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                 put_page(sent_page);
 348         }
 349
 350         /* Zero header and leave csum up to XDP layers */
 351         hdr = xdp->data;
 352         memset(hdr, 0, vi->hdr_len);
 353
 354         num_sg = 1;
 355         sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
 356         err = virtqueue_add_outbuf(sq->vq, sq->sg, num_sg,
 357                                    xdp->data, GFP_ATOMIC);
 358         if (unlikely(err)) {
 359                 put_page(page);
 360                 return; // On error abort to avoid unnecessary kick
 361         }
 362
 363         virtqueue_kick(sq->vq);
 364 }
 365
 366 static u32 do_xdp_prog(struct virtnet_info *vi,
 367                        struct receive_queue *rq,
 368                        struct bpf_prog *xdp_prog,
 369                        struct page *page, int offset, int len)
 370 {
 371         int hdr_padded_len;
 372         struct xdp_buff xdp;
 373         unsigned int qp;
 374         u32 act;
 375         u8 *buf;
 376
 377         buf = page_address(page) + offset;
 378
 379         if (vi->mergeable_rx_bufs)
 380                 hdr_padded_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 381         else
 382                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 383
 384         xdp.data = buf + hdr_padded_len;
 385         xdp.data_end = xdp.data + (len - vi->hdr_len);
 386
 387         act = bpf_prog_run_xdp(xdp_prog, &xdp);
 388         switch (act) {
 389         case XDP_PASS:
 390                 return XDP_PASS;
 391         case XDP_TX:
 392                 qp = vi->curr_queue_pairs -
 393                         vi->xdp_queue_pairs +
 394                         smp_processor_id();
 395                 xdp.data = buf + (vi->mergeable_rx_bufs ? 0 : 4);
 396                 virtnet_xdp_xmit(vi, rq, &vi->sq[qp], &xdp);
 397                 return XDP_TX;
 398         default:
 399                 bpf_warn_invalid_xdp_action(act);
 400         case XDP_ABORTED:
 401         case XDP_DROP:
 402                 return XDP_DROP;
 403         }
 404 }
 405
 406 static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
 407 {
 408         struct sk_buff * skb = buf;
 409
 410         len -= vi->hdr_len;
 411         skb_trim(skb, len);
 412
 413         return skb;
 414 }
 415
 416 static struct sk_buff *receive_big(struct net_device *dev,
 417                                    struct virtnet_info *vi,
 418                                    struct receive_queue *rq,
 419                                    void *buf,
 420                                    unsigned int len)
 421 {
 422         struct page *page = buf;
 423         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
 424
 425         if (unlikely(!skb))
 426                 goto err;
 427
 428         return skb;
 429
 430 err:
 431         dev->stats.rx_dropped++;
 432         give_pages(rq, page);
 433         return NULL;
 434 }
 435
 436 /* The conditions to enable XDP should preclude the underlying device from
 437  * sending packets across multiple buffers (num_buf > 1). However per spec
 438  * it does not appear to be illegal to do so but rather just against convention.
 439  * So in order to avoid making a system unresponsive the packets are pushed
 440  * into a page and the XDP program is run. This will be extremely slow and we
 441  * push a warning to the user to fix this as soon as possible. Fixing this may
 442  * require resolving the underlying hardware to determine why multiple buffers
 443  * are being received or simply loading the XDP program in the ingress stack
 444  * after the skb is built because there is no advantage to running it here
 445  * anymore.
 446  */
 447 static struct page *xdp_linearize_page(struct receive_queue *rq,
 448                                        u16 *num_buf,
 449                                        struct page *p,
 450                                        int offset,
 451                                        unsigned int *len)
 452 {
 453         struct page *page = alloc_page(GFP_ATOMIC);
 454         unsigned int page_off = 0;
 455
 456         if (!page)
 457                 return NULL;
 458
 459         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 460         page_off += *len;
 461
 462         while (--*num_buf) {
 463                 unsigned int buflen;
 464                 unsigned long ctx;
 465                 void *buf;
 466                 int off;
 467
 468                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &buflen);
 469                 if (unlikely(!ctx))
 470                         goto err_buf;
 471
 472                 buf = mergeable_ctx_to_buf_address(ctx);
 473                 p = virt_to_head_page(buf);
 474                 off = buf - page_address(p);
 475
 476                 /* guard against a misconfigured or uncooperative backend that
 477                  * is sending packet larger than the MTU.
 478                  */
 479                 if ((page_off + buflen) > PAGE_SIZE) {
 480                         put_page(p);
 481                         goto err_buf;
 482                 }
 483
 484                 memcpy(page_address(page) + page_off,
 485                        page_address(p) + off, buflen);
 486                 page_off += buflen;
 487                 put_page(p);
 488         }
 489
 490         *len = page_off;
 491         return page;
 492 err_buf:
 493         __free_pages(page, 0);
 494         return NULL;
 495 }
 496
 497 static struct sk_buff *receive_mergeable(struct net_device *dev,
 498                                          struct virtnet_info *vi,
 499                                          struct receive_queue *rq,
 500                                          unsigned long ctx,
 501                                          unsigned int len)
 502 {
 503         void *buf = mergeable_ctx_to_buf_address(ctx);
 504         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 505         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 506         struct page *page = virt_to_head_page(buf);
 507         int offset = buf - page_address(page);
 508         struct sk_buff *head_skb, *curr_skb;
 509         struct bpf_prog *xdp_prog;
 510         unsigned int truesize;
 511
 512         head_skb = NULL;
 513
 514         rcu_read_lock();
 515         xdp_prog = rcu_dereference(rq->xdp_prog);
 516         if (xdp_prog) {
 517                 struct page *xdp_page;
 518                 u32 act;
 519
 520                 /* This happens when rx buffer size is underestimated */
 521                 if (unlikely(num_buf > 1)) {
 522                         /* linearize data for XDP */
 523                         xdp_page = xdp_linearize_page(rq, &num_buf,
 524                                                       page, offset, &len);
 525                         if (!xdp_page)
 526                                 goto err_xdp;
 527                         offset = 0;
 528                 } else {
 529                         xdp_page = page;
 530                 }
 531
 532                 /* Transient failure which in theory could occur if
 533                  * in-flight packets from before XDP was enabled reach
 534                  * the receive path after XDP is loaded. In practice I
 535                  * was not able to create this condition.
 536                  */
 537                 if (unlikely(hdr->hdr.gso_type))
 538                         goto err_xdp;
 539
 540                 act = do_xdp_prog(vi, rq, xdp_prog, xdp_page, offset, len);
 541                 switch (act) {
 542                 case XDP_PASS:
 543                         /* We can only create skb based on xdp_page. */
 544                         if (unlikely(xdp_page != page)) {
 545                                 rcu_read_unlock();
 546                                 put_page(page);
 547                                 head_skb = page_to_skb(vi, rq, xdp_page,
 548                                                        0, len, PAGE_SIZE);
 549                                 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 550                                 return head_skb;
 551                         }
 552                         break;
 553                 case XDP_TX:
 554                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 555                         if (unlikely(xdp_page != page))
 556                                 goto err_xdp;
 557                         rcu_read_unlock();
 558                         goto xdp_xmit;
 559                 case XDP_DROP:
 560                 default:
 561                         if (unlikely(xdp_page != page))
 562                                 __free_pages(xdp_page, 0);
 563                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
 564                         goto err_xdp;
 565                 }
 566         }
 567         rcu_read_unlock();
 568
 569         truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 570         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
 571         curr_skb = head_skb;
 572
 573         if (unlikely(!curr_skb))
 574                 goto err_skb;
 575         while (--num_buf) {
 576                 int num_skb_frags;
 577
 578                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 579                 if (unlikely(!ctx)) {
 580                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
 581                                  dev->name, num_buf,
 582                                  virtio16_to_cpu(vi->vdev,
 583                                                  hdr->num_buffers));
 584                         dev->stats.rx_length_errors++;
 585                         goto err_buf;
 586                 }
 587
 588                 buf = mergeable_ctx_to_buf_address(ctx);
 589                 page = virt_to_head_page(buf);
 590
 591                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
 592                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
 593                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
 594
 595                         if (unlikely(!nskb))
 596                                 goto err_skb;
 597                         if (curr_skb == head_skb)
 598                                 skb_shinfo(curr_skb)->frag_list = nskb;
 599                         else
 600                                 curr_skb->next = nskb;
 601                         curr_skb = nskb;
 602                         head_skb->truesize += nskb->truesize;
 603                         num_skb_frags = 0;
 604                 }
 605                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
 606                 if (curr_skb != head_skb) {
 607                         head_skb->data_len += len;
 608                         head_skb->len += len;
 609                         head_skb->truesize += truesize;
 610                 }
 611                 offset = buf - page_address(page);
 612                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
 613                         put_page(page);
 614                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
 615                                              len, truesize);
 616                 } else {
 617                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
 618                                         offset, len, truesize);
 619                 }
 620         }
 621
 622         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
 623         return head_skb;
 624
 625 err_xdp:
 626         rcu_read_unlock();
 627 err_skb:
 628         put_page(page);
 629         while (--num_buf) {
 630                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
 631                 if (unlikely(!ctx)) {
 632                         pr_debug("%s: rx error: %d buffers missing\n",
 633                                  dev->name, num_buf);
 634                         dev->stats.rx_length_errors++;
 635                         break;
 636                 }
 637                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
 638                 put_page(page);
 639         }
 640 err_buf:
 641         dev->stats.rx_dropped++;
 642         dev_kfree_skb(head_skb);
 643 xdp_xmit:
 644         return NULL;
 645 }
 646
 647 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
 648                         void *buf, unsigned int len)
 649 {
 650         struct net_device *dev = vi->dev;
 651         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
 652         struct sk_buff *skb;
 653         struct virtio_net_hdr_mrg_rxbuf *hdr;
 654
 655         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
 656                 pr_debug("%s: short packet %i\n", dev->name, len);
 657                 dev->stats.rx_length_errors++;
 658                 if (vi->mergeable_rx_bufs) {
 659                         unsigned long ctx = (unsigned long)buf;
 660                         void *base = mergeable_ctx_to_buf_address(ctx);
 661                         put_page(virt_to_head_page(base));
 662                 } else if (vi->big_packets) {
 663                         give_pages(rq, buf);
 664                 } else {
 665                         dev_kfree_skb(buf);
 666                 }
 667                 return;
 668         }
 669
 670         if (vi->mergeable_rx_bufs)
 671                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
 672         else if (vi->big_packets)
 673                 skb = receive_big(dev, vi, rq, buf, len);
 674         else
 675                 skb = receive_small(vi, buf, len);
 676
 677         if (unlikely(!skb))
 678                 return;
 679
 680         hdr = skb_vnet_hdr(skb);
 681
 682         u64_stats_update_begin(&stats->rx_syncp);
 683         stats->rx_bytes += skb->len;
 684         stats->rx_packets++;
 685         u64_stats_update_end(&stats->rx_syncp);
 686
 687         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
 688                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 689
 690         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
 691                                   virtio_is_little_endian(vi->vdev))) {
 692                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
 693                                      dev->name, hdr->hdr.gso_type,
 694                                      hdr->hdr.gso_size);
 695                 goto frame_err;
 696         }
 697
 698         skb->protocol = eth_type_trans(skb, dev);
 699         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
 700                  ntohs(skb->protocol), skb->len, skb->pkt_type);
 701
 702         napi_gro_receive(&rq->napi, skb);
 703         return;
 704
 705 frame_err:
 706         dev->stats.rx_frame_errors++;
 707         dev_kfree_skb(skb);
 708 }
 709
 710 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
 711                              gfp_t gfp)
 712 {
 713         struct sk_buff *skb;
 714         struct virtio_net_hdr_mrg_rxbuf *hdr;
 715         int err;
 716
 717         skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
 718         if (unlikely(!skb))
 719                 return -ENOMEM;
 720
 721         skb_put(skb, GOOD_PACKET_LEN);
 722
 723         hdr = skb_vnet_hdr(skb);
 724         sg_init_table(rq->sg, 2);
 725         sg_set_buf(rq->sg, hdr, vi->hdr_len);
 726         skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
 727
 728         err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
 729         if (err < 0)
 730                 dev_kfree_skb(skb);
 731
 732         return err;
 733 }
 734
 735 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
 736                            gfp_t gfp)
 737 {
 738         struct page *first, *list = NULL;
 739         char *p;
 740         int i, err, offset;
 741
 742         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
 743
 744         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
 745         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
 746                 first = get_a_page(rq, gfp);
 747                 if (!first) {
 748                         if (list)
 749                                 give_pages(rq, list);
 750                         return -ENOMEM;
 751                 }
 752                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
 753
 754                 /* chain new page in list head to match sg */
 755                 first->private = (unsigned long)list;
 756                 list = first;
 757         }
 758
 759         first = get_a_page(rq, gfp);
 760         if (!first) {
 761                 give_pages(rq, list);
 762                 return -ENOMEM;
 763         }
 764         p = page_address(first);
 765
 766         /* rq->sg[0], rq->sg[1] share the same page */
 767         /* a separated rq->sg[0] for header - required in case !any_header_sg */
 768         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
 769
 770         /* rq->sg[1] for data packet, from offset */
 771         offset = sizeof(struct padded_vnet_hdr);
 772         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
 773
 774         /* chain first in list head */
 775         first->private = (unsigned long)list;
 776         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
 777                                   first, gfp);
 778         if (err < 0)
 779                 give_pages(rq, first);
 780
 781         return err;
 782 }
 783
 784 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
 785 {
 786         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 787         unsigned int len;
 788
 789         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
 790                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
 791         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
 792 }
 793
 794 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
 795 {
 796         struct page_frag *alloc_frag = &rq->alloc_frag;
 797         char *buf;
 798         unsigned long ctx;
 799         int err;
 800         unsigned int len, hole;
 801
 802         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
 803         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
 804                 return -ENOMEM;
 805
 806         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
 807         ctx = mergeable_buf_to_ctx(buf, len);
 808         get_page(alloc_frag->page);
 809         alloc_frag->offset += len;
 810         hole = alloc_frag->size - alloc_frag->offset;
 811         if (hole < len) {
 812                 /* To avoid internal fragmentation, if there is very likely not
 813                  * enough space for another buffer, add the remaining space to
 814                  * the current buffer. This extra space is not included in
 815                  * the truesize stored in ctx.
 816                  */
 817                 len += hole;
 818                 alloc_frag->offset += hole;
 819         }
 820
 821         sg_init_one(rq->sg, buf, len);
 822         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
 823         if (err < 0)
 824                 put_page(virt_to_head_page(buf));
 825
 826         return err;
 827 }
 828
 829 /*
 830  * Returns false if we couldn't fill entirely (OOM).
 831  *
 832  * Normally run in the receive path, but can also be run from ndo_open
 833  * before we're receiving packets, or from refill_work which is
 834  * careful to disable receiving (using napi_disable).
 835  */
 836 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
 837                           gfp_t gfp)
 838 {
 839         int err;
 840         bool oom;
 841
 842         gfp |= __GFP_COLD;
 843         do {
 844                 if (vi->mergeable_rx_bufs)
 845                         err = add_recvbuf_mergeable(rq, gfp);
 846                 else if (vi->big_packets)
 847                         err = add_recvbuf_big(vi, rq, gfp);
 848                 else
 849                         err = add_recvbuf_small(vi, rq, gfp);
 850
 851                 oom = err == -ENOMEM;
 852                 if (err)
 853                         break;
 854         } while (rq->vq->num_free);
 855         virtqueue_kick(rq->vq);
 856         return !oom;
 857 }
 858
 859 static void skb_recv_done(struct virtqueue *rvq)
 860 {
 861         struct virtnet_info *vi = rvq->vdev->priv;
 862         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 863
 864         /* Schedule NAPI, Suppress further interrupts if successful. */
 865         if (napi_schedule_prep(&rq->napi)) {
 866                 virtqueue_disable_cb(rvq);
 867                 __napi_schedule(&rq->napi);
 868         }
 869 }
 870
 871 static void virtnet_napi_enable(struct receive_queue *rq)
 872 {
 873         napi_enable(&rq->napi);
 874
 875         /* If all buffers were filled by other side before we napi_enabled, we
 876          * won't get another interrupt, so process any outstanding packets
 877          * now.  virtnet_poll wants re-enable the queue, so we disable here.
 878          * We synchronize against interrupts via NAPI_STATE_SCHED */
 879         if (napi_schedule_prep(&rq->napi)) {
 880                 virtqueue_disable_cb(rq->vq);
 881                 local_bh_disable();
 882                 __napi_schedule(&rq->napi);
 883                 local_bh_enable();
 884         }
 885 }
 886
 887 static void refill_work(struct work_struct *work)
 888 {
 889         struct virtnet_info *vi =
 890                 container_of(work, struct virtnet_info, refill.work);
 891         bool still_empty;
 892         int i;
 893
 894         for (i = 0; i < vi->curr_queue_pairs; i++) {
 895                 struct receive_queue *rq = &vi->rq[i];
 896
 897                 napi_disable(&rq->napi);
 898                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
 899                 virtnet_napi_enable(rq);
 900
 901                 /* In theory, this can happen: if we don't get any buffers in
 902                  * we will *never* try to fill again.
 903                  */
 904                 if (still_empty)
 905                         schedule_delayed_work(&vi->refill, HZ/2);
 906         }
 907 }
 908
 909 static int virtnet_receive(struct receive_queue *rq, int budget)
 910 {
 911         struct virtnet_info *vi = rq->vq->vdev->priv;
 912         unsigned int len, received = 0;
 913         void *buf;
 914
 915         while (received < budget &&
 916                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
 917                 receive_buf(vi, rq, buf, len);
 918                 received++;
 919         }
 920
 921         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
 922                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 923                         schedule_delayed_work(&vi->refill, 0);
 924         }
 925
 926         return received;
 927 }
 928
 929 static int virtnet_poll(struct napi_struct *napi, int budget)
 930 {
 931         struct receive_queue *rq =
 932                 container_of(napi, struct receive_queue, napi);
 933         unsigned int r, received;
 934
 935         received = virtnet_receive(rq, budget);
 936
 937         /* Out of packets? */
 938         if (received < budget) {
 939                 r = virtqueue_enable_cb_prepare(rq->vq);
 940                 napi_complete_done(napi, received);
 941                 if (unlikely(virtqueue_poll(rq->vq, r)) &&
 942                     napi_schedule_prep(napi)) {
 943                         virtqueue_disable_cb(rq->vq);
 944                         __napi_schedule(napi);
 945                 }
 946         }
 947
 948         return received;
 949 }
 950
 951 #ifdef CONFIG_NET_RX_BUSY_POLL
 952 /* must be called with local_bh_disable()d */
 953 static int virtnet_busy_poll(struct napi_struct *napi)
 954 {
 955         struct receive_queue *rq =
 956                 container_of(napi, struct receive_queue, napi);
 957         struct virtnet_info *vi = rq->vq->vdev->priv;
 958         int r, received = 0, budget = 4;
 959
 960         if (!(vi->status & VIRTIO_NET_S_LINK_UP))
 961                 return LL_FLUSH_FAILED;
 962
 963         if (!napi_schedule_prep(napi))
 964                 return LL_FLUSH_BUSY;
 965
 966         virtqueue_disable_cb(rq->vq);
 967
 968 again:
 969         received += virtnet_receive(rq, budget);
 970
 971         r = virtqueue_enable_cb_prepare(rq->vq);
 972         clear_bit(NAPI_STATE_SCHED, &napi->state);
 973         if (unlikely(virtqueue_poll(rq->vq, r)) &&
 974             napi_schedule_prep(napi)) {
 975                 virtqueue_disable_cb(rq->vq);
 976                 if (received < budget) {
 977                         budget -= received;
 978                         goto again;
 979                 } else {
 980                         __napi_schedule(napi);
 981                 }
 982         }
 983
 984         return received;
 985 }
 986 #endif  /* CONFIG_NET_RX_BUSY_POLL */
 987
 988 static int virtnet_open(struct net_device *dev)
 989 {
 990         struct virtnet_info *vi = netdev_priv(dev);
 991         int i;
 992
 993         for (i = 0; i < vi->max_queue_pairs; i++) {
 994                 if (i < vi->curr_queue_pairs)
 995                         /* Make sure we have some buffers: if oom use wq. */
 996                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
 997                                 schedule_delayed_work(&vi->refill, 0);
 998                 virtnet_napi_enable(&vi->rq[i]);
 999         }
1000
1001         return 0;
1002 }
1003
1004 static void free_old_xmit_skbs(struct send_queue *sq)
1005 {
1006         struct sk_buff *skb;
1007         unsigned int len;
1008         struct virtnet_info *vi = sq->vq->vdev->priv;
1009         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1010
1011         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1012                 pr_debug("Sent skb %p\n", skb);
1013
1014                 u64_stats_update_begin(&stats->tx_syncp);
1015                 stats->tx_bytes += skb->len;
1016                 stats->tx_packets++;
1017                 u64_stats_update_end(&stats->tx_syncp);
1018
1019                 dev_kfree_skb_any(skb);
1020         }
1021 }
1022
1023 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1024 {
1025         struct virtio_net_hdr_mrg_rxbuf *hdr;
1026         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1027         struct virtnet_info *vi = sq->vq->vdev->priv;
1028         unsigned num_sg;
1029         unsigned hdr_len = vi->hdr_len;
1030         bool can_push;
1031
1032         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1033
1034         can_push = vi->any_header_sg &&
1035                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1036                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1037         /* Even if we can, don't push here yet as this would skew
1038          * csum_start offset below. */
1039         if (can_push)
1040                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1041         else
1042                 hdr = skb_vnet_hdr(skb);
1043
1044         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1045                                     virtio_is_little_endian(vi->vdev)))
1046                 BUG();
1047
1048         if (vi->mergeable_rx_bufs)
1049                 hdr->num_buffers = 0;
1050
1051         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1052         if (can_push) {
1053                 __skb_push(skb, hdr_len);
1054                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1055                 /* Pull header back to avoid skew in tx bytes calculations. */
1056                 __skb_pull(skb, hdr_len);
1057         } else {
1058                 sg_set_buf(sq->sg, hdr, hdr_len);
1059                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
1060         }
1061         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1062 }
1063
1064 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1065 {
1066         struct virtnet_info *vi = netdev_priv(dev);
1067         int qnum = skb_get_queue_mapping(skb);
1068         struct send_queue *sq = &vi->sq[qnum];
1069         int err;
1070         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1071         bool kick = !skb->xmit_more;
1072
1073         /* Free up any pending old buffers before queueing new ones. */
1074         free_old_xmit_skbs(sq);
1075
1076         /* timestamp packet in software */
1077         skb_tx_timestamp(skb);
1078
1079         /* Try to transmit */
1080         err = xmit_skb(sq, skb);
1081
1082         /* This should not happen! */
1083         if (unlikely(err)) {
1084                 dev->stats.tx_fifo_errors++;
1085                 if (net_ratelimit())
1086                         dev_warn(&dev->dev,
1087                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1088                 dev->stats.tx_dropped++;
1089                 dev_kfree_skb_any(skb);
1090                 return NETDEV_TX_OK;
1091         }
1092
1093         /* Don't wait up for transmitted skbs to be freed. */
1094         skb_orphan(skb);
1095         nf_reset(skb);
1096
1097         /* If running out of space, stop queue to avoid getting packets that we
1098          * are then unable to transmit.
1099          * An alternative would be to force queuing layer to requeue the skb by
1100          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1101          * returned in a normal path of operation: it means that driver is not
1102          * maintaining the TX queue stop/start state properly, and causes
1103          * the stack to do a non-trivial amount of useless work.
1104          * Since most packets only take 1 or 2 ring slots, stopping the queue
1105          * early means 16 slots are typically wasted.
1106          */
1107         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1108                 netif_stop_subqueue(dev, qnum);
1109                 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1110                         /* More just got used, free them then recheck. */
1111                         free_old_xmit_skbs(sq);
1112                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1113                                 netif_start_subqueue(dev, qnum);
1114                                 virtqueue_disable_cb(sq->vq);
1115                         }
1116                 }
1117         }
1118
1119         if (kick || netif_xmit_stopped(txq))
1120                 virtqueue_kick(sq->vq);
1121
1122         return NETDEV_TX_OK;
1123 }
1124
1125 /*
1126  * Send command via the control virtqueue and check status.  Commands
1127  * supported by the hypervisor, as indicated by feature bits, should
1128  * never fail unless improperly formatted.
1129  */
1130 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1131                                  struct scatterlist *out)
1132 {
1133         struct scatterlist *sgs[4], hdr, stat;
1134         unsigned out_num = 0, tmp;
1135
1136         /* Caller should know better */
1137         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1138
1139         vi->ctrl_status = ~0;
1140         vi->ctrl_hdr.class = class;
1141         vi->ctrl_hdr.cmd = cmd;
1142         /* Add header */
1143         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1144         sgs[out_num++] = &hdr;
1145
1146         if (out)
1147                 sgs[out_num++] = out;
1148
1149         /* Add return status. */
1150         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1151         sgs[out_num] = &stat;
1152
1153         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1154         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1155
1156         if (unlikely(!virtqueue_kick(vi->cvq)))
1157                 return vi->ctrl_status == VIRTIO_NET_OK;
1158
1159         /* Spin for a response, the kick causes an ioport write, trapping
1160          * into the hypervisor, so the request should be handled immediately.
1161          */
1162         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1163                !virtqueue_is_broken(vi->cvq))
1164                 cpu_relax();
1165
1166         return vi->ctrl_status == VIRTIO_NET_OK;
1167 }
1168
1169 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1170 {
1171         struct virtnet_info *vi = netdev_priv(dev);
1172         struct virtio_device *vdev = vi->vdev;
1173         int ret;
1174         struct sockaddr *addr;
1175         struct scatterlist sg;
1176
1177         addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1178         if (!addr)
1179                 return -ENOMEM;
1180         memcpy(addr, p, sizeof(*addr));
1181
1182         ret = eth_prepare_mac_addr_change(dev, addr);
1183         if (ret)
1184                 goto out;
1185
1186         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1187                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1188                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1189                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1190                         dev_warn(&vdev->dev,
1191                                  "Failed to set mac address by vq command.\n");
1192                         ret = -EINVAL;
1193                         goto out;
1194                 }
1195         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1196                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1197                 unsigned int i;
1198
1199                 /* Naturally, this has an atomicity problem. */
1200                 for (i = 0; i < dev->addr_len; i++)
1201                         virtio_cwrite8(vdev,
1202                                        offsetof(struct virtio_net_config, mac) +
1203                                        i, addr->sa_data[i]);
1204         }
1205
1206         eth_commit_mac_addr_change(dev, p);
1207         ret = 0;
1208
1209 out:
1210         kfree(addr);
1211         return ret;
1212 }
1213
1214 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1215                                                struct rtnl_link_stats64 *tot)
1216 {
1217         struct virtnet_info *vi = netdev_priv(dev);
1218         int cpu;
1219         unsigned int start;
1220
1221         for_each_possible_cpu(cpu) {
1222                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1223                 u64 tpackets, tbytes, rpackets, rbytes;
1224
1225                 do {
1226                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1227                         tpackets = stats->tx_packets;
1228                         tbytes   = stats->tx_bytes;
1229                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1230
1231                 do {
1232                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1233                         rpackets = stats->rx_packets;
1234                         rbytes   = stats->rx_bytes;
1235                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1236
1237                 tot->rx_packets += rpackets;
1238                 tot->tx_packets += tpackets;
1239                 tot->rx_bytes   += rbytes;
1240                 tot->tx_bytes   += tbytes;
1241         }
1242
1243         tot->tx_dropped = dev->stats.tx_dropped;
1244         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1245         tot->rx_dropped = dev->stats.rx_dropped;
1246         tot->rx_length_errors = dev->stats.rx_length_errors;
1247         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1248
1249         return tot;
1250 }
1251
1252 #ifdef CONFIG_NET_POLL_CONTROLLER
1253 static void virtnet_netpoll(struct net_device *dev)
1254 {
1255         struct virtnet_info *vi = netdev_priv(dev);
1256         int i;
1257
1258         for (i = 0; i < vi->curr_queue_pairs; i++)
1259                 napi_schedule(&vi->rq[i].napi);
1260 }
1261 #endif
1262
1263 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1264 {
1265         rtnl_lock();
1266         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1267                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1268                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1269         rtnl_unlock();
1270 }
1271
1272 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1273 {
1274         struct scatterlist sg;
1275         struct net_device *dev = vi->dev;
1276
1277         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1278                 return 0;
1279
1280         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1281         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1282
1283         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1284                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1285                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1286                          queue_pairs);
1287                 return -EINVAL;
1288         } else {
1289                 vi->curr_queue_pairs = queue_pairs;
1290                 /* virtnet_open() will refill when device is going to up. */
1291                 if (dev->flags & IFF_UP)
1292                         schedule_delayed_work(&vi->refill, 0);
1293         }
1294
1295         return 0;
1296 }
1297
1298 static int virtnet_close(struct net_device *dev)
1299 {
1300         struct virtnet_info *vi = netdev_priv(dev);
1301         int i;
1302
1303         /* Make sure refill_work doesn't re-enable napi! */
1304         cancel_delayed_work_sync(&vi->refill);
1305
1306         for (i = 0; i < vi->max_queue_pairs; i++)
1307                 napi_disable(&vi->rq[i].napi);
1308
1309         return 0;
1310 }
1311
1312 static void virtnet_set_rx_mode(struct net_device *dev)
1313 {
1314         struct virtnet_info *vi = netdev_priv(dev);
1315         struct scatterlist sg[2];
1316         struct virtio_net_ctrl_mac *mac_data;
1317         struct netdev_hw_addr *ha;
1318         int uc_count;
1319         int mc_count;
1320         void *buf;
1321         int i;
1322
1323         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1324         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1325                 return;
1326
1327         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1328         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1329
1330         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1331
1332         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1333                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1334                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1335                          vi->ctrl_promisc ? "en" : "dis");
1336
1337         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1338
1339         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1340                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1341                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1342                          vi->ctrl_allmulti ? "en" : "dis");
1343
1344         uc_count = netdev_uc_count(dev);
1345         mc_count = netdev_mc_count(dev);
1346         /* MAC filter - use one buffer for both lists */
1347         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1348                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1349         mac_data = buf;
1350         if (!buf)
1351                 return;
1352
1353         sg_init_table(sg, 2);
1354
1355         /* Store the unicast list and count in the front of the buffer */
1356         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1357         i = 0;
1358         netdev_for_each_uc_addr(ha, dev)
1359                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1360
1361         sg_set_buf(&sg[0], mac_data,
1362                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1363
1364         /* multicast list and count fill the end */
1365         mac_data = (void *)&mac_data->macs[uc_count][0];
1366
1367         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1368         i = 0;
1369         netdev_for_each_mc_addr(ha, dev)
1370                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1371
1372         sg_set_buf(&sg[1], mac_data,
1373                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1374
1375         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1376                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1377                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1378
1379         kfree(buf);
1380 }
1381
1382 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1383                                    __be16 proto, u16 vid)
1384 {
1385         struct virtnet_info *vi = netdev_priv(dev);
1386         struct scatterlist sg;
1387
1388         vi->ctrl_vid = vid;
1389         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1390
1391         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1392                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1393                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1394         return 0;
1395 }
1396
1397 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1398                                     __be16 proto, u16 vid)
1399 {
1400         struct virtnet_info *vi = netdev_priv(dev);
1401         struct scatterlist sg;
1402
1403         vi->ctrl_vid = vid;
1404         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1405
1406         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1407                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1408                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1409         return 0;
1410 }
1411
1412 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1413 {
1414         int i;
1415
1416         if (vi->affinity_hint_set) {
1417                 for (i = 0; i < vi->max_queue_pairs; i++) {
1418                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1419                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1420                 }
1421
1422                 vi->affinity_hint_set = false;
1423         }
1424 }
1425
1426 static void virtnet_set_affinity(struct virtnet_info *vi)
1427 {
1428         int i;
1429         int cpu;
1430
1431         /* In multiqueue mode, when the number of cpu is equal to the number of
1432          * queue pairs, we let the queue pairs to be private to one cpu by
1433          * setting the affinity hint to eliminate the contention.
1434          */
1435         if (vi->curr_queue_pairs == 1 ||
1436             vi->max_queue_pairs != num_online_cpus()) {
1437                 virtnet_clean_affinity(vi, -1);
1438                 return;
1439         }
1440
1441         i = 0;
1442         for_each_online_cpu(cpu) {
1443                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1444                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1445                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1446                 i++;
1447         }
1448
1449         vi->affinity_hint_set = true;
1450 }
1451
1452 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1453 {
1454         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1455                                                    node);
1456         virtnet_set_affinity(vi);
1457         return 0;
1458 }
1459
1460 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1461 {
1462         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1463                                                    node_dead);
1464         virtnet_set_affinity(vi);
1465         return 0;
1466 }
1467
1468 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1469 {
1470         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1471                                                    node);
1472
1473         virtnet_clean_affinity(vi, cpu);
1474         return 0;
1475 }
1476
1477 static enum cpuhp_state virtionet_online;
1478
1479 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1480 {
1481         int ret;
1482
1483         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1484         if (ret)
1485                 return ret;
1486         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1487                                                &vi->node_dead);
1488         if (!ret)
1489                 return ret;
1490         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1491         return ret;
1492 }
1493
1494 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1495 {
1496         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1497         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1498                                             &vi->node_dead);
1499 }
1500
1501 static void virtnet_get_ringparam(struct net_device *dev,
1502                                 struct ethtool_ringparam *ring)
1503 {
1504         struct virtnet_info *vi = netdev_priv(dev);
1505
1506         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1507         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1508         ring->rx_pending = ring->rx_max_pending;
1509         ring->tx_pending = ring->tx_max_pending;
1510 }
1511
1512
1513 static void virtnet_get_drvinfo(struct net_device *dev,
1514                                 struct ethtool_drvinfo *info)
1515 {
1516         struct virtnet_info *vi = netdev_priv(dev);
1517         struct virtio_device *vdev = vi->vdev;
1518
1519         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1520         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1521         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1522
1523 }
1524
1525 /* TODO: Eliminate OOO packets during switching */
1526 static int virtnet_set_channels(struct net_device *dev,
1527                                 struct ethtool_channels *channels)
1528 {
1529         struct virtnet_info *vi = netdev_priv(dev);
1530         u16 queue_pairs = channels->combined_count;
1531         int err;
1532
1533         /* We don't support separate rx/tx channels.
1534          * We don't allow setting 'other' channels.
1535          */
1536         if (channels->rx_count || channels->tx_count || channels->other_count)
1537                 return -EINVAL;
1538
1539         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1540                 return -EINVAL;
1541
1542         /* For now we don't support modifying channels while XDP is loaded
1543          * also when XDP is loaded all RX queues have XDP programs so we only
1544          * need to check a single RX queue.
1545          */
1546         if (vi->rq[0].xdp_prog)
1547                 return -EINVAL;
1548
1549         get_online_cpus();
1550         err = virtnet_set_queues(vi, queue_pairs);
1551         if (!err) {
1552                 netif_set_real_num_tx_queues(dev, queue_pairs);
1553                 netif_set_real_num_rx_queues(dev, queue_pairs);
1554
1555                 virtnet_set_affinity(vi);
1556         }
1557         put_online_cpus();
1558
1559         return err;
1560 }
1561
1562 static void virtnet_get_channels(struct net_device *dev,
1563                                  struct ethtool_channels *channels)
1564 {
1565         struct virtnet_info *vi = netdev_priv(dev);
1566
1567         channels->combined_count = vi->curr_queue_pairs;
1568         channels->max_combined = vi->max_queue_pairs;
1569         channels->max_other = 0;
1570         channels->rx_count = 0;
1571         channels->tx_count = 0;
1572         channels->other_count = 0;
1573 }
1574
1575 /* Check if the user is trying to change anything besides speed/duplex */
1576 static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1577 {
1578         struct ethtool_cmd diff1 = *cmd;
1579         struct ethtool_cmd diff2 = {};
1580
1581         /* cmd is always set so we need to clear it, validate the port type
1582          * and also without autonegotiation we can ignore advertising
1583          */
1584         ethtool_cmd_speed_set(&diff1, 0);
1585         diff2.port = PORT_OTHER;
1586         diff1.advertising = 0;
1587         diff1.duplex = 0;
1588         diff1.cmd = 0;
1589
1590         return !memcmp(&diff1, &diff2, sizeof(diff1));
1591 }
1592
1593 static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1594 {
1595         struct virtnet_info *vi = netdev_priv(dev);
1596         u32 speed;
1597
1598         speed = ethtool_cmd_speed(cmd);
1599         /* don't allow custom speed and duplex */
1600         if (!ethtool_validate_speed(speed) ||
1601             !ethtool_validate_duplex(cmd->duplex) ||
1602             !virtnet_validate_ethtool_cmd(cmd))
1603                 return -EINVAL;
1604         vi->speed = speed;
1605         vi->duplex = cmd->duplex;
1606
1607         return 0;
1608 }
1609
1610 static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1611 {
1612         struct virtnet_info *vi = netdev_priv(dev);
1613
1614         ethtool_cmd_speed_set(cmd, vi->speed);
1615         cmd->duplex = vi->duplex;
1616         cmd->port = PORT_OTHER;
1617
1618         return 0;
1619 }
1620
1621 static void virtnet_init_settings(struct net_device *dev)
1622 {
1623         struct virtnet_info *vi = netdev_priv(dev);
1624
1625         vi->speed = SPEED_UNKNOWN;
1626         vi->duplex = DUPLEX_UNKNOWN;
1627 }
1628
1629 static const struct ethtool_ops virtnet_ethtool_ops = {
1630         .get_drvinfo = virtnet_get_drvinfo,
1631         .get_link = ethtool_op_get_link,
1632         .get_ringparam = virtnet_get_ringparam,
1633         .set_channels = virtnet_set_channels,
1634         .get_channels = virtnet_get_channels,
1635         .get_ts_info = ethtool_op_get_ts_info,
1636         .get_settings = virtnet_get_settings,
1637         .set_settings = virtnet_set_settings,
1638 };
1639
1640 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog)
1641 {
1642         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1643         struct virtnet_info *vi = netdev_priv(dev);
1644         struct bpf_prog *old_prog;
1645         u16 xdp_qp = 0, curr_qp;
1646         int i, err;
1647
1648         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1649             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1650             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
1651             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO)) {
1652                 netdev_warn(dev, "can't set XDP while host is implementing LRO, disable LRO first\n");
1653                 return -EOPNOTSUPP;
1654         }
1655
1656         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1657                 netdev_warn(dev, "XDP expects header/data in single page, any_header_sg required\n");
1658                 return -EINVAL;
1659         }
1660
1661         if (dev->mtu > max_sz) {
1662                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1663                 return -EINVAL;
1664         }
1665
1666         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1667         if (prog)
1668                 xdp_qp = nr_cpu_ids;
1669
1670         /* XDP requires extra queues for XDP_TX */
1671         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1672                 netdev_warn(dev, "request %i queues but max is %i\n",
1673                             curr_qp + xdp_qp, vi->max_queue_pairs);
1674                 return -ENOMEM;
1675         }
1676
1677         err = virtnet_set_queues(vi, curr_qp + xdp_qp);
1678         if (err) {
1679                 dev_warn(&dev->dev, "XDP Device queue allocation failure.\n");
1680                 return err;
1681         }
1682
1683         if (prog) {
1684                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1685                 if (IS_ERR(prog)) {
1686                         virtnet_set_queues(vi, curr_qp);
1687                         return PTR_ERR(prog);
1688                 }
1689         }
1690
1691         vi->xdp_queue_pairs = xdp_qp;
1692         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
1693
1694         for (i = 0; i < vi->max_queue_pairs; i++) {
1695                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1696                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
1697                 if (old_prog)
1698                         bpf_prog_put(old_prog);
1699         }
1700
1701         return 0;
1702 }
1703
1704 static bool virtnet_xdp_query(struct net_device *dev)
1705 {
1706         struct virtnet_info *vi = netdev_priv(dev);
1707         int i;
1708
1709         for (i = 0; i < vi->max_queue_pairs; i++) {
1710                 if (vi->rq[i].xdp_prog)
1711                         return true;
1712         }
1713         return false;
1714 }
1715
1716 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
1717 {
1718         switch (xdp->command) {
1719         case XDP_SETUP_PROG:
1720                 return virtnet_xdp_set(dev, xdp->prog);
1721         case XDP_QUERY_PROG:
1722                 xdp->prog_attached = virtnet_xdp_query(dev);
1723                 return 0;
1724         default:
1725                 return -EINVAL;
1726         }
1727 }
1728
1729 static const struct net_device_ops virtnet_netdev = {
1730         .ndo_open            = virtnet_open,
1731         .ndo_stop            = virtnet_close,
1732         .ndo_start_xmit      = start_xmit,
1733         .ndo_validate_addr   = eth_validate_addr,
1734         .ndo_set_mac_address = virtnet_set_mac_address,
1735         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1736         .ndo_get_stats64     = virtnet_stats,
1737         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1738         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1739 #ifdef CONFIG_NET_POLL_CONTROLLER
1740         .ndo_poll_controller = virtnet_netpoll,
1741 #endif
1742 #ifdef CONFIG_NET_RX_BUSY_POLL
1743         .ndo_busy_poll          = virtnet_busy_poll,
1744 #endif
1745         .ndo_xdp                = virtnet_xdp,
1746 };
1747
1748 static void virtnet_config_changed_work(struct work_struct *work)
1749 {
1750         struct virtnet_info *vi =
1751                 container_of(work, struct virtnet_info, config_work);
1752         u16 v;
1753
1754         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1755                                  struct virtio_net_config, status, &v) < 0)
1756                 return;
1757
1758         if (v & VIRTIO_NET_S_ANNOUNCE) {
1759                 netdev_notify_peers(vi->dev);
1760                 virtnet_ack_link_announce(vi);
1761         }
1762
1763         /* Ignore unknown (future) status bits */
1764         v &= VIRTIO_NET_S_LINK_UP;
1765
1766         if (vi->status == v)
1767                 return;
1768
1769         vi->status = v;
1770
1771         if (vi->status & VIRTIO_NET_S_LINK_UP) {
1772                 netif_carrier_on(vi->dev);
1773                 netif_tx_wake_all_queues(vi->dev);
1774         } else {
1775                 netif_carrier_off(vi->dev);
1776                 netif_tx_stop_all_queues(vi->dev);
1777         }
1778 }
1779
1780 static void virtnet_config_changed(struct virtio_device *vdev)
1781 {
1782         struct virtnet_info *vi = vdev->priv;
1783
1784         schedule_work(&vi->config_work);
1785 }
1786
1787 static void virtnet_free_queues(struct virtnet_info *vi)
1788 {
1789         int i;
1790
1791         for (i = 0; i < vi->max_queue_pairs; i++) {
1792                 napi_hash_del(&vi->rq[i].napi);
1793                 netif_napi_del(&vi->rq[i].napi);
1794         }
1795
1796         /* We called napi_hash_del() before netif_napi_del(),
1797          * we need to respect an RCU grace period before freeing vi->rq
1798          */
1799         synchronize_net();
1800
1801         kfree(vi->rq);
1802         kfree(vi->sq);
1803 }
1804
1805 static void free_receive_bufs(struct virtnet_info *vi)
1806 {
1807         struct bpf_prog *old_prog;
1808         int i;
1809
1810         rtnl_lock();
1811         for (i = 0; i < vi->max_queue_pairs; i++) {
1812                 while (vi->rq[i].pages)
1813                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1814
1815                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1816                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
1817                 if (old_prog)
1818                         bpf_prog_put(old_prog);
1819         }
1820         rtnl_unlock();
1821 }
1822
1823 static void free_receive_page_frags(struct virtnet_info *vi)
1824 {
1825         int i;
1826         for (i = 0; i < vi->max_queue_pairs; i++)
1827                 if (vi->rq[i].alloc_frag.page)
1828                         put_page(vi->rq[i].alloc_frag.page);
1829 }
1830
1831 static bool is_xdp_queue(struct virtnet_info *vi, int q)
1832 {
1833         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1834                 return false;
1835         else if (q < vi->curr_queue_pairs)
1836                 return true;
1837         else
1838                 return false;
1839 }
1840
1841 static void free_unused_bufs(struct virtnet_info *vi)
1842 {
1843         void *buf;
1844         int i;
1845
1846         for (i = 0; i < vi->max_queue_pairs; i++) {
1847                 struct virtqueue *vq = vi->sq[i].vq;
1848                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1849                         if (!is_xdp_queue(vi, i))
1850                                 dev_kfree_skb(buf);
1851                         else
1852                                 put_page(virt_to_head_page(buf));
1853                 }
1854         }
1855
1856         for (i = 0; i < vi->max_queue_pairs; i++) {
1857                 struct virtqueue *vq = vi->rq[i].vq;
1858
1859                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1860                         if (vi->mergeable_rx_bufs) {
1861                                 unsigned long ctx = (unsigned long)buf;
1862                                 void *base = mergeable_ctx_to_buf_address(ctx);
1863                                 put_page(virt_to_head_page(base));
1864                         } else if (vi->big_packets) {
1865                                 give_pages(&vi->rq[i], buf);
1866                         } else {
1867                                 dev_kfree_skb(buf);
1868                         }
1869                 }
1870         }
1871 }
1872
1873 static void virtnet_del_vqs(struct virtnet_info *vi)
1874 {
1875         struct virtio_device *vdev = vi->vdev;
1876
1877         virtnet_clean_affinity(vi, -1);
1878
1879         vdev->config->del_vqs(vdev);
1880
1881         virtnet_free_queues(vi);
1882 }
1883
1884 static int virtnet_find_vqs(struct virtnet_info *vi)
1885 {
1886         vq_callback_t **callbacks;
1887         struct virtqueue **vqs;
1888         int ret = -ENOMEM;
1889         int i, total_vqs;
1890         const char **names;
1891
1892         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1893          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1894          * possible control vq.
1895          */
1896         total_vqs = vi->max_queue_pairs * 2 +
1897                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1898
1899         /* Allocate space for find_vqs parameters */
1900         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1901         if (!vqs)
1902                 goto err_vq;
1903         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1904         if (!callbacks)
1905                 goto err_callback;
1906         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1907         if (!names)
1908                 goto err_names;
1909
1910         /* Parameters for control virtqueue, if any */
1911         if (vi->has_cvq) {
1912                 callbacks[total_vqs - 1] = NULL;
1913                 names[total_vqs - 1] = "control";
1914         }
1915
1916         /* Allocate/initialize parameters for send/receive virtqueues */
1917         for (i = 0; i < vi->max_queue_pairs; i++) {
1918                 callbacks[rxq2vq(i)] = skb_recv_done;
1919                 callbacks[txq2vq(i)] = skb_xmit_done;
1920                 sprintf(vi->rq[i].name, "input.%d", i);
1921                 sprintf(vi->sq[i].name, "output.%d", i);
1922                 names[rxq2vq(i)] = vi->rq[i].name;
1923                 names[txq2vq(i)] = vi->sq[i].name;
1924         }
1925
1926         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1927                                          names);
1928         if (ret)
1929                 goto err_find;
1930
1931         if (vi->has_cvq) {
1932                 vi->cvq = vqs[total_vqs - 1];
1933                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1934                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1935         }
1936
1937         for (i = 0; i < vi->max_queue_pairs; i++) {
1938                 vi->rq[i].vq = vqs[rxq2vq(i)];
1939                 vi->sq[i].vq = vqs[txq2vq(i)];
1940         }
1941
1942         kfree(names);
1943         kfree(callbacks);
1944         kfree(vqs);
1945
1946         return 0;
1947
1948 err_find:
1949         kfree(names);
1950 err_names:
1951         kfree(callbacks);
1952 err_callback:
1953         kfree(vqs);
1954 err_vq:
1955         return ret;
1956 }
1957
1958 static int virtnet_alloc_queues(struct virtnet_info *vi)
1959 {
1960         int i;
1961
1962         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1963         if (!vi->sq)
1964                 goto err_sq;
1965         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1966         if (!vi->rq)
1967                 goto err_rq;
1968
1969         INIT_DELAYED_WORK(&vi->refill, refill_work);
1970         for (i = 0; i < vi->max_queue_pairs; i++) {
1971                 vi->rq[i].pages = NULL;
1972                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1973                                napi_weight);
1974
1975                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1976                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
1977                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1978         }
1979
1980         return 0;
1981
1982 err_rq:
1983         kfree(vi->sq);
1984 err_sq:
1985         return -ENOMEM;
1986 }
1987
1988 static int init_vqs(struct virtnet_info *vi)
1989 {
1990         int ret;
1991
1992         /* Allocate send & receive queues */
1993         ret = virtnet_alloc_queues(vi);
1994         if (ret)
1995                 goto err;
1996
1997         ret = virtnet_find_vqs(vi);
1998         if (ret)
1999                 goto err_free;
2000
2001         get_online_cpus();
2002         virtnet_set_affinity(vi);
2003         put_online_cpus();
2004
2005         return 0;
2006
2007 err_free:
2008         virtnet_free_queues(vi);
2009 err:
2010         return ret;
2011 }
2012
2013 #ifdef CONFIG_SYSFS
2014 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2015                 struct rx_queue_attribute *attribute, char *buf)
2016 {
2017         struct virtnet_info *vi = netdev_priv(queue->dev);
2018         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2019         struct ewma_pkt_len *avg;
2020
2021         BUG_ON(queue_index >= vi->max_queue_pairs);
2022         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2023         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
2024 }
2025
2026 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2027         __ATTR_RO(mergeable_rx_buffer_size);
2028
2029 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2030         &mergeable_rx_buffer_size_attribute.attr,
2031         NULL
2032 };
2033
2034 static const struct attribute_group virtio_net_mrg_rx_group = {
2035         .name = "virtio_net",
2036         .attrs = virtio_net_mrg_rx_attrs
2037 };
2038 #endif
2039
2040 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2041                                     unsigned int fbit,
2042                                     const char *fname, const char *dname)
2043 {
2044         if (!virtio_has_feature(vdev, fbit))
2045                 return false;
2046
2047         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2048                 fname, dname);
2049
2050         return true;
2051 }
2052
2053 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2054         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2055
2056 static bool virtnet_validate_features(struct virtio_device *vdev)
2057 {
2058         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2059             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2060                              "VIRTIO_NET_F_CTRL_VQ") ||
2061              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2062                              "VIRTIO_NET_F_CTRL_VQ") ||
2063              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2064                              "VIRTIO_NET_F_CTRL_VQ") ||
2065              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2066              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2067                              "VIRTIO_NET_F_CTRL_VQ"))) {
2068                 return false;
2069         }
2070
2071         return true;
2072 }
2073
2074 #define MIN_MTU ETH_MIN_MTU
2075 #define MAX_MTU ETH_MAX_MTU
2076
2077 static int virtnet_probe(struct virtio_device *vdev)
2078 {
2079         int i, err;
2080         struct net_device *dev;
2081         struct virtnet_info *vi;
2082         u16 max_queue_pairs;
2083         int mtu;
2084
2085         if (!vdev->config->get) {
2086                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2087                         __func__);
2088                 return -EINVAL;
2089         }
2090
2091         if (!virtnet_validate_features(vdev))
2092                 return -EINVAL;
2093
2094         /* Find if host supports multiqueue virtio_net device */
2095         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2096                                    struct virtio_net_config,
2097                                    max_virtqueue_pairs, &max_queue_pairs);
2098
2099         /* We need at least 2 queue's */
2100         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2101             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2102             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2103                 max_queue_pairs = 1;
2104
2105         /* Allocate ourselves a network device with room for our info */
2106         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2107         if (!dev)
2108                 return -ENOMEM;
2109
2110         /* Set up network device as normal. */
2111         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2112         dev->netdev_ops = &virtnet_netdev;
2113         dev->features = NETIF_F_HIGHDMA;
2114
2115         dev->ethtool_ops = &virtnet_ethtool_ops;
2116         SET_NETDEV_DEV(dev, &vdev->dev);
2117
2118         /* Do we support "hardware" checksums? */
2119         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2120                 /* This opens up the world of extra features. */
2121                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2122                 if (csum)
2123                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2124
2125                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2126                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
2127                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2128                 }
2129                 /* Individual feature bits: what can host handle? */
2130                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2131                         dev->hw_features |= NETIF_F_TSO;
2132                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2133                         dev->hw_features |= NETIF_F_TSO6;
2134                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2135                         dev->hw_features |= NETIF_F_TSO_ECN;
2136                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
2137                         dev->hw_features |= NETIF_F_UFO;
2138
2139                 dev->features |= NETIF_F_GSO_ROBUST;
2140
2141                 if (gso)
2142                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
2143                 /* (!csum && gso) case will be fixed by register_netdev() */
2144         }
2145         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2146                 dev->features |= NETIF_F_RXCSUM;
2147
2148         dev->vlan_features = dev->features;
2149
2150         /* MTU range: 68 - 65535 */
2151         dev->min_mtu = MIN_MTU;
2152         dev->max_mtu = MAX_MTU;
2153
2154         /* Configuration may specify what MAC to use.  Otherwise random. */
2155         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2156                 virtio_cread_bytes(vdev,
2157                                    offsetof(struct virtio_net_config, mac),
2158                                    dev->dev_addr, dev->addr_len);
2159         else
2160                 eth_hw_addr_random(dev);
2161
2162         /* Set up our device-specific information */
2163         vi = netdev_priv(dev);
2164         vi->dev = dev;
2165         vi->vdev = vdev;
2166         vdev->priv = vi;
2167         vi->stats = alloc_percpu(struct virtnet_stats);
2168         err = -ENOMEM;
2169         if (vi->stats == NULL)
2170                 goto free;
2171
2172         for_each_possible_cpu(i) {
2173                 struct virtnet_stats *virtnet_stats;
2174                 virtnet_stats = per_cpu_ptr(vi->stats, i);
2175                 u64_stats_init(&virtnet_stats->tx_syncp);
2176                 u64_stats_init(&virtnet_stats->rx_syncp);
2177         }
2178
2179         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2180
2181         /* If we can receive ANY GSO packets, we must allocate large ones. */
2182         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2183             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2184             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2185             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2186                 vi->big_packets = true;
2187
2188         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2189                 vi->mergeable_rx_bufs = true;
2190
2191         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2192             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2193                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2194         else
2195                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2196
2197         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2198             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2199                 vi->any_header_sg = true;
2200
2201         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2202                 vi->has_cvq = true;
2203
2204         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2205                 mtu = virtio_cread16(vdev,
2206                                      offsetof(struct virtio_net_config,
2207                                               mtu));
2208                 if (mtu < dev->min_mtu) {
2209                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2210                 } else {
2211                         dev->mtu = mtu;
2212                         dev->max_mtu = mtu;
2213                 }
2214         }
2215
2216         if (vi->any_header_sg)
2217                 dev->needed_headroom = vi->hdr_len;
2218
2219         /* Enable multiqueue by default */
2220         if (num_online_cpus() >= max_queue_pairs)
2221                 vi->curr_queue_pairs = max_queue_pairs;
2222         else
2223                 vi->curr_queue_pairs = num_online_cpus();
2224         vi->max_queue_pairs = max_queue_pairs;
2225
2226         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2227         err = init_vqs(vi);
2228         if (err)
2229                 goto free_stats;
2230
2231 #ifdef CONFIG_SYSFS
2232         if (vi->mergeable_rx_bufs)
2233                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2234 #endif
2235         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2236         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2237
2238         virtnet_init_settings(dev);
2239
2240         err = register_netdev(dev);
2241         if (err) {
2242                 pr_debug("virtio_net: registering device failed\n");
2243                 goto free_vqs;
2244         }
2245
2246         virtio_device_ready(vdev);
2247
2248         err = virtnet_cpu_notif_add(vi);
2249         if (err) {
2250                 pr_debug("virtio_net: registering cpu notifier failed\n");
2251                 goto free_unregister_netdev;
2252         }
2253
2254         rtnl_lock();
2255         virtnet_set_queues(vi, vi->curr_queue_pairs);
2256         rtnl_unlock();
2257
2258         /* Assume link up if device can't report link status,
2259            otherwise get link status from config. */
2260         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2261                 netif_carrier_off(dev);
2262                 schedule_work(&vi->config_work);
2263         } else {
2264                 vi->status = VIRTIO_NET_S_LINK_UP;
2265                 netif_carrier_on(dev);
2266         }
2267
2268         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2269                  dev->name, max_queue_pairs);
2270
2271         return 0;
2272
2273 free_unregister_netdev:
2274         vi->vdev->config->reset(vdev);
2275
2276         unregister_netdev(dev);
2277 free_vqs:
2278         cancel_delayed_work_sync(&vi->refill);
2279         free_receive_page_frags(vi);
2280         virtnet_del_vqs(vi);
2281 free_stats:
2282         free_percpu(vi->stats);
2283 free:
2284         free_netdev(dev);
2285         return err;
2286 }
2287
2288 static void remove_vq_common(struct virtnet_info *vi)
2289 {
2290         vi->vdev->config->reset(vi->vdev);
2291
2292         /* Free unused buffers in both send and recv, if any. */
2293         free_unused_bufs(vi);
2294
2295         free_receive_bufs(vi);
2296
2297         free_receive_page_frags(vi);
2298
2299         virtnet_del_vqs(vi);
2300 }
2301
2302 static void virtnet_remove(struct virtio_device *vdev)
2303 {
2304         struct virtnet_info *vi = vdev->priv;
2305
2306         virtnet_cpu_notif_remove(vi);
2307
2308         /* Make sure no work handler is accessing the device. */
2309         flush_work(&vi->config_work);
2310
2311         unregister_netdev(vi->dev);
2312
2313         remove_vq_common(vi);
2314
2315         free_percpu(vi->stats);
2316         free_netdev(vi->dev);
2317 }
2318
2319 #ifdef CONFIG_PM_SLEEP
2320 static int virtnet_freeze(struct virtio_device *vdev)
2321 {
2322         struct virtnet_info *vi = vdev->priv;
2323         int i;
2324
2325         virtnet_cpu_notif_remove(vi);
2326
2327         /* Make sure no work handler is accessing the device */
2328         flush_work(&vi->config_work);
2329
2330         netif_device_detach(vi->dev);
2331         cancel_delayed_work_sync(&vi->refill);
2332
2333         if (netif_running(vi->dev)) {
2334                 for (i = 0; i < vi->max_queue_pairs; i++)
2335                         napi_disable(&vi->rq[i].napi);
2336         }
2337
2338         remove_vq_common(vi);
2339
2340         return 0;
2341 }
2342
2343 static int virtnet_restore(struct virtio_device *vdev)
2344 {
2345         struct virtnet_info *vi = vdev->priv;
2346         int err, i;
2347
2348         err = init_vqs(vi);
2349         if (err)
2350                 return err;
2351
2352         virtio_device_ready(vdev);
2353
2354         if (netif_running(vi->dev)) {
2355                 for (i = 0; i < vi->curr_queue_pairs; i++)
2356                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2357                                 schedule_delayed_work(&vi->refill, 0);
2358
2359                 for (i = 0; i < vi->max_queue_pairs; i++)
2360                         virtnet_napi_enable(&vi->rq[i]);
2361         }
2362
2363         netif_device_attach(vi->dev);
2364
2365         rtnl_lock();
2366         virtnet_set_queues(vi, vi->curr_queue_pairs);
2367         rtnl_unlock();
2368
2369         err = virtnet_cpu_notif_add(vi);
2370         if (err)
2371                 return err;
2372
2373         return 0;
2374 }
2375 #endif
2376
2377 static struct virtio_device_id id_table[] = {
2378         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2379         { 0 },
2380 };
2381
2382 #define VIRTNET_FEATURES \
2383         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2384         VIRTIO_NET_F_MAC, \
2385         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2386         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2387         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2388         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2389         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2390         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2391         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2392         VIRTIO_NET_F_MTU
2393
2394 static unsigned int features[] = {
2395         VIRTNET_FEATURES,
2396 };
2397
2398 static unsigned int features_legacy[] = {
2399         VIRTNET_FEATURES,
2400         VIRTIO_NET_F_GSO,
2401         VIRTIO_F_ANY_LAYOUT,
2402 };
2403
2404 static struct virtio_driver virtio_net_driver = {
2405         .feature_table = features,
2406         .feature_table_size = ARRAY_SIZE(features),
2407         .feature_table_legacy = features_legacy,
2408         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2409         .driver.name =  KBUILD_MODNAME,
2410         .driver.owner = THIS_MODULE,
2411         .id_table =     id_table,
2412         .probe =        virtnet_probe,
2413         .remove =       virtnet_remove,
2414         .config_changed = virtnet_config_changed,
2415 #ifdef CONFIG_PM_SLEEP
2416         .freeze =       virtnet_freeze,
2417         .restore =      virtnet_restore,
2418 #endif
2419 };
2420
2421 static __init int virtio_net_driver_init(void)
2422 {
2423         int ret;
2424
2425         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "AP_VIRT_NET_ONLINE",
2426                                       virtnet_cpu_online,
2427                                       virtnet_cpu_down_prep);
2428         if (ret < 0)
2429                 goto out;
2430         virtionet_online = ret;
2431         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "VIRT_NET_DEAD",
2432                                       NULL, virtnet_cpu_dead);
2433         if (ret)
2434                 goto err_dead;
2435
2436         ret = register_virtio_driver(&virtio_net_driver);
2437         if (ret)
2438                 goto err_virtio;
2439         return 0;
2440 err_virtio:
2441         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2442 err_dead:
2443         cpuhp_remove_multi_state(virtionet_online);
2444 out:
2445         return ret;
2446 }
2447 module_init(virtio_net_driver_init);
2448
2449 static __exit void virtio_net_driver_exit(void)
2450 {
2451         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2452         cpuhp_remove_multi_state(virtionet_online);
2453         unregister_virtio_driver(&virtio_net_driver);
2454 }
2455 module_exit(virtio_net_driver_exit);
2456
2457 MODULE_DEVICE_TABLE(virtio, id_table);
2458 MODULE_DESCRIPTION("Virtio network driver");
2459 MODULE_LICENSE("GPL");