
 824         if (napi_schedule_prep(&rq->napi)) {
 825                 virtqueue_disable_cb(rvq);
 826                 __napi_schedule(&rq->napi);
 827         }
 828 }
 829
 830 static void virtnet_napi_enable(struct receive_queue *rq)
 831 {
 832         napi_enable(&rq->napi);
 833
 834         /* If all buffers were filled by other side before we napi_enabled, we
 835          * won't get another interrupt, so process any outstanding packets
 836          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 837          * We synchronize against interrupts via NAPI_STATE_SCHED */
 838         if (napi_schedule_prep(&rq->napi)) {
 839                 virtqueue_disable_cb(rq->vq);
 840                 local_bh_disable();
 841                 __napi_schedule(&rq->napi);
 842                 local_bh_enable();
 843         }
 844 }
 845
 846 static void refill_work(struct work_struct *work)
 847 {
 848         struct virtnet_info *vi =
 849                 container_of(work, struct virtnet_info, refill.work);
 850         bool still_empty;
 851         int i;
 852
 853         for (i = 0; i < vi->curr_queue_pairs; i++) {
 854                 struct receive_queue *rq = &vi->rq[i];
 855
 856                 napi_disable(&rq->napi);
 857                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
 858                 virtnet_napi_enable(rq);
 859
 860                 /* In theory, this can happen: if we don't get any buffers in
 861                  * we will *never* try to fill again.
 862                  */
 863                 if (still_empty)
 864                         schedule_delayed_work(&vi->refill, HZ/2);
 865         }
 866 }
 867
 868 static int virtnet_receive(struct receive_queue *rq, int budget)
 869 {
 870         struct virtnet_info *vi = rq->vq->vdev->priv;
 871         unsigned int len, received = 0;
 872         void *buf;
 873
 874         while (received < budget &&
 875                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 876                 receive_buf(vi, rq, buf, len);
 877                 received++;
 878         }
 879
 880         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 881                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 882                         schedule_delayed_work(&vi->refill, 0);
 883         }
 884
 885         return received;
 886 }
 887
 888 static int virtnet_poll(struct napi_struct *napi, int budget)
 889 {
 890         struct receive_queue *rq =
 891                 container_of(napi, struct receive_queue, napi);
 892         unsigned int r, received;
 893
 894         received = virtnet_receive(rq, budget);
 895
 896         /* Out of packets? */
 897         if (received < budget) {
 898                 r = virtqueue_enable_cb_prepare(rq->vq);
 899                 napi_complete_done(napi, received);
 900                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 901                     napi_schedule_prep(napi)) {
 902                         virtqueue_disable_cb(rq->vq);
 903                         __napi_schedule(napi);
 904                 }
 905         }
 906
 907         return received;
 908 }
 909
 910 #ifdef CONFIG_NET_RX_BUSY_POLL
 911 /* must be called with local_bh_disable()d */
 912 static int virtnet_busy_poll(struct napi_struct *napi)
 913 {
 914         struct receive_queue *rq =
 915                 container_of(napi, struct receive_queue, napi);
 916         struct virtnet_info *vi = rq->vq->vdev->priv;
 917         int r, received = 0, budget = 4;
 918
 919         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
 920                 return LL_FLUSH_FAILED;
 921
 922         if (!napi_schedule_prep(napi))
 923                 return LL_FLUSH_BUSY;
 924
 925         virtqueue_disable_cb(rq->vq);
 926
 927 again:
 928         received += virtnet_receive(rq, budget);
 929
 930         r = virtqueue_enable_cb_prepare(rq->vq);
 931         clear_bit(NAPI_STATE_SCHED, &napi->state);
 932         if (unlikely(virtqueue_poll(rq->vq, r)) &&
 933             napi_schedule_prep(napi)) {
 934                 virtqueue_disable_cb(rq->vq);
 935                 if (received < budget) {
 936                         budget -= received;
 937                         goto again;
 938                 } else {
 939                         __napi_schedule(napi);
 940                 }
 941         }
 942
 943         return received;
 944 }
 945 #endif  /* CONFIG_NET_RX_BUSY_POLL */
 946
 947 static int virtnet_open(struct net_device *dev)
 948 {
 949         struct virtnet_info *vi = netdev_priv(dev);
 950         int i;
 951
 952         for (i = 0; i < vi->max_queue_pairs; i++) {
 953                 if (i < vi->curr_queue_pairs)
 954                         /* Make sure we have some buffers: if oom use wq. */
 955                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
 956                                 schedule_delayed_work(&vi->refill, 0);
 957                 virtnet_napi_enable(&vi->rq[i]);
 958         }
 959
 960         return 0;
 961 }
 962
 963 static void free_old_xmit_skbs(struct send_queue *sq)
 964 {
 965         struct sk_buff *skb;
 966         unsigned int len;
 967         struct virtnet_info *vi = sq->vq->vdev->priv;
 968         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 969
 970         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 971                 pr_debug("Sent skb %p\n", skb);
 972
 973                 u64_stats_update_begin(&stats->tx_syncp);
 974                 stats->tx_bytes += skb->len;
 975                 stats->tx_packets++;
 976                 u64_stats_update_end(&stats->tx_syncp);
 977
 978                 dev_kfree_skb_any(skb);
 979         }
 980 }
 981
 982 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 983 {
 984         struct virtio_net_hdr_mrg_rxbuf *hdr;
 985         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
 986         struct virtnet_info *vi = sq->vq->vdev->priv;
 987         unsigned num_sg;
 988         unsigned hdr_len = vi->hdr_len;
 989         bool can_push;
 990
 991         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
 992
 993         can_push = vi->any_header_sg &&
 994                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
 995                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
 996         /* Even if we can, don't push here yet as this would skew
 997          * csum_start offset below. */
 998         if (can_push)
 999                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1000         else
1001                 hdr = skb_vnet_hdr(skb);
1002
1003         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1004                                     virtio_is_little_endian(vi->vdev)))
1005                 BUG();
1006
1007         if (vi->mergeable_rx_bufs)
1008                 hdr->num_buffers = 0;
1009
1010         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1011         if (can_push) {
1012                 __skb_push(skb, hdr_len);
1013                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1014                 /* Pull header back to avoid skew in tx bytes calculations. */
1015                 __skb_pull(skb, hdr_len);
1016         } else {
1017                 sg_set_buf(sq->sg, hdr, hdr_len);
1018                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
1019         }
1020         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1021 }
1022
1023 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1024 {
1025         struct virtnet_info *vi = netdev_priv(dev);
1026         int qnum = skb_get_queue_mapping(skb);
1027         struct send_queue *sq = &vi->sq[qnum];
1028         int err;
1029         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1030         bool kick = !skb->xmit_more;
1031
1032         /* Free up any pending old buffers before queueing new ones. */
1033         free_old_xmit_skbs(sq);
1034
1035         /* timestamp packet in software */
1036         skb_tx_timestamp(skb);
1037
1038         /* Try to transmit */
1039         err = xmit_skb(sq, skb);
1040
1041         /* This should not happen! */
1042         if (unlikely(err)) {
1043                 dev->stats.tx_fifo_errors++;
1044                 if (net_ratelimit())
1045                         dev_warn(&dev->dev,
1046                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1047                 dev->stats.tx_dropped++;
1048                 dev_kfree_skb_any(skb);
1049                 return NETDEV_TX_OK;
1050         }
1051
1052         /* Don't wait up for transmitted skbs to be freed. */
1053         skb_orphan(skb);
1054         nf_reset(skb);
1055
1056         /* If running out of space, stop queue to avoid getting packets that we
1057          * are then unable to transmit.
1058          * An alternative would be to force queuing layer to requeue the skb by
1059          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1060          * returned in a normal path of operation: it means that driver is not
1061          * maintaining the TX queue stop/start state properly, and causes
1062          * the stack to do a non-trivial amount of useless work.
1063          * Since most packets only take 1 or 2 ring slots, stopping the queue
1064          * early means 16 slots are typically wasted.
1065          */
1066         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1067                 netif_stop_subqueue(dev, qnum);
1068                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1069                         /* More just got used, free them then recheck. */
1070                         free_old_xmit_skbs(sq);
1071                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1072                                 netif_start_subqueue(dev, qnum);
1073                                 virtqueue_disable_cb(sq->vq);
1074                         }
1075                 }
1076         }
1077
1078         if (kick || netif_xmit_stopped(txq))
1079                 virtqueue_kick(sq->vq);
1080
1081         return NETDEV_TX_OK;
1082 }
1083
1084 /*
1085  * Send command via the control virtqueue and check status.  Commands
1086  * supported by the hypervisor, as indicated by feature bits, should
1087  * never fail unless improperly formatted.
1088  */
1089 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1090                                  struct scatterlist *out)
1091 {
1092         struct scatterlist *sgs[4], hdr, stat;
1093         unsigned out_num = 0, tmp;
1094
1095         /* Caller should know better */
1096         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1097
1098         vi->ctrl_status = ~0;
1099         vi->ctrl_hdr.class = class;
1100         vi->ctrl_hdr.cmd = cmd;
1101         /* Add header */
1102         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1103         sgs[out_num++] = &hdr;
1104
1105         if (out)
1106                 sgs[out_num++] = out;
1107
1108         /* Add return status. */
1109         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1110         sgs[out_num] = &stat;
1111
1112         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1113         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1114
1115         if (unlikely(!virtqueue_kick(vi->cvq)))
1116                 return vi->ctrl_status == VIRTIO_NET_OK;
1117
1118         /* Spin for a response, the kick causes an ioport write, trapping
1119          * into the hypervisor, so the request should be handled immediately.
1120          */
1121         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1122                !virtqueue_is_broken(vi->cvq))
1123                 cpu_relax();
1124
1125         return vi->ctrl_status == VIRTIO_NET_OK;
1126 }
1127
1128 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1129 {
1130         struct virtnet_info *vi = netdev_priv(dev);
1131         struct virtio_device *vdev = vi->vdev;
1132         int ret;
1133         struct sockaddr *addr;
1134         struct scatterlist sg;
1135
1136         addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1137         if (!addr)
1138                 return -ENOMEM;
1139         memcpy(addr, p, sizeof(*addr));
1140
1141         ret = eth_prepare_mac_addr_change(dev, addr);
1142         if (ret)
1143                 goto out;
1144
1145         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1146                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1147                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1148                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1149                         dev_warn(&vdev->dev,
1150                                  "Failed to set mac address by vq command.\n");
1151                         ret = -EINVAL;
1152                         goto out;
1153                 }
1154         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1155                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1156                 unsigned int i;
1157
1158                 /* Naturally, this has an atomicity problem. */
1159                 for (i = 0; i < dev->addr_len; i++)
1160                         virtio_cwrite8(vdev,
1161                                        offsetof(struct virtio_net_config, mac) +
1162                                        i, addr->sa_data[i]);
1163         }
1164
1165         eth_commit_mac_addr_change(dev, p);
1166         ret = 0;
1167
1168 out:
1169         kfree(addr);
1170         return ret;
1171 }
1172
1173 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1174                                                struct rtnl_link_stats64 *tot)
1175 {
1176         struct virtnet_info *vi = netdev_priv(dev);
1177         int cpu;
1178         unsigned int start;
1179
1180         for_each_possible_cpu(cpu) {
1181                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1182                 u64 tpackets, tbytes, rpackets, rbytes;
1183
1184                 do {
1185                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1186                         tpackets = stats->tx_packets;
1187                         tbytes   = stats->tx_bytes;
1188                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1189
1190                 do {
1191                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1192                         rpackets = stats->rx_packets;
1193                         rbytes   = stats->rx_bytes;
1194                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1195
1196                 tot->rx_packets += rpackets;
1197                 tot->tx_packets += tpackets;
1198                 tot->rx_bytes   += rbytes;
1199                 tot->tx_bytes   += tbytes;
1200         }
1201
1202         tot->tx_dropped = dev->stats.tx_dropped;
1203         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1204         tot->rx_dropped = dev->stats.rx_dropped;
1205         tot->rx_length_errors = dev->stats.rx_length_errors;
1206         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1207
1208         return tot;
1209 }
1210
1211 #ifdef CONFIG_NET_POLL_CONTROLLER
1212 static void virtnet_netpoll(struct net_device *dev)
1213 {
1214         struct virtnet_info *vi = netdev_priv(dev);
1215         int i;
1216
1217         for (i = 0; i < vi->curr_queue_pairs; i++)
1218                 napi_schedule(&vi->rq[i].napi);
1219 }
1220 #endif
1221
1222 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1223 {
1224         rtnl_lock();
1225         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1226                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1227                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1228         rtnl_unlock();
1229 }
1230
1231 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1232 {
1233         struct scatterlist sg;
1234         struct net_device *dev = vi->dev;
1235
1236         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1237                 return 0;
1238
1239         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1240         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1241
1242         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1243                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1244                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1245                          queue_pairs);
1246                 return -EINVAL;
1247         } else {
1248                 vi->curr_queue_pairs = queue_pairs;
1249                 /* virtnet_open() will refill when device is going to up. */
1250                 if (dev->flags & IFF_UP)
1251                         schedule_delayed_work(&vi->refill, 0);
1252         }
1253
1254         return 0;
1255 }
1256
1257 static int virtnet_close(struct net_device *dev)
1258 {
1259         struct virtnet_info *vi = netdev_priv(dev);
1260         int i;
1261
1262         /* Make sure refill_work doesn't re-enable napi! */
1263         cancel_delayed_work_sync(&vi->refill);
1264
1265         for (i = 0; i < vi->max_queue_pairs; i++)
1266                 napi_disable(&vi->rq[i].napi);
1267
1268         return 0;
1269 }
1270
1271 static void virtnet_set_rx_mode(struct net_device *dev)
1272 {
1273         struct virtnet_info *vi = netdev_priv(dev);
1274         struct scatterlist sg[2];
1275         struct virtio_net_ctrl_mac *mac_data;
1276         struct netdev_hw_addr *ha;
1277         int uc_count;
1278         int mc_count;
1279         void *buf;
1280         int i;
1281
1282         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1283         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1284                 return;
1285
1286         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1287         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1288
1289         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1290
1291         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1292                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1293                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1294                          vi->ctrl_promisc ? "en" : "dis");
1295
1296         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1297
1298         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1299                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1300                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1301                          vi->ctrl_allmulti ? "en" : "dis");
1302
1303         uc_count = netdev_uc_count(dev);
1304         mc_count = netdev_mc_count(dev);
1305         /* MAC filter - use one buffer for both lists */
1306         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1307                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1308         mac_data = buf;
1309         if (!buf)
1310                 return;
1311
1312         sg_init_table(sg, 2);
1313
1314         /* Store the unicast list and count in the front of the buffer */
1315         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1316         i = 0;
1317         netdev_for_each_uc_addr(ha, dev)
1318                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1319
1320         sg_set_buf(&sg[0], mac_data,
1321                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1322
1323         /* multicast list and count fill the end */
1324         mac_data = (void *)&mac_data->macs[uc_count][0];
1325
1326         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1327         i = 0;
1328         netdev_for_each_mc_addr(ha, dev)
1329                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1330
1331         sg_set_buf(&sg[1], mac_data,
1332                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1333
1334         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1335                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1336                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1337
1338         kfree(buf);
1339 }
1340
1341 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1342                                    __be16 proto, u16 vid)
1343 {
1344         struct virtnet_info *vi = netdev_priv(dev);
1345         struct scatterlist sg;
1346
1347         vi->ctrl_vid = vid;
1348         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1349
1350         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1351                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1352                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1353         return 0;
1354 }
1355
1356 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1357                                     __be16 proto, u16 vid)
1358 {
1359         struct virtnet_info *vi = netdev_priv(dev);
1360         struct scatterlist sg;
1361
1362         vi->ctrl_vid = vid;
1363         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1364
1365         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1366                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1367                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1368         return 0;
1369 }
1370
1371 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1372 {
1373         int i;
1374
1375         if (vi->affinity_hint_set) {
1376                 for (i = 0; i < vi->max_queue_pairs; i++) {
1377                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1378                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1379                 }
1380
1381                 vi->affinity_hint_set = false;
1382         }
1383 }
1384
1385 static void virtnet_set_affinity(struct virtnet_info *vi)
1386 {
1387         int i;
1388         int cpu;
1389
1390         /* In multiqueue mode, when the number of cpu is equal to the number of
1391          * queue pairs, we let the queue pairs to be private to one cpu by
1392          * setting the affinity hint to eliminate the contention.
1393          */
1394         if (vi->curr_queue_pairs == 1 ||
1395             vi->max_queue_pairs != num_online_cpus()) {
1396                 virtnet_clean_affinity(vi, -1);
1397                 return;
1398         }
1399
1400         i = 0;
1401         for_each_online_cpu(cpu) {
1402                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1403                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1404                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1405                 i++;
1406         }
1407
1408         vi->affinity_hint_set = true;
1409 }
1410
1411 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1412 {
1413         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1414                                                    node);
1415         virtnet_set_affinity(vi);
1416         return 0;
1417 }
1418
1419 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1420 {
1421         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1422                                                    node_dead);
1423         virtnet_set_affinity(vi);
1424         return 0;
1425 }
1426
1427 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1428 {
1429         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1430                                                    node);
1431
1432         virtnet_clean_affinity(vi, cpu);
1433         return 0;
1434 }
1435
1436 static enum cpuhp_state virtionet_online;
1437
1438 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1439 {
1440         int ret;
1441
1442         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1443         if (ret)
1444                 return ret;
1445         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1446                                                &vi->node_dead);
1447         if (!ret)
1448                 return ret;
1449         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1450         return ret;
1451 }
1452
1453 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1454 {
1455         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1456         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1457                                             &vi->node_dead);
1458 }
1459
1460 static void virtnet_get_ringparam(struct net_device *dev,
1461                                 struct ethtool_ringparam *ring)
1462 {
1463         struct virtnet_info *vi = netdev_priv(dev);
1464
1465         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1466         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1467         ring->rx_pending = ring->rx_max_pending;
1468         ring->tx_pending = ring->tx_max_pending;
1469 }
1470
1471
1472 static void virtnet_get_drvinfo(struct net_device *dev,
1473                                 struct ethtool_drvinfo *info)
1474 {
1475         struct virtnet_info *vi = netdev_priv(dev);
1476         struct virtio_device *vdev = vi->vdev;
1477
1478         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1479         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1480         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1481
1482 }
1483
1484 /* TODO: Eliminate OOO packets during switching */
1485 static int virtnet_set_channels(struct net_device *dev,
1486                                 struct ethtool_channels *channels)
1487 {
1488         struct virtnet_info *vi = netdev_priv(dev);
1489         u16 queue_pairs = channels->combined_count;
1490         int err;
1491
1492         /* We don't support separate rx/tx channels.
1493          * We don't allow setting 'other' channels.
1494          */
1495         if (channels->rx_count || channels->tx_count || channels->other_count)
1496                 return -EINVAL;
1497
1498         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1499                 return -EINVAL;
1500
1501         /* For now we don't support modifying channels while XDP is loaded
1502          * also when XDP is loaded all RX queues have XDP programs so we only
1503          * need to check a single RX queue.
1504          */
1505         if (vi->rq[0].xdp_prog)
1506                 return -EINVAL;
1507
1508         get_online_cpus();
1509         err = virtnet_set_queues(vi, queue_pairs);
1510         if (!err) {
1511                 netif_set_real_num_tx_queues(dev, queue_pairs);
1512                 netif_set_real_num_rx_queues(dev, queue_pairs);
1513
1514                 virtnet_set_affinity(vi);
1515         }
1516         put_online_cpus();
1517
1518         return err;
1519 }
1520
1521 static void virtnet_get_channels(struct net_device *dev,
1522                                  struct ethtool_channels *channels)
1523 {
1524         struct virtnet_info *vi = netdev_priv(dev);
1525
1526         channels->combined_count = vi->curr_queue_pairs;
1527         channels->max_combined = vi->max_queue_pairs;
1528         channels->max_other = 0;
1529         channels->rx_count = 0;
1530         channels->tx_count = 0;
1531         channels->other_count = 0;
1532 }
1533
1534 /* Check if the user is trying to change anything besides speed/duplex */
1535 static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1536 {
1537         struct ethtool_cmd diff1 = *cmd;
1538         struct ethtool_cmd diff2 = {};
1539
1540         /* cmd is always set so we need to clear it, validate the port type
1541          * and also without autonegotiation we can ignore advertising
1542          */
1543         ethtool_cmd_speed_set(&diff1, 0);
1544         diff2.port = PORT_OTHER;
1545         diff1.advertising = 0;
1546         diff1.duplex = 0;
1547         diff1.cmd = 0;
1548
1549         return !memcmp(&diff1, &diff2, sizeof(diff1));
1550 }
1551
1552 static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1553 {
1554         struct virtnet_info *vi = netdev_priv(dev);
1555         u32 speed;
1556
1557         speed = ethtool_cmd_speed(cmd);
1558         /* don't allow custom speed and duplex */
1559         if (!ethtool_validate_speed(speed) ||
1560             !ethtool_validate_duplex(cmd->duplex) ||
1561             !virtnet_validate_ethtool_cmd(cmd))
1562                 return -EINVAL;
1563         vi->speed = speed;
1564         vi->duplex = cmd->duplex;
1565
1566         return 0;
1567 }
1568
1569 static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1570 {
1571         struct virtnet_info *vi = netdev_priv(dev);
1572
1573         ethtool_cmd_speed_set(cmd, vi->speed);
1574         cmd->duplex = vi->duplex;
1575         cmd->port = PORT_OTHER;
1576
1577         return 0;
1578 }
1579
1580 static void virtnet_init_settings(struct net_device *dev)
1581 {
1582         struct virtnet_info *vi = netdev_priv(dev);
1583
1584         vi->speed = SPEED_UNKNOWN;
1585         vi->duplex = DUPLEX_UNKNOWN;
1586 }
1587
1588 static const struct ethtool_ops virtnet_ethtool_ops = {
1589         .get_drvinfo = virtnet_get_drvinfo,
1590         .get_link = ethtool_op_get_link,
1591         .get_ringparam = virtnet_get_ringparam,
1592         .set_channels = virtnet_set_channels,
1593         .get_channels = virtnet_get_channels,
1594         .get_ts_info = ethtool_op_get_ts_info,
1595         .get_settings = virtnet_get_settings,
1596         .set_settings = virtnet_set_settings,
1597 };
1598
1599 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog)
1600 {
1601         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1602         struct virtnet_info *vi = netdev_priv(dev);
1603         struct bpf_prog *old_prog;
1604         u16 xdp_qp = 0, curr_qp;
1605         int i, err;
1606
1607         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1608             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6)) {
1609                 netdev_warn(dev, "can't set XDP while host is implementing LRO, disable LRO first\n");
1610                 return -EOPNOTSUPP;
1611         }
1612
1613         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1614                 netdev_warn(dev, "XDP expects header/data in single page, any_header_sg required\n");
1615                 return -EINVAL;
1616         }
1617
1618         if (dev->mtu > max_sz) {
1619                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1620                 return -EINVAL;
1621         }
1622
1623         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1624         if (prog)
1625                 xdp_qp = nr_cpu_ids;
1626
1627         /* XDP requires extra queues for XDP_TX */
1628         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1629                 netdev_warn(dev, "request %i queues but max is %i\n",
1630                             curr_qp + xdp_qp, vi->max_queue_pairs);
1631                 return -ENOMEM;
1632         }
1633
1634         err = virtnet_set_queues(vi, curr_qp + xdp_qp);
1635         if (err) {
1636                 dev_warn(&dev->dev, "XDP Device queue allocation failure.\n");
1637                 return err;
1638         }
1639
1640         if (prog) {
1641                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1642                 if (IS_ERR(prog)) {
1643                         virtnet_set_queues(vi, curr_qp);
1644                         return PTR_ERR(prog);
1645                 }
1646         }
1647
1648         vi->xdp_queue_pairs = xdp_qp;
1649         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
1650
1651         for (i = 0; i < vi->max_queue_pairs; i++) {
1652                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1653                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
1654                 if (old_prog)
1655                         bpf_prog_put(old_prog);
1656         }
1657
1658         return 0;
1659 }
1660
1661 static bool virtnet_xdp_query(struct net_device *dev)
1662 {
1663         struct virtnet_info *vi = netdev_priv(dev);
1664         int i;
1665
1666         for (i = 0; i < vi->max_queue_pairs; i++) {
1667                 if (vi->rq[i].xdp_prog)
1668                         return true;
1669         }
1670         return false;
1671 }
1672
1673 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
1674 {
1675         switch (xdp->command) {
1676         case XDP_SETUP_PROG:
1677                 return virtnet_xdp_set(dev, xdp->prog);
1678         case XDP_QUERY_PROG:
1679                 xdp->prog_attached = virtnet_xdp_query(dev);
1680                 return 0;
1681         default:
1682                 return -EINVAL;
1683         }
1684 }
1685
1686 static const struct net_device_ops virtnet_netdev = {
1687         .ndo_open            = virtnet_open,
1688         .ndo_stop            = virtnet_close,
1689         .ndo_start_xmit      = start_xmit,
1690         .ndo_validate_addr   = eth_validate_addr,
1691         .ndo_set_mac_address = virtnet_set_mac_address,
1692         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1693         .ndo_get_stats64     = virtnet_stats,
1694         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1695         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1696 #ifdef CONFIG_NET_POLL_CONTROLLER
1697         .ndo_poll_controller = virtnet_netpoll,
1698 #endif
1699 #ifdef CONFIG_NET_RX_BUSY_POLL
1700         .ndo_busy_poll          = virtnet_busy_poll,
1701 #endif
1702         .ndo_xdp                = virtnet_xdp,
1703 };
1704
1705 static void virtnet_config_changed_work(struct work_struct *work)
1706 {
1707         struct virtnet_info *vi =
1708                 container_of(work, struct virtnet_info, config_work);
1709         u16 v;
1710
1711         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1712                                  struct virtio_net_config, status, &v) < 0)
1713                 return;
1714
1715         if (v & VIRTIO_NET_S_ANNOUNCE) {
1716                 netdev_notify_peers(vi->dev);
1717                 virtnet_ack_link_announce(vi);
1718         }
1719
1720         /* Ignore unknown (future) status bits */
1721         v &= VIRTIO_NET_S_LINK_UP;
1722
1723         if (vi->status == v)
1724                 return;
1725
1726         vi->status = v;
1727
1728         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1729                 netif_carrier_on(vi->dev);
1730                 netif_tx_wake_all_queues(vi->dev);
1731         } else {
1732                 netif_carrier_off(vi->dev);
1733                 netif_tx_stop_all_queues(vi->dev);
1734         }
1735 }
1736
1737 static void virtnet_config_changed(struct virtio_device *vdev)
1738 {
1739         struct virtnet_info *vi = vdev->priv;
1740
1741         schedule_work(&vi->config_work);
1742 }
1743
1744 static void virtnet_free_queues(struct virtnet_info *vi)
1745 {
1746         int i;
1747
1748         for (i = 0; i < vi->max_queue_pairs; i++) {
1749                 napi_hash_del(&vi->rq[i].napi);
1750                 netif_napi_del(&vi->rq[i].napi);
1751         }
1752
1753         /* We called napi_hash_del() before netif_napi_del(),
1754          * we need to respect an RCU grace period before freeing vi->rq
1755          */
1756         synchronize_net();
1757
1758         kfree(vi->rq);
1759         kfree(vi->sq);
1760 }
1761
1762 static void free_receive_bufs(struct virtnet_info *vi)
1763 {
1764         struct bpf_prog *old_prog;
1765         int i;
1766
1767         rtnl_lock();
1768         for (i = 0; i < vi->max_queue_pairs; i++) {
1769                 while (vi->rq[i].pages)
1770                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1771
1772                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1773                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
1774                 if (old_prog)
1775                         bpf_prog_put(old_prog);
1776         }
1777         rtnl_unlock();
1778 }
1779
1780 static void free_receive_page_frags(struct virtnet_info *vi)
1781 {
1782         int i;
1783         for (i = 0; i < vi->max_queue_pairs; i++)
1784                 if (vi->rq[i].alloc_frag.page)
1785                         put_page(vi->rq[i].alloc_frag.page);
1786 }
1787
1788 static bool is_xdp_queue(struct virtnet_info *vi, int q)
1789 {
1790         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1791                 return false;
1792         else if (q < vi->curr_queue_pairs)
1793                 return true;
1794         else
1795                 return false;
1796 }
1797
1798 static void free_unused_bufs(struct virtnet_info *vi)
1799 {
1800         void *buf;
1801         int i;
1802
1803         for (i = 0; i < vi->max_queue_pairs; i++) {
1804                 struct virtqueue *vq = vi->sq[i].vq;
1805                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1806                         if (!is_xdp_queue(vi, i))
1807                                 dev_kfree_skb(buf);
1808                         else
1809                                 put_page(virt_to_head_page(buf));
1810                 }
1811         }
1812
1813         for (i = 0; i < vi->max_queue_pairs; i++) {
1814                 struct virtqueue *vq = vi->rq[i].vq;
1815
1816                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1817                         if (vi->mergeable_rx_bufs) {
1818                                 unsigned long ctx = (unsigned long)buf;
1819                                 void *base = mergeable_ctx_to_buf_address(ctx);
1820                                 put_page(virt_to_head_page(base));
1821                         } else if (vi->big_packets) {
1822                                 give_pages(&vi->rq[i], buf);
1823                         } else {
1824                                 dev_kfree_skb(buf);
1825                         }
1826                 }
1827         }
1828 }
1829
1830 static void virtnet_del_vqs(struct virtnet_info *vi)
1831 {
1832         struct virtio_device *vdev = vi->vdev;
1833
1834         virtnet_clean_affinity(vi, -1);
1835
1836         vdev->config->del_vqs(vdev);
1837
1838         virtnet_free_queues(vi);
1839 }
1840
1841 static int virtnet_find_vqs(struct virtnet_info *vi)
1842 {
1843         vq_callback_t **callbacks;
1844         struct virtqueue **vqs;
1845         int ret = -ENOMEM;
1846         int i, total_vqs;
1847         const char **names;
1848
1849         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1850          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1851          * possible control vq.
1852          */
1853         total_vqs = vi->max_queue_pairs * 2 +
1854                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1855
1856         /* Allocate space for find_vqs parameters */
1857         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1858         if (!vqs)
1859                 goto err_vq;
1860         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1861         if (!callbacks)
1862                 goto err_callback;
1863         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1864         if (!names)
1865                 goto err_names;
1866
1867         /* Parameters for control virtqueue, if any */
1868         if (vi->has_cvq) {
1869                 callbacks[total_vqs - 1] = NULL;
1870                 names[total_vqs - 1] = "control";
1871         }
1872
1873         /* Allocate/initialize parameters for send/receive virtqueues */
1874         for (i = 0; i < vi->max_queue_pairs; i++) {
1875                 callbacks[rxq2vq(i)] = skb_recv_done;
1876                 callbacks[txq2vq(i)] = skb_xmit_done;
1877                 sprintf(vi->rq[i].name, "input.%d", i);
1878                 sprintf(vi->sq[i].name, "output.%d", i);
1879                 names[rxq2vq(i)] = vi->rq[i].name;
1880                 names[txq2vq(i)] = vi->sq[i].name;
1881         }
1882
1883         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1884                                          names);
1885         if (ret)
1886                 goto err_find;
1887
1888         if (vi->has_cvq) {
1889                 vi->cvq = vqs[total_vqs - 1];
1890                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1891                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1892         }
1893
1894         for (i = 0; i < vi->max_queue_pairs; i++) {
1895                 vi->rq[i].vq = vqs[rxq2vq(i)];
1896                 vi->sq[i].vq = vqs[txq2vq(i)];
1897         }
1898
1899         kfree(names);
1900         kfree(callbacks);
1901         kfree(vqs);
1902
1903         return 0;
1904
1905 err_find:
1906         kfree(names);
1907 err_names:
1908         kfree(callbacks);
1909 err_callback:
1910         kfree(vqs);
1911 err_vq:
1912         return ret;
1913 }
1914
1915 static int virtnet_alloc_queues(struct virtnet_info *vi)
1916 {
1917         int i;
1918
1919         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1920         if (!vi->sq)
1921                 goto err_sq;
1922         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1923         if (!vi->rq)
1924                 goto err_rq;
1925
1926         INIT_DELAYED_WORK(&vi->refill, refill_work);
1927         for (i = 0; i < vi->max_queue_pairs; i++) {
1928                 vi->rq[i].pages = NULL;
1929                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1930                                napi_weight);
1931
1932                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1933                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
1934                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1935         }
1936
1937         return 0;
1938
1939 err_rq:
1940         kfree(vi->sq);
1941 err_sq:
1942         return -ENOMEM;
1943 }
1944
1945 static int init_vqs(struct virtnet_info *vi)
1946 {
1947         int ret;
1948
1949         /* Allocate send & receive queues */
1950         ret = virtnet_alloc_queues(vi);
1951         if (ret)
1952                 goto err;
1953
1954         ret = virtnet_find_vqs(vi);
1955         if (ret)
1956                 goto err_free;
1957
1958         get_online_cpus();
1959         virtnet_set_affinity(vi);
1960         put_online_cpus();
1961
1962         return 0;
1963
1964 err_free:
1965         virtnet_free_queues(vi);
1966 err:
1967         return ret;
1968 }
1969
1970 #ifdef CONFIG_SYSFS
1971 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1972                 struct rx_queue_attribute *attribute, char *buf)
1973 {
1974         struct virtnet_info *vi = netdev_priv(queue->dev);
1975         unsigned int queue_index = get_netdev_rx_queue_index(queue);
1976         struct ewma_pkt_len *avg;
1977
1978         BUG_ON(queue_index >= vi->max_queue_pairs);
1979         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1980         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1981 }
1982
1983 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1984         __ATTR_RO(mergeable_rx_buffer_size);
1985
1986 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1987         &mergeable_rx_buffer_size_attribute.attr,
1988         NULL
1989 };
1990
1991 static const struct attribute_group virtio_net_mrg_rx_group = {
1992         .name = "virtio_net",
1993         .attrs = virtio_net_mrg_rx_attrs
1994 };
1995 #endif
1996
1997 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1998                                     unsigned int fbit,
1999                                     const char *fname, const char *dname)
2000 {
2001         if (!virtio_has_feature(vdev, fbit))
2002                 return false;
2003
2004         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2005                 fname, dname);
2006
2007         return true;
2008 }
2009
2010 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2011         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2012
2013 static bool virtnet_validate_features(struct virtio_device *vdev)
2014 {
2015         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2016             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2017                              "VIRTIO_NET_F_CTRL_VQ") ||
2018              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2019                              "VIRTIO_NET_F_CTRL_VQ") ||
2020              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2021                              "VIRTIO_NET_F_CTRL_VQ") ||
2022              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2023              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2024                              "VIRTIO_NET_F_CTRL_VQ"))) {
2025                 return false;
2026         }
2027
2028         return true;
2029 }
2030
2031 #define MIN_MTU ETH_MIN_MTU
2032 #define MAX_MTU ETH_MAX_MTU
2033
2034 static int virtnet_probe(struct virtio_device *vdev)
2035 {
2036         int i, err;
2037         struct net_device *dev;
2038         struct virtnet_info *vi;
2039         u16 max_queue_pairs;
2040         int mtu;
2041
2042         if (!vdev->config->get) {
2043                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2044                         __func__);
2045                 return -EINVAL;
2046         }
2047
2048         if (!virtnet_validate_features(vdev))
2049                 return -EINVAL;
2050
2051         /* Find if host supports multiqueue virtio_net device */
2052         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2053                                    struct virtio_net_config,
2054                                    max_virtqueue_pairs, &max_queue_pairs);
2055
2056         /* We need at least 2 queue's */
2057         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2058             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2059             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2060                 max_queue_pairs = 1;
2061
2062         /* Allocate ourselves a network device with room for our info */
2063         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2064         if (!dev)
2065                 return -ENOMEM;
2066
2067         /* Set up network device as normal. */
2068         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2069         dev->netdev_ops = &virtnet_netdev;
2070         dev->features = NETIF_F_HIGHDMA;
2071
2072         dev->ethtool_ops = &virtnet_ethtool_ops;
2073         SET_NETDEV_DEV(dev, &vdev->dev);
2074
2075         /* Do we support "hardware" checksums? */
2076         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2077                 /* This opens up the world of extra features. */
2078                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2079                 if (csum)
2080                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2081
2082                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2083                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
2084                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2085                 }
2086                 /* Individual feature bits: what can host handle? */
2087                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2088                         dev->hw_features |= NETIF_F_TSO;
2089                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2090                         dev->hw_features |= NETIF_F_TSO6;
2091                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2092                         dev->hw_features |= NETIF_F_TSO_ECN;
2093                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
2094                         dev->hw_features |= NETIF_F_UFO;
2095
2096                 dev->features |= NETIF_F_GSO_ROBUST;
2097
2098                 if (gso)
2099                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
2100                 /* (!csum && gso) case will be fixed by register_netdev() */
2101         }
2102         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2103                 dev->features |= NETIF_F_RXCSUM;
2104
2105         dev->vlan_features = dev->features;
2106
2107         /* MTU range: 68 - 65535 */
2108         dev->min_mtu = MIN_MTU;
2109         dev->max_mtu = MAX_MTU;
2110
2111         /* Configuration may specify what MAC to use.  Otherwise random. */
2112         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2113                 virtio_cread_bytes(vdev,
2114                                    offsetof(struct virtio_net_config, mac),
2115                                    dev->dev_addr, dev->addr_len);
2116         else
2117                 eth_hw_addr_random(dev);
2118
2119         /* Set up our device-specific information */
2120         vi = netdev_priv(dev);
2121         vi->dev = dev;
2122         vi->vdev = vdev;
2123         vdev->priv = vi;
2124         vi->stats = alloc_percpu(struct virtnet_stats);
2125         err = -ENOMEM;
2126         if (vi->stats == NULL)
2127                 goto free;
2128
2129         for_each_possible_cpu(i) {
2130                 struct virtnet_stats *virtnet_stats;
2131                 virtnet_stats = per_cpu_ptr(vi->stats, i);
2132                 u64_stats_init(&virtnet_stats->tx_syncp);
2133                 u64_stats_init(&virtnet_stats->rx_syncp);
2134         }
2135
2136         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2137
2138         /* If we can receive ANY GSO packets, we must allocate large ones. */
2139         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2140             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2141             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2142             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2143                 vi->big_packets = true;
2144
2145         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2146                 vi->mergeable_rx_bufs = true;
2147
2148         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2149             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2150                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2151         else
2152                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2153
2154         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2155             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2156                 vi->any_header_sg = true;
2157
2158         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2159                 vi->has_cvq = true;
2160
2161         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2162                 mtu = virtio_cread16(vdev,
2163                                      offsetof(struct virtio_net_config,
2164                                               mtu));
2165                 if (mtu < dev->min_mtu) {
2166                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2167                 } else {
2168                         dev->mtu = mtu;
2169                         dev->max_mtu = mtu;
2170                 }
2171         }
2172
2173         if (vi->any_header_sg)
2174                 dev->needed_headroom = vi->hdr_len;
2175
2176         /* Enable multiqueue by default */
2177         if (num_online_cpus() >= max_queue_pairs)
2178                 vi->curr_queue_pairs = max_queue_pairs;
2179         else
2180                 vi->curr_queue_pairs = num_online_cpus();
2181         vi->max_queue_pairs = max_queue_pairs;
2182
2183         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2184         err = init_vqs(vi);
2185         if (err)
2186                 goto free_stats;
2187
2188 #ifdef CONFIG_SYSFS
2189         if (vi->mergeable_rx_bufs)
2190                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2191 #endif
2192         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2193         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2194
2195         virtnet_init_settings(dev);
2196
2197         err = register_netdev(dev);
2198         if (err) {
2199                 pr_debug("virtio_net: registering device failed\n");
2200                 goto free_vqs;
2201         }
2202
2203         virtio_device_ready(vdev);
2204
2205         err = virtnet_cpu_notif_add(vi);
2206         if (err) {
2207                 pr_debug("virtio_net: registering cpu notifier failed\n");
2208                 goto free_unregister_netdev;
2209         }
2210
2211         rtnl_lock();
2212         virtnet_set_queues(vi, vi->curr_queue_pairs);
2213         rtnl_unlock();
2214
2215         /* Assume link up if device can't report link status,
2216            otherwise get link status from config. */
2217         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2218                 netif_carrier_off(dev);
2219                 schedule_work(&vi->config_work);
2220         } else {
2221                 vi->status = VIRTIO_NET_S_LINK_UP;
2222                 netif_carrier_on(dev);
2223         }
2224
2225         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2226                  dev->name, max_queue_pairs);
2227
2228         return 0;
2229
2230 free_unregister_netdev:
2231         vi->vdev->config->reset(vdev);
2232
2233         unregister_netdev(dev);
2234 free_vqs:
2235         cancel_delayed_work_sync(&vi->refill);
2236         free_receive_page_frags(vi);
2237         virtnet_del_vqs(vi);
2238 free_stats:
2239         free_percpu(vi->stats);
2240 free:
2241         free_netdev(dev);
2242         return err;
2243 }
2244
2245 static void remove_vq_common(struct virtnet_info *vi)
2246 {
2247         vi->vdev->config->reset(vi->vdev);
2248
2249         /* Free unused buffers in both send and recv, if any. */
2250         free_unused_bufs(vi);
2251
2252         free_receive_bufs(vi);
2253
2254         free_receive_page_frags(vi);
2255
2256         virtnet_del_vqs(vi);
2257 }
2258
2259 static void virtnet_remove(struct virtio_device *vdev)
2260 {
2261         struct virtnet_info *vi = vdev->priv;
2262
2263         virtnet_cpu_notif_remove(vi);
2264
2265         /* Make sure no work handler is accessing the device. */
2266         flush_work(&vi->config_work);
2267
2268         unregister_netdev(vi->dev);
2269
2270         remove_vq_common(vi);
2271
2272         free_percpu(vi->stats);
2273         free_netdev(vi->dev);
2274 }
2275
2276 #ifdef CONFIG_PM_SLEEP
2277 static int virtnet_freeze(struct virtio_device *vdev)
2278 {
2279         struct virtnet_info *vi = vdev->priv;
2280         int i;
2281
2282         virtnet_cpu_notif_remove(vi);
2283
2284         /* Make sure no work handler is accessing the device */
2285         flush_work(&vi->config_work);
2286
2287         netif_device_detach(vi->dev);
2288         cancel_delayed_work_sync(&vi->refill);
2289
2290         if (netif_running(vi->dev)) {
2291                 for (i = 0; i < vi->max_queue_pairs; i++)
2292                         napi_disable(&vi->rq[i].napi);
2293         }
2294
2295         remove_vq_common(vi);
2296
2297         return 0;
2298 }
2299
2300 static int virtnet_restore(struct virtio_device *vdev)
2301 {
2302         struct virtnet_info *vi = vdev->priv;
2303         int err, i;
2304
2305         err = init_vqs(vi);
2306         if (err)
2307                 return err;
2308
2309         virtio_device_ready(vdev);
2310
2311         if (netif_running(vi->dev)) {
2312                 for (i = 0; i < vi->curr_queue_pairs; i++)
2313                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2314                                 schedule_delayed_work(&vi->refill, 0);
2315
2316                 for (i = 0; i < vi->max_queue_pairs; i++)
2317                         virtnet_napi_enable(&vi->rq[i]);
2318         }
2319
2320         netif_device_attach(vi->dev);
2321
2322         rtnl_lock();
2323         virtnet_set_queues(vi, vi->curr_queue_pairs);
2324         rtnl_unlock();
2325
2326         err = virtnet_cpu_notif_add(vi);
2327         if (err)
2328                 return err;
2329
2330         return 0;
2331 }
2332 #endif
2333
2334 static struct virtio_device_id id_table[] = {
2335         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2336         { 0 },
2337 };
2338
2339 #define VIRTNET_FEATURES \
2340         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2341         VIRTIO_NET_F_MAC, \
2342         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2343         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2344         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2345         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2346         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2347         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2348         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2349         VIRTIO_NET_F_MTU
2350
2351 static unsigned int features[] = {
2352         VIRTNET_FEATURES,
2353 };
2354
2355 static unsigned int features_legacy[] = {
2356         VIRTNET_FEATURES,
2357         VIRTIO_NET_F_GSO,
2358         VIRTIO_F_ANY_LAYOUT,
2359 };
2360
2361 static struct virtio_driver virtio_net_driver = {
2362         .feature_table = features,
2363         .feature_table_size = ARRAY_SIZE(features),
2364         .feature_table_legacy = features_legacy,
2365         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2366         .driver.name =  KBUILD_MODNAME,
2367         .driver.owner = THIS_MODULE,
2368         .id_table =     id_table,
2369         .probe =        virtnet_probe,
2370         .remove =       virtnet_remove,
2371         .config_changed = virtnet_config_changed,
2372 #ifdef CONFIG_PM_SLEEP
2373         .freeze =       virtnet_freeze,
2374         .restore =      virtnet_restore,
2375 #endif
2376 };
2377
2378 static __init int virtio_net_driver_init(void)
2379 {
2380         int ret;
2381
2382         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "AP_VIRT_NET_ONLINE",
2383                                       virtnet_cpu_online,
2384                                       virtnet_cpu_down_prep);
2385         if (ret < 0)
2386                 goto out;
2387         virtionet_online = ret;
2388         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "VIRT_NET_DEAD",
2389                                       NULL, virtnet_cpu_dead);
2390         if (ret)
2391                 goto err_dead;
2392
2393         ret = register_virtio_driver(&virtio_net_driver);
2394         if (ret)
2395                 goto err_virtio;
2396         return 0;
2397 err_virtio:
2398         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2399 err_dead:
2400         cpuhp_remove_multi_state(virtionet_online);
2401 out:
2402         return ret;
2403 }
2404 module_init(virtio_net_driver_init);
2405
2406 static __exit void virtio_net_driver_exit(void)
2407 {
2408         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2409         cpuhp_remove_multi_state(virtionet_online);
2410         unregister_virtio_driver(&virtio_net_driver);
2411 }
2412 module_exit(virtio_net_driver_exit);
2413
2414 MODULE_DEVICE_TABLE(virtio, id_table);
2415 MODULE_DESCRIPTION("Virtio network driver");
2416 MODULE_LICENSE("GPL");