drivers/net/virtio_net.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /* A network driver using virtio.
   3  *
   4  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
   5  */
   6 //#define DEBUG
   7 #include <linux/netdevice.h>
   8 #include <linux/etherdevice.h>
   9 #include <linux/ethtool.h>
  10 #include <linux/module.h>
  11 #include <linux/virtio.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/bpf.h>
  14 #include <linux/bpf_trace.h>
  15 #include <linux/scatterlist.h>
  16 #include <linux/if_vlan.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/average.h>
  20 #include <linux/filter.h>
  21 #include <linux/kernel.h>
  22 #include <net/route.h>
  23 #include <net/xdp.h>
  24 #include <net/net_failover.h>
  25 #include <net/netdev_rx_queue.h>
  26
  27 static int napi_weight = NAPI_POLL_WEIGHT;
  28 module_param(napi_weight, int, 0444);
  29
  30 static bool csum = true, gso = true, napi_tx = true;
  31 module_param(csum, bool, 0444);
  32 module_param(gso, bool, 0444);
  33 module_param(napi_tx, bool, 0644);
  34
  35 /* FIXME: MTU in config. */
  36 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  37 #define GOOD_COPY_LEN   128
  38
  39 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  40
  41 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  42 #define VIRTIO_XDP_HEADROOM 256
  43
  44 /* Separating two types of XDP xmit */
  45 #define VIRTIO_XDP_TX           BIT(0)
  46 #define VIRTIO_XDP_REDIR        BIT(1)
  47
  48 #define VIRTIO_XDP_FLAG BIT(0)
  49
  50 /* RX packet size EWMA. The average packet size is used to determine the packet
  51  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  52  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  53  * term, transient changes in packet size.
  54  */
  55 DECLARE_EWMA(pkt_len, 0, 64)
  56
  57 #define VIRTNET_DRIVER_VERSION "1.0.0"
  58
  59 static const unsigned long guest_offloads[] = {
  60         VIRTIO_NET_F_GUEST_TSO4,
  61         VIRTIO_NET_F_GUEST_TSO6,
  62         VIRTIO_NET_F_GUEST_ECN,
  63         VIRTIO_NET_F_GUEST_UFO,
  64         VIRTIO_NET_F_GUEST_CSUM,
  65         VIRTIO_NET_F_GUEST_USO4,
  66         VIRTIO_NET_F_GUEST_USO6,
  67         VIRTIO_NET_F_GUEST_HDRLEN
  68 };
  69
  70 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
  71                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
  72                                 (1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
  73                                 (1ULL << VIRTIO_NET_F_GUEST_UFO)  | \
  74                                 (1ULL << VIRTIO_NET_F_GUEST_USO4) | \
  75                                 (1ULL << VIRTIO_NET_F_GUEST_USO6))
  76
  77 struct virtnet_stat_desc {
  78         char desc[ETH_GSTRING_LEN];
  79         size_t offset;
  80 };
  81
  82 struct virtnet_sq_stats {
  83         struct u64_stats_sync syncp;
  84         u64_stats_t packets;
  85         u64_stats_t bytes;
  86         u64_stats_t xdp_tx;
  87         u64_stats_t xdp_tx_drops;
  88         u64_stats_t kicks;
  89         u64_stats_t tx_timeouts;
  90 };
  91
  92 struct virtnet_rq_stats {
  93         struct u64_stats_sync syncp;
  94         u64_stats_t packets;
  95         u64_stats_t bytes;
  96         u64_stats_t drops;
  97         u64_stats_t xdp_packets;
  98         u64_stats_t xdp_tx;
  99         u64_stats_t xdp_redirects;
 100         u64_stats_t xdp_drops;
 101         u64_stats_t kicks;
 102 };
 103
 104 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
 105 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
 106
 107 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
 108         { "packets",            VIRTNET_SQ_STAT(packets) },
 109         { "bytes",              VIRTNET_SQ_STAT(bytes) },
 110         { "xdp_tx",             VIRTNET_SQ_STAT(xdp_tx) },
 111         { "xdp_tx_drops",       VIRTNET_SQ_STAT(xdp_tx_drops) },
 112         { "kicks",              VIRTNET_SQ_STAT(kicks) },
 113         { "tx_timeouts",        VIRTNET_SQ_STAT(tx_timeouts) },
 114 };
 115
 116 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
 117         { "packets",            VIRTNET_RQ_STAT(packets) },
 118         { "bytes",              VIRTNET_RQ_STAT(bytes) },
 119         { "drops",              VIRTNET_RQ_STAT(drops) },
 120         { "xdp_packets",        VIRTNET_RQ_STAT(xdp_packets) },
 121         { "xdp_tx",             VIRTNET_RQ_STAT(xdp_tx) },
 122         { "xdp_redirects",      VIRTNET_RQ_STAT(xdp_redirects) },
 123         { "xdp_drops",          VIRTNET_RQ_STAT(xdp_drops) },
 124         { "kicks",              VIRTNET_RQ_STAT(kicks) },
 125 };
 126
 127 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 128 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 129
 130 struct virtnet_interrupt_coalesce {
 131         u32 max_packets;
 132         u32 max_usecs;
 133 };
 134
 135 /* The dma information of pages allocated at a time. */
 136 struct virtnet_rq_dma {
 137         dma_addr_t addr;
 138         u32 ref;
 139         u16 len;
 140         u16 need_sync;
 141 };
 142
 143 /* Internal representation of a send virtqueue */
 144 struct send_queue {
 145         /* Virtqueue associated with this send _queue */
 146         struct virtqueue *vq;
 147
 148         /* TX: fragments + linear part + virtio header */
 149         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 150
 151         /* Name of the send queue: output.$index */
 152         char name[16];
 153
 154         struct virtnet_sq_stats stats;
 155
 156         struct virtnet_interrupt_coalesce intr_coal;
 157
 158         struct napi_struct napi;
 159
 160         /* Record whether sq is in reset state. */
 161         bool reset;
 162 };
 163
 164 /* Internal representation of a receive virtqueue */
 165 struct receive_queue {
 166         /* Virtqueue associated with this receive_queue */
 167         struct virtqueue *vq;
 168
 169         struct napi_struct napi;
 170
 171         struct bpf_prog __rcu *xdp_prog;
 172
 173         struct virtnet_rq_stats stats;
 174
 175         struct virtnet_interrupt_coalesce intr_coal;
 176
 177         /* Chain pages by the private ptr. */
 178         struct page *pages;
 179
 180         /* Average packet length for mergeable receive buffers. */
 181         struct ewma_pkt_len mrg_avg_pkt_len;
 182
 183         /* Page frag for packet buffer allocation. */
 184         struct page_frag alloc_frag;
 185
 186         /* RX: fragments + linear part + virtio header */
 187         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 188
 189         /* Min single buffer size for mergeable buffers case. */
 190         unsigned int min_buf_len;
 191
 192         /* Name of this receive queue: input.$index */
 193         char name[16];
 194
 195         struct xdp_rxq_info xdp_rxq;
 196
 197         /* Record the last dma info to free after new pages is allocated. */
 198         struct virtnet_rq_dma *last_dma;
 199
 200         /* Do dma by self */
 201         bool do_dma;
 202 };
 203
 204 /* This structure can contain rss message with maximum settings for indirection table and keysize
 205  * Note, that default structure that describes RSS configuration virtio_net_rss_config
 206  * contains same info but can't handle table values.
 207  * In any case, structure would be passed to virtio hw through sg_buf split by parts
 208  * because table sizes may be differ according to the device configuration.
 209  */
 210 #define VIRTIO_NET_RSS_MAX_KEY_SIZE     40
 211 #define VIRTIO_NET_RSS_MAX_TABLE_LEN    128
 212 struct virtio_net_ctrl_rss {
 213         u32 hash_types;
 214         u16 indirection_table_mask;
 215         u16 unclassified_queue;
 216         u16 indirection_table[VIRTIO_NET_RSS_MAX_TABLE_LEN];
 217         u16 max_tx_vq;
 218         u8 hash_key_length;
 219         u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE];
 220 };
 221
 222 /* Control VQ buffers: protected by the rtnl lock */
 223 struct control_buf {
 224         struct virtio_net_ctrl_hdr hdr;
 225         virtio_net_ctrl_ack status;
 226         struct virtio_net_ctrl_mq mq;
 227         u8 promisc;
 228         u8 allmulti;
 229         __virtio16 vid;
 230         __virtio64 offloads;
 231         struct virtio_net_ctrl_rss rss;
 232         struct virtio_net_ctrl_coal_tx coal_tx;
 233         struct virtio_net_ctrl_coal_rx coal_rx;
 234         struct virtio_net_ctrl_coal_vq coal_vq;
 235 };
 236
 237 struct virtnet_info {
 238         struct virtio_device *vdev;
 239         struct virtqueue *cvq;
 240         struct net_device *dev;
 241         struct send_queue *sq;
 242         struct receive_queue *rq;
 243         unsigned int status;
 244
 245         /* Max # of queue pairs supported by the device */
 246         u16 max_queue_pairs;
 247
 248         /* # of queue pairs currently used by the driver */
 249         u16 curr_queue_pairs;
 250
 251         /* # of XDP queue pairs currently used by the driver */
 252         u16 xdp_queue_pairs;
 253
 254         /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
 255         bool xdp_enabled;
 256
 257         /* I like... big packets and I cannot lie! */
 258         bool big_packets;
 259
 260         /* number of sg entries allocated for big packets */
 261         unsigned int big_packets_num_skbfrags;
 262
 263         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 264         bool mergeable_rx_bufs;
 265
 266         /* Host supports rss and/or hash report */
 267         bool has_rss;
 268         bool has_rss_hash_report;
 269         u8 rss_key_size;
 270         u16 rss_indir_table_size;
 271         u32 rss_hash_types_supported;
 272         u32 rss_hash_types_saved;
 273
 274         /* Has control virtqueue */
 275         bool has_cvq;
 276
 277         /* Host can handle any s/g split between our header and packet data */
 278         bool any_header_sg;
 279
 280         /* Packet virtio header size */
 281         u8 hdr_len;
 282
 283         /* Work struct for delayed refilling if we run low on memory. */
 284         struct delayed_work refill;
 285
 286         /* Is delayed refill enabled? */
 287         bool refill_enabled;
 288
 289         /* The lock to synchronize the access to refill_enabled */
 290         spinlock_t refill_lock;
 291
 292         /* Work struct for config space updates */
 293         struct work_struct config_work;
 294
 295         /* Does the affinity hint is set for virtqueues? */
 296         bool affinity_hint_set;
 297
 298         /* CPU hotplug instances for online & dead */
 299         struct hlist_node node;
 300         struct hlist_node node_dead;
 301
 302         struct control_buf *ctrl;
 303
 304         /* Ethtool settings */
 305         u8 duplex;
 306         u32 speed;
 307
 308         /* Interrupt coalescing settings */
 309         struct virtnet_interrupt_coalesce intr_coal_tx;
 310         struct virtnet_interrupt_coalesce intr_coal_rx;
 311
 312         unsigned long guest_offloads;
 313         unsigned long guest_offloads_capable;
 314
 315         /* failover when STANDBY feature enabled */
 316         struct failover *failover;
 317 };
 318
 319 struct padded_vnet_hdr {
 320         struct virtio_net_hdr_v1_hash hdr;
 321         /*
 322          * hdr is in a separate sg buffer, and data sg buffer shares same page
 323          * with this header sg. This padding makes next sg 16 byte aligned
 324          * after the header.
 325          */
 326         char padding[12];
 327 };
 328
 329 struct virtio_net_common_hdr {
 330         union {
 331                 struct virtio_net_hdr hdr;
 332                 struct virtio_net_hdr_mrg_rxbuf mrg_hdr;
 333                 struct virtio_net_hdr_v1_hash hash_v1_hdr;
 334         };
 335 };
 336
 337 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf);
 338
 339 static bool is_xdp_frame(void *ptr)
 340 {
 341         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 342 }
 343
 344 static void *xdp_to_ptr(struct xdp_frame *ptr)
 345 {
 346         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 347 }
 348
 349 static struct xdp_frame *ptr_to_xdp(void *ptr)
 350 {
 351         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 352 }
 353
 354 /* Converting between virtqueue no. and kernel tx/rx queue no.
 355  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 356  */
 357 static int vq2txq(struct virtqueue *vq)
 358 {
 359         return (vq->index - 1) / 2;
 360 }
 361
 362 static int txq2vq(int txq)
 363 {
 364         return txq * 2 + 1;
 365 }
 366
 367 static int vq2rxq(struct virtqueue *vq)
 368 {
 369         return vq->index / 2;
 370 }
 371
 372 static int rxq2vq(int rxq)
 373 {
 374         return rxq * 2;
 375 }
 376
 377 static inline struct virtio_net_common_hdr *
 378 skb_vnet_common_hdr(struct sk_buff *skb)
 379 {
 380         return (struct virtio_net_common_hdr *)skb->cb;
 381 }
 382
 383 /*
 384  * private is used to chain pages for big packets, put the whole
 385  * most recent used list in the beginning for reuse
 386  */
 387 static void give_pages(struct receive_queue *rq, struct page *page)
 388 {
 389         struct page *end;
 390
 391         /* Find end of list, sew whole thing into vi->rq.pages. */
 392         for (end = page; end->private; end = (struct page *)end->private);
 393         end->private = (unsigned long)rq->pages;
 394         rq->pages = page;
 395 }
 396
 397 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 398 {
 399         struct page *p = rq->pages;
 400
 401         if (p) {
 402                 rq->pages = (struct page *)p->private;
 403                 /* clear private here, it is used to chain pages */
 404                 p->private = 0;
 405         } else
 406                 p = alloc_page(gfp_mask);
 407         return p;
 408 }
 409
 410 static void virtnet_rq_free_buf(struct virtnet_info *vi,
 411                                 struct receive_queue *rq, void *buf)
 412 {
 413         if (vi->mergeable_rx_bufs)
 414                 put_page(virt_to_head_page(buf));
 415         else if (vi->big_packets)
 416                 give_pages(rq, buf);
 417         else
 418                 put_page(virt_to_head_page(buf));
 419 }
 420
 421 static void enable_delayed_refill(struct virtnet_info *vi)
 422 {
 423         spin_lock_bh(&vi->refill_lock);
 424         vi->refill_enabled = true;
 425         spin_unlock_bh(&vi->refill_lock);
 426 }
 427
 428 static void disable_delayed_refill(struct virtnet_info *vi)
 429 {
 430         spin_lock_bh(&vi->refill_lock);
 431         vi->refill_enabled = false;
 432         spin_unlock_bh(&vi->refill_lock);
 433 }
 434
 435 static void virtqueue_napi_schedule(struct napi_struct *napi,
 436                                     struct virtqueue *vq)
 437 {
 438         if (napi_schedule_prep(napi)) {
 439                 virtqueue_disable_cb(vq);
 440                 __napi_schedule(napi);
 441         }
 442 }
 443
 444 static void virtqueue_napi_complete(struct napi_struct *napi,
 445                                     struct virtqueue *vq, int processed)
 446 {
 447         int opaque;
 448
 449         opaque = virtqueue_enable_cb_prepare(vq);
 450         if (napi_complete_done(napi, processed)) {
 451                 if (unlikely(virtqueue_poll(vq, opaque)))
 452                         virtqueue_napi_schedule(napi, vq);
 453         } else {
 454                 virtqueue_disable_cb(vq);
 455         }
 456 }
 457
 458 static void skb_xmit_done(struct virtqueue *vq)
 459 {
 460         struct virtnet_info *vi = vq->vdev->priv;
 461         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 462
 463         /* Suppress further interrupts. */
 464         virtqueue_disable_cb(vq);
 465
 466         if (napi->weight)
 467                 virtqueue_napi_schedule(napi, vq);
 468         else
 469                 /* We were probably waiting for more output buffers. */
 470                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 471 }
 472
 473 #define MRG_CTX_HEADER_SHIFT 22
 474 static void *mergeable_len_to_ctx(unsigned int truesize,
 475                                   unsigned int headroom)
 476 {
 477         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 478 }
 479
 480 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 481 {
 482         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 483 }
 484
 485 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 486 {
 487         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 488 }
 489
 490 static struct sk_buff *virtnet_build_skb(void *buf, unsigned int buflen,
 491                                          unsigned int headroom,
 492                                          unsigned int len)
 493 {
 494         struct sk_buff *skb;
 495
 496         skb = build_skb(buf, buflen);
 497         if (unlikely(!skb))
 498                 return NULL;
 499
 500         skb_reserve(skb, headroom);
 501         skb_put(skb, len);
 502
 503         return skb;
 504 }
 505
 506 /* Called from bottom half context */
 507 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 508                                    struct receive_queue *rq,
 509                                    struct page *page, unsigned int offset,
 510                                    unsigned int len, unsigned int truesize,
 511                                    unsigned int headroom)
 512 {
 513         struct sk_buff *skb;
 514         struct virtio_net_common_hdr *hdr;
 515         unsigned int copy, hdr_len, hdr_padded_len;
 516         struct page *page_to_free = NULL;
 517         int tailroom, shinfo_size;
 518         char *p, *hdr_p, *buf;
 519
 520         p = page_address(page) + offset;
 521         hdr_p = p;
 522
 523         hdr_len = vi->hdr_len;
 524         if (vi->mergeable_rx_bufs)
 525                 hdr_padded_len = hdr_len;
 526         else
 527                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 528
 529         buf = p - headroom;
 530         len -= hdr_len;
 531         offset += hdr_padded_len;
 532         p += hdr_padded_len;
 533         tailroom = truesize - headroom  - hdr_padded_len - len;
 534
 535         shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 536
 537         /* copy small packet so we can reuse these pages */
 538         if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
 539                 skb = virtnet_build_skb(buf, truesize, p - buf, len);
 540                 if (unlikely(!skb))
 541                         return NULL;
 542
 543                 page = (struct page *)page->private;
 544                 if (page)
 545                         give_pages(rq, page);
 546                 goto ok;
 547         }
 548
 549         /* copy small packet so we can reuse these pages for small data */
 550         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 551         if (unlikely(!skb))
 552                 return NULL;
 553
 554         /* Copy all frame if it fits skb->head, otherwise
 555          * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
 556          */
 557         if (len <= skb_tailroom(skb))
 558                 copy = len;
 559         else
 560                 copy = ETH_HLEN;
 561         skb_put_data(skb, p, copy);
 562
 563         len -= copy;
 564         offset += copy;
 565
 566         if (vi->mergeable_rx_bufs) {
 567                 if (len)
 568                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 569                 else
 570                         page_to_free = page;
 571                 goto ok;
 572         }
 573
 574         /*
 575          * Verify that we can indeed put this data into a skb.
 576          * This is here to handle cases when the device erroneously
 577          * tries to receive more than is possible. This is usually
 578          * the case of a broken device.
 579          */
 580         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 581                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 582                 dev_kfree_skb(skb);
 583                 return NULL;
 584         }
 585         BUG_ON(offset >= PAGE_SIZE);
 586         while (len) {
 587                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 588                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 589                                 frag_size, truesize);
 590                 len -= frag_size;
 591                 page = (struct page *)page->private;
 592                 offset = 0;
 593         }
 594
 595         if (page)
 596                 give_pages(rq, page);
 597
 598 ok:
 599         hdr = skb_vnet_common_hdr(skb);
 600         memcpy(hdr, hdr_p, hdr_len);
 601         if (page_to_free)
 602                 put_page(page_to_free);
 603
 604         return skb;
 605 }
 606
 607 static void virtnet_rq_unmap(struct receive_queue *rq, void *buf, u32 len)
 608 {
 609         struct page *page = virt_to_head_page(buf);
 610         struct virtnet_rq_dma *dma;
 611         void *head;
 612         int offset;
 613
 614         head = page_address(page);
 615
 616         dma = head;
 617
 618         --dma->ref;
 619
 620         if (dma->need_sync && len) {
 621                 offset = buf - (head + sizeof(*dma));
 622
 623                 virtqueue_dma_sync_single_range_for_cpu(rq->vq, dma->addr,
 624                                                         offset, len,
 625                                                         DMA_FROM_DEVICE);
 626         }
 627
 628         if (dma->ref)
 629                 return;
 630
 631         virtqueue_dma_unmap_single_attrs(rq->vq, dma->addr, dma->len,
 632                                          DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
 633         put_page(page);
 634 }
 635
 636 static void *virtnet_rq_get_buf(struct receive_queue *rq, u32 *len, void **ctx)
 637 {
 638         void *buf;
 639
 640         buf = virtqueue_get_buf_ctx(rq->vq, len, ctx);
 641         if (buf && rq->do_dma)
 642                 virtnet_rq_unmap(rq, buf, *len);
 643
 644         return buf;
 645 }
 646
 647 static void virtnet_rq_init_one_sg(struct receive_queue *rq, void *buf, u32 len)
 648 {
 649         struct virtnet_rq_dma *dma;
 650         dma_addr_t addr;
 651         u32 offset;
 652         void *head;
 653
 654         if (!rq->do_dma) {
 655                 sg_init_one(rq->sg, buf, len);
 656                 return;
 657         }
 658
 659         head = page_address(rq->alloc_frag.page);
 660
 661         offset = buf - head;
 662
 663         dma = head;
 664
 665         addr = dma->addr - sizeof(*dma) + offset;
 666
 667         sg_init_table(rq->sg, 1);
 668         rq->sg[0].dma_address = addr;
 669         rq->sg[0].length = len;
 670 }
 671
 672 static void *virtnet_rq_alloc(struct receive_queue *rq, u32 size, gfp_t gfp)
 673 {
 674         struct page_frag *alloc_frag = &rq->alloc_frag;
 675         struct virtnet_rq_dma *dma;
 676         void *buf, *head;
 677         dma_addr_t addr;
 678
 679         if (unlikely(!skb_page_frag_refill(size, alloc_frag, gfp)))
 680                 return NULL;
 681
 682         head = page_address(alloc_frag->page);
 683
 684         if (rq->do_dma) {
 685                 dma = head;
 686
 687                 /* new pages */
 688                 if (!alloc_frag->offset) {
 689                         if (rq->last_dma) {
 690                                 /* Now, the new page is allocated, the last dma
 691                                  * will not be used. So the dma can be unmapped
 692                                  * if the ref is 0.
 693                                  */
 694                                 virtnet_rq_unmap(rq, rq->last_dma, 0);
 695                                 rq->last_dma = NULL;
 696                         }
 697
 698                         dma->len = alloc_frag->size - sizeof(*dma);
 699
 700                         addr = virtqueue_dma_map_single_attrs(rq->vq, dma + 1,
 701                                                               dma->len, DMA_FROM_DEVICE, 0);
 702                         if (virtqueue_dma_mapping_error(rq->vq, addr))
 703                                 return NULL;
 704
 705                         dma->addr = addr;
 706                         dma->need_sync = virtqueue_dma_need_sync(rq->vq, addr);
 707
 708                         /* Add a reference to dma to prevent the entire dma from
 709                          * being released during error handling. This reference
 710                          * will be freed after the pages are no longer used.
 711                          */
 712                         get_page(alloc_frag->page);
 713                         dma->ref = 1;
 714                         alloc_frag->offset = sizeof(*dma);
 715
 716                         rq->last_dma = dma;
 717                 }
 718
 719                 ++dma->ref;
 720         }
 721
 722         buf = head + alloc_frag->offset;
 723
 724         get_page(alloc_frag->page);
 725         alloc_frag->offset += size;
 726
 727         return buf;
 728 }
 729
 730 static void virtnet_rq_set_premapped(struct virtnet_info *vi)
 731 {
 732         int i;
 733
 734         /* disable for big mode */
 735         if (!vi->mergeable_rx_bufs && vi->big_packets)
 736                 return;
 737
 738         for (i = 0; i < vi->max_queue_pairs; i++) {
 739                 if (virtqueue_set_dma_premapped(vi->rq[i].vq))
 740                         continue;
 741
 742                 vi->rq[i].do_dma = true;
 743         }
 744 }
 745
 746 static void virtnet_rq_unmap_free_buf(struct virtqueue *vq, void *buf)
 747 {
 748         struct virtnet_info *vi = vq->vdev->priv;
 749         struct receive_queue *rq;
 750         int i = vq2rxq(vq);
 751
 752         rq = &vi->rq[i];
 753
 754         if (rq->do_dma)
 755                 virtnet_rq_unmap(rq, buf, 0);
 756
 757         virtnet_rq_free_buf(vi, rq, buf);
 758 }
 759
 760 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
 761 {
 762         unsigned int len;
 763         unsigned int packets = 0;
 764         unsigned int bytes = 0;
 765         void *ptr;
 766
 767         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 768                 if (likely(!is_xdp_frame(ptr))) {
 769                         struct sk_buff *skb = ptr;
 770
 771                         pr_debug("Sent skb %p\n", skb);
 772
 773                         bytes += skb->len;
 774                         napi_consume_skb(skb, in_napi);
 775                 } else {
 776                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 777
 778                         bytes += xdp_get_frame_len(frame);
 779                         xdp_return_frame(frame);
 780                 }
 781                 packets++;
 782         }
 783
 784         /* Avoid overhead when no packets have been processed
 785          * happens when called speculatively from start_xmit.
 786          */
 787         if (!packets)
 788                 return;
 789
 790         u64_stats_update_begin(&sq->stats.syncp);
 791         u64_stats_add(&sq->stats.bytes, bytes);
 792         u64_stats_add(&sq->stats.packets, packets);
 793         u64_stats_update_end(&sq->stats.syncp);
 794 }
 795
 796 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
 797 {
 798         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
 799                 return false;
 800         else if (q < vi->curr_queue_pairs)
 801                 return true;
 802         else
 803                 return false;
 804 }
 805
 806 static void check_sq_full_and_disable(struct virtnet_info *vi,
 807                                       struct net_device *dev,
 808                                       struct send_queue *sq)
 809 {
 810         bool use_napi = sq->napi.weight;
 811         int qnum;
 812
 813         qnum = sq - vi->sq;
 814
 815         /* If running out of space, stop queue to avoid getting packets that we
 816          * are then unable to transmit.
 817          * An alternative would be to force queuing layer to requeue the skb by
 818          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
 819          * returned in a normal path of operation: it means that driver is not
 820          * maintaining the TX queue stop/start state properly, and causes
 821          * the stack to do a non-trivial amount of useless work.
 822          * Since most packets only take 1 or 2 ring slots, stopping the queue
 823          * early means 16 slots are typically wasted.
 824          */
 825         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
 826                 netif_stop_subqueue(dev, qnum);
 827                 if (use_napi) {
 828                         if (unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
 829                                 virtqueue_napi_schedule(&sq->napi, sq->vq);
 830                 } else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 831                         /* More just got used, free them then recheck. */
 832                         free_old_xmit_skbs(sq, false);
 833                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 834                                 netif_start_subqueue(dev, qnum);
 835                                 virtqueue_disable_cb(sq->vq);
 836                         }
 837                 }
 838         }
 839 }
 840
 841 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 842                                    struct send_queue *sq,
 843                                    struct xdp_frame *xdpf)
 844 {
 845         struct virtio_net_hdr_mrg_rxbuf *hdr;
 846         struct skb_shared_info *shinfo;
 847         u8 nr_frags = 0;
 848         int err, i;
 849
 850         if (unlikely(xdpf->headroom < vi->hdr_len))
 851                 return -EOVERFLOW;
 852
 853         if (unlikely(xdp_frame_has_frags(xdpf))) {
 854                 shinfo = xdp_get_shared_info_from_frame(xdpf);
 855                 nr_frags = shinfo->nr_frags;
 856         }
 857
 858         /* In wrapping function virtnet_xdp_xmit(), we need to free
 859          * up the pending old buffers, where we need to calculate the
 860          * position of skb_shared_info in xdp_get_frame_len() and
 861          * xdp_return_frame(), which will involve to xdpf->data and
 862          * xdpf->headroom. Therefore, we need to update the value of
 863          * headroom synchronously here.
 864          */
 865         xdpf->headroom -= vi->hdr_len;
 866         xdpf->data -= vi->hdr_len;
 867         /* Zero header and leave csum up to XDP layers */
 868         hdr = xdpf->data;
 869         memset(hdr, 0, vi->hdr_len);
 870         xdpf->len   += vi->hdr_len;
 871
 872         sg_init_table(sq->sg, nr_frags + 1);
 873         sg_set_buf(sq->sg, xdpf->data, xdpf->len);
 874         for (i = 0; i < nr_frags; i++) {
 875                 skb_frag_t *frag = &shinfo->frags[i];
 876
 877                 sg_set_page(&sq->sg[i + 1], skb_frag_page(frag),
 878                             skb_frag_size(frag), skb_frag_off(frag));
 879         }
 880
 881         err = virtqueue_add_outbuf(sq->vq, sq->sg, nr_frags + 1,
 882                                    xdp_to_ptr(xdpf), GFP_ATOMIC);
 883         if (unlikely(err))
 884                 return -ENOSPC; /* Caller handle free/refcnt */
 885
 886         return 0;
 887 }
 888
 889 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
 890  * the current cpu, so it does not need to be locked.
 891  *
 892  * Here we use marco instead of inline functions because we have to deal with
 893  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
 894  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
 895  * functions to perfectly solve these three problems at the same time.
 896  */
 897 #define virtnet_xdp_get_sq(vi) ({                                       \
 898         int cpu = smp_processor_id();                                   \
 899         struct netdev_queue *txq;                                       \
 900         typeof(vi) v = (vi);                                            \
 901         unsigned int qp;                                                \
 902                                                                         \
 903         if (v->curr_queue_pairs > nr_cpu_ids) {                         \
 904                 qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
 905                 qp += cpu;                                              \
 906                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 907                 __netif_tx_acquire(txq);                                \
 908         } else {                                                        \
 909                 qp = cpu % v->curr_queue_pairs;                         \
 910                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 911                 __netif_tx_lock(txq, cpu);                              \
 912         }                                                               \
 913         v->sq + qp;                                                     \
 914 })
 915
 916 #define virtnet_xdp_put_sq(vi, q) {                                     \
 917         struct netdev_queue *txq;                                       \
 918         typeof(vi) v = (vi);                                            \
 919                                                                         \
 920         txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
 921         if (v->curr_queue_pairs > nr_cpu_ids)                           \
 922                 __netif_tx_release(txq);                                \
 923         else                                                            \
 924                 __netif_tx_unlock(txq);                                 \
 925 }
 926
 927 static int virtnet_xdp_xmit(struct net_device *dev,
 928                             int n, struct xdp_frame **frames, u32 flags)
 929 {
 930         struct virtnet_info *vi = netdev_priv(dev);
 931         struct receive_queue *rq = vi->rq;
 932         struct bpf_prog *xdp_prog;
 933         struct send_queue *sq;
 934         unsigned int len;
 935         int packets = 0;
 936         int bytes = 0;
 937         int nxmit = 0;
 938         int kicks = 0;
 939         void *ptr;
 940         int ret;
 941         int i;
 942
 943         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 944          * indicate XDP resources have been successfully allocated.
 945          */
 946         xdp_prog = rcu_access_pointer(rq->xdp_prog);
 947         if (!xdp_prog)
 948                 return -ENXIO;
 949
 950         sq = virtnet_xdp_get_sq(vi);
 951
 952         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 953                 ret = -EINVAL;
 954                 goto out;
 955         }
 956
 957         /* Free up any pending old buffers before queueing new ones. */
 958         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 959                 if (likely(is_xdp_frame(ptr))) {
 960                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 961
 962                         bytes += xdp_get_frame_len(frame);
 963                         xdp_return_frame(frame);
 964                 } else {
 965                         struct sk_buff *skb = ptr;
 966
 967                         bytes += skb->len;
 968                         napi_consume_skb(skb, false);
 969                 }
 970                 packets++;
 971         }
 972
 973         for (i = 0; i < n; i++) {
 974                 struct xdp_frame *xdpf = frames[i];
 975
 976                 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
 977                         break;
 978                 nxmit++;
 979         }
 980         ret = nxmit;
 981
 982         if (!is_xdp_raw_buffer_queue(vi, sq - vi->sq))
 983                 check_sq_full_and_disable(vi, dev, sq);
 984
 985         if (flags & XDP_XMIT_FLUSH) {
 986                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
 987                         kicks = 1;
 988         }
 989 out:
 990         u64_stats_update_begin(&sq->stats.syncp);
 991         u64_stats_add(&sq->stats.bytes, bytes);
 992         u64_stats_add(&sq->stats.packets, packets);
 993         u64_stats_add(&sq->stats.xdp_tx, n);
 994         u64_stats_add(&sq->stats.xdp_tx_drops, n - nxmit);
 995         u64_stats_add(&sq->stats.kicks, kicks);
 996         u64_stats_update_end(&sq->stats.syncp);
 997
 998         virtnet_xdp_put_sq(vi, sq);
 999         return ret;
1000 }
1001
1002 static void put_xdp_frags(struct xdp_buff *xdp)
1003 {
1004         struct skb_shared_info *shinfo;
1005         struct page *xdp_page;
1006         int i;
1007
1008         if (xdp_buff_has_frags(xdp)) {
1009                 shinfo = xdp_get_shared_info_from_buff(xdp);
1010                 for (i = 0; i < shinfo->nr_frags; i++) {
1011                         xdp_page = skb_frag_page(&shinfo->frags[i]);
1012                         put_page(xdp_page);
1013                 }
1014         }
1015 }
1016
1017 static int virtnet_xdp_handler(struct bpf_prog *xdp_prog, struct xdp_buff *xdp,
1018                                struct net_device *dev,
1019                                unsigned int *xdp_xmit,
1020                                struct virtnet_rq_stats *stats)
1021 {
1022         struct xdp_frame *xdpf;
1023         int err;
1024         u32 act;
1025
1026         act = bpf_prog_run_xdp(xdp_prog, xdp);
1027         u64_stats_inc(&stats->xdp_packets);
1028
1029         switch (act) {
1030         case XDP_PASS:
1031                 return act;
1032
1033         case XDP_TX:
1034                 u64_stats_inc(&stats->xdp_tx);
1035                 xdpf = xdp_convert_buff_to_frame(xdp);
1036                 if (unlikely(!xdpf)) {
1037                         netdev_dbg(dev, "convert buff to frame failed for xdp\n");
1038                         return XDP_DROP;
1039                 }
1040
1041                 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
1042                 if (unlikely(!err)) {
1043                         xdp_return_frame_rx_napi(xdpf);
1044                 } else if (unlikely(err < 0)) {
1045                         trace_xdp_exception(dev, xdp_prog, act);
1046                         return XDP_DROP;
1047                 }
1048                 *xdp_xmit |= VIRTIO_XDP_TX;
1049                 return act;
1050
1051         case XDP_REDIRECT:
1052                 u64_stats_inc(&stats->xdp_redirects);
1053                 err = xdp_do_redirect(dev, xdp, xdp_prog);
1054                 if (err)
1055                         return XDP_DROP;
1056
1057                 *xdp_xmit |= VIRTIO_XDP_REDIR;
1058                 return act;
1059
1060         default:
1061                 bpf_warn_invalid_xdp_action(dev, xdp_prog, act);
1062                 fallthrough;
1063         case XDP_ABORTED:
1064                 trace_xdp_exception(dev, xdp_prog, act);
1065                 fallthrough;
1066         case XDP_DROP:
1067                 return XDP_DROP;
1068         }
1069 }
1070
1071 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
1072 {
1073         return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
1074 }
1075
1076 /* We copy the packet for XDP in the following cases:
1077  *
1078  * 1) Packet is scattered across multiple rx buffers.
1079  * 2) Headroom space is insufficient.
1080  *
1081  * This is inefficient but it's a temporary condition that
1082  * we hit right after XDP is enabled and until queue is refilled
1083  * with large buffers with sufficient headroom - so it should affect
1084  * at most queue size packets.
1085  * Afterwards, the conditions to enable
1086  * XDP should preclude the underlying device from sending packets
1087  * across multiple buffers (num_buf > 1), and we make sure buffers
1088  * have enough headroom.
1089  */
1090 static struct page *xdp_linearize_page(struct receive_queue *rq,
1091                                        int *num_buf,
1092                                        struct page *p,
1093                                        int offset,
1094                                        int page_off,
1095                                        unsigned int *len)
1096 {
1097         int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1098         struct page *page;
1099
1100         if (page_off + *len + tailroom > PAGE_SIZE)
1101                 return NULL;
1102
1103         page = alloc_page(GFP_ATOMIC);
1104         if (!page)
1105                 return NULL;
1106
1107         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
1108         page_off += *len;
1109
1110         while (--*num_buf) {
1111                 unsigned int buflen;
1112                 void *buf;
1113                 int off;
1114
1115                 buf = virtnet_rq_get_buf(rq, &buflen, NULL);
1116                 if (unlikely(!buf))
1117                         goto err_buf;
1118
1119                 p = virt_to_head_page(buf);
1120                 off = buf - page_address(p);
1121
1122                 /* guard against a misconfigured or uncooperative backend that
1123                  * is sending packet larger than the MTU.
1124                  */
1125                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
1126                         put_page(p);
1127                         goto err_buf;
1128                 }
1129
1130                 memcpy(page_address(page) + page_off,
1131                        page_address(p) + off, buflen);
1132                 page_off += buflen;
1133                 put_page(p);
1134         }
1135
1136         /* Headroom does not contribute to packet length */
1137         *len = page_off - VIRTIO_XDP_HEADROOM;
1138         return page;
1139 err_buf:
1140         __free_pages(page, 0);
1141         return NULL;
1142 }
1143
1144 static struct sk_buff *receive_small_build_skb(struct virtnet_info *vi,
1145                                                unsigned int xdp_headroom,
1146                                                void *buf,
1147                                                unsigned int len)
1148 {
1149         unsigned int header_offset;
1150         unsigned int headroom;
1151         unsigned int buflen;
1152         struct sk_buff *skb;
1153
1154         header_offset = VIRTNET_RX_PAD + xdp_headroom;
1155         headroom = vi->hdr_len + header_offset;
1156         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1157                 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1158
1159         skb = virtnet_build_skb(buf, buflen, headroom, len);
1160         if (unlikely(!skb))
1161                 return NULL;
1162
1163         buf += header_offset;
1164         memcpy(skb_vnet_common_hdr(skb), buf, vi->hdr_len);
1165
1166         return skb;
1167 }
1168
1169 static struct sk_buff *receive_small_xdp(struct net_device *dev,
1170                                          struct virtnet_info *vi,
1171                                          struct receive_queue *rq,
1172                                          struct bpf_prog *xdp_prog,
1173                                          void *buf,
1174                                          unsigned int xdp_headroom,
1175                                          unsigned int len,
1176                                          unsigned int *xdp_xmit,
1177                                          struct virtnet_rq_stats *stats)
1178 {
1179         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
1180         unsigned int headroom = vi->hdr_len + header_offset;
1181         struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
1182         struct page *page = virt_to_head_page(buf);
1183         struct page *xdp_page;
1184         unsigned int buflen;
1185         struct xdp_buff xdp;
1186         struct sk_buff *skb;
1187         unsigned int metasize = 0;
1188         u32 act;
1189
1190         if (unlikely(hdr->hdr.gso_type))
1191                 goto err_xdp;
1192
1193         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1194                 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1195
1196         if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
1197                 int offset = buf - page_address(page) + header_offset;
1198                 unsigned int tlen = len + vi->hdr_len;
1199                 int num_buf = 1;
1200
1201                 xdp_headroom = virtnet_get_headroom(vi);
1202                 header_offset = VIRTNET_RX_PAD + xdp_headroom;
1203                 headroom = vi->hdr_len + header_offset;
1204                 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1205                         SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1206                 xdp_page = xdp_linearize_page(rq, &num_buf, page,
1207                                               offset, header_offset,
1208                                               &tlen);
1209                 if (!xdp_page)
1210                         goto err_xdp;
1211
1212                 buf = page_address(xdp_page);
1213                 put_page(page);
1214                 page = xdp_page;
1215         }
1216
1217         xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
1218         xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
1219                          xdp_headroom, len, true);
1220
1221         act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
1222
1223         switch (act) {
1224         case XDP_PASS:
1225                 /* Recalculate length in case bpf program changed it */
1226                 len = xdp.data_end - xdp.data;
1227                 metasize = xdp.data - xdp.data_meta;
1228                 break;
1229
1230         case XDP_TX:
1231         case XDP_REDIRECT:
1232                 goto xdp_xmit;
1233
1234         default:
1235                 goto err_xdp;
1236         }
1237
1238         skb = virtnet_build_skb(buf, buflen, xdp.data - buf, len);
1239         if (unlikely(!skb))
1240                 goto err;
1241
1242         if (metasize)
1243                 skb_metadata_set(skb, metasize);
1244
1245         return skb;
1246
1247 err_xdp:
1248         u64_stats_inc(&stats->xdp_drops);
1249 err:
1250         u64_stats_inc(&stats->drops);
1251         put_page(page);
1252 xdp_xmit:
1253         return NULL;
1254 }
1255
1256 static struct sk_buff *receive_small(struct net_device *dev,
1257                                      struct virtnet_info *vi,
1258                                      struct receive_queue *rq,
1259                                      void *buf, void *ctx,
1260                                      unsigned int len,
1261                                      unsigned int *xdp_xmit,
1262                                      struct virtnet_rq_stats *stats)
1263 {
1264         unsigned int xdp_headroom = (unsigned long)ctx;
1265         struct page *page = virt_to_head_page(buf);
1266         struct sk_buff *skb;
1267
1268         len -= vi->hdr_len;
1269         u64_stats_add(&stats->bytes, len);
1270
1271         if (unlikely(len > GOOD_PACKET_LEN)) {
1272                 pr_debug("%s: rx error: len %u exceeds max size %d\n",
1273                          dev->name, len, GOOD_PACKET_LEN);
1274                 DEV_STATS_INC(dev, rx_length_errors);
1275                 goto err;
1276         }
1277
1278         if (unlikely(vi->xdp_enabled)) {
1279                 struct bpf_prog *xdp_prog;
1280
1281                 rcu_read_lock();
1282                 xdp_prog = rcu_dereference(rq->xdp_prog);
1283                 if (xdp_prog) {
1284                         skb = receive_small_xdp(dev, vi, rq, xdp_prog, buf,
1285                                                 xdp_headroom, len, xdp_xmit,
1286                                                 stats);
1287                         rcu_read_unlock();
1288                         return skb;
1289                 }
1290                 rcu_read_unlock();
1291         }
1292
1293         skb = receive_small_build_skb(vi, xdp_headroom, buf, len);
1294         if (likely(skb))
1295                 return skb;
1296
1297 err:
1298         u64_stats_inc(&stats->drops);
1299         put_page(page);
1300         return NULL;
1301 }
1302
1303 static struct sk_buff *receive_big(struct net_device *dev,
1304                                    struct virtnet_info *vi,
1305                                    struct receive_queue *rq,
1306                                    void *buf,
1307                                    unsigned int len,
1308                                    struct virtnet_rq_stats *stats)
1309 {
1310         struct page *page = buf;
1311         struct sk_buff *skb =
1312                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, 0);
1313
1314         u64_stats_add(&stats->bytes, len - vi->hdr_len);
1315         if (unlikely(!skb))
1316                 goto err;
1317
1318         return skb;
1319
1320 err:
1321         u64_stats_inc(&stats->drops);
1322         give_pages(rq, page);
1323         return NULL;
1324 }
1325
1326 static void mergeable_buf_free(struct receive_queue *rq, int num_buf,
1327                                struct net_device *dev,
1328                                struct virtnet_rq_stats *stats)
1329 {
1330         struct page *page;
1331         void *buf;
1332         int len;
1333
1334         while (num_buf-- > 1) {
1335                 buf = virtnet_rq_get_buf(rq, &len, NULL);
1336                 if (unlikely(!buf)) {
1337                         pr_debug("%s: rx error: %d buffers missing\n",
1338                                  dev->name, num_buf);
1339                         DEV_STATS_INC(dev, rx_length_errors);
1340                         break;
1341                 }
1342                 u64_stats_add(&stats->bytes, len);
1343                 page = virt_to_head_page(buf);
1344                 put_page(page);
1345         }
1346 }
1347
1348 /* Why not use xdp_build_skb_from_frame() ?
1349  * XDP core assumes that xdp frags are PAGE_SIZE in length, while in
1350  * virtio-net there are 2 points that do not match its requirements:
1351  *  1. The size of the prefilled buffer is not fixed before xdp is set.
1352  *  2. xdp_build_skb_from_frame() does more checks that we don't need,
1353  *     like eth_type_trans() (which virtio-net does in receive_buf()).
1354  */
1355 static struct sk_buff *build_skb_from_xdp_buff(struct net_device *dev,
1356                                                struct virtnet_info *vi,
1357                                                struct xdp_buff *xdp,
1358                                                unsigned int xdp_frags_truesz)
1359 {
1360         struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
1361         unsigned int headroom, data_len;
1362         struct sk_buff *skb;
1363         int metasize;
1364         u8 nr_frags;
1365
1366         if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
1367                 pr_debug("Error building skb as missing reserved tailroom for xdp");
1368                 return NULL;
1369         }
1370
1371         if (unlikely(xdp_buff_has_frags(xdp)))
1372                 nr_frags = sinfo->nr_frags;
1373
1374         skb = build_skb(xdp->data_hard_start, xdp->frame_sz);
1375         if (unlikely(!skb))
1376                 return NULL;
1377
1378         headroom = xdp->data - xdp->data_hard_start;
1379         data_len = xdp->data_end - xdp->data;
1380         skb_reserve(skb, headroom);
1381         __skb_put(skb, data_len);
1382
1383         metasize = xdp->data - xdp->data_meta;
1384         metasize = metasize > 0 ? metasize : 0;
1385         if (metasize)
1386                 skb_metadata_set(skb, metasize);
1387
1388         if (unlikely(xdp_buff_has_frags(xdp)))
1389                 xdp_update_skb_shared_info(skb, nr_frags,
1390                                            sinfo->xdp_frags_size,
1391                                            xdp_frags_truesz,
1392                                            xdp_buff_is_frag_pfmemalloc(xdp));
1393
1394         return skb;
1395 }
1396
1397 /* TODO: build xdp in big mode */
1398 static int virtnet_build_xdp_buff_mrg(struct net_device *dev,
1399                                       struct virtnet_info *vi,
1400                                       struct receive_queue *rq,
1401                                       struct xdp_buff *xdp,
1402                                       void *buf,
1403                                       unsigned int len,
1404                                       unsigned int frame_sz,
1405                                       int *num_buf,
1406                                       unsigned int *xdp_frags_truesize,
1407                                       struct virtnet_rq_stats *stats)
1408 {
1409         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1410         unsigned int headroom, tailroom, room;
1411         unsigned int truesize, cur_frag_size;
1412         struct skb_shared_info *shinfo;
1413         unsigned int xdp_frags_truesz = 0;
1414         struct page *page;
1415         skb_frag_t *frag;
1416         int offset;
1417         void *ctx;
1418
1419         xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
1420         xdp_prepare_buff(xdp, buf - VIRTIO_XDP_HEADROOM,
1421                          VIRTIO_XDP_HEADROOM + vi->hdr_len, len - vi->hdr_len, true);
1422
1423         if (!*num_buf)
1424                 return 0;
1425
1426         if (*num_buf > 1) {
1427                 /* If we want to build multi-buffer xdp, we need
1428                  * to specify that the flags of xdp_buff have the
1429                  * XDP_FLAGS_HAS_FRAG bit.
1430                  */
1431                 if (!xdp_buff_has_frags(xdp))
1432                         xdp_buff_set_frags_flag(xdp);
1433
1434                 shinfo = xdp_get_shared_info_from_buff(xdp);
1435                 shinfo->nr_frags = 0;
1436                 shinfo->xdp_frags_size = 0;
1437         }
1438
1439         if (*num_buf > MAX_SKB_FRAGS + 1)
1440                 return -EINVAL;
1441
1442         while (--*num_buf > 0) {
1443                 buf = virtnet_rq_get_buf(rq, &len, &ctx);
1444                 if (unlikely(!buf)) {
1445                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1446                                  dev->name, *num_buf,
1447                                  virtio16_to_cpu(vi->vdev, hdr->num_buffers));
1448                         DEV_STATS_INC(dev, rx_length_errors);
1449                         goto err;
1450                 }
1451
1452                 u64_stats_add(&stats->bytes, len);
1453                 page = virt_to_head_page(buf);
1454                 offset = buf - page_address(page);
1455
1456                 truesize = mergeable_ctx_to_truesize(ctx);
1457                 headroom = mergeable_ctx_to_headroom(ctx);
1458                 tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1459                 room = SKB_DATA_ALIGN(headroom + tailroom);
1460
1461                 cur_frag_size = truesize;
1462                 xdp_frags_truesz += cur_frag_size;
1463                 if (unlikely(len > truesize - room || cur_frag_size > PAGE_SIZE)) {
1464                         put_page(page);
1465                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1466                                  dev->name, len, (unsigned long)(truesize - room));
1467                         DEV_STATS_INC(dev, rx_length_errors);
1468                         goto err;
1469                 }
1470
1471                 frag = &shinfo->frags[shinfo->nr_frags++];
1472                 skb_frag_fill_page_desc(frag, page, offset, len);
1473                 if (page_is_pfmemalloc(page))
1474                         xdp_buff_set_frag_pfmemalloc(xdp);
1475
1476                 shinfo->xdp_frags_size += len;
1477         }
1478
1479         *xdp_frags_truesize = xdp_frags_truesz;
1480         return 0;
1481
1482 err:
1483         put_xdp_frags(xdp);
1484         return -EINVAL;
1485 }
1486
1487 static void *mergeable_xdp_get_buf(struct virtnet_info *vi,
1488                                    struct receive_queue *rq,
1489                                    struct bpf_prog *xdp_prog,
1490                                    void *ctx,
1491                                    unsigned int *frame_sz,
1492                                    int *num_buf,
1493                                    struct page **page,
1494                                    int offset,
1495                                    unsigned int *len,
1496                                    struct virtio_net_hdr_mrg_rxbuf *hdr)
1497 {
1498         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1499         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1500         struct page *xdp_page;
1501         unsigned int xdp_room;
1502
1503         /* Transient failure which in theory could occur if
1504          * in-flight packets from before XDP was enabled reach
1505          * the receive path after XDP is loaded.
1506          */
1507         if (unlikely(hdr->hdr.gso_type))
1508                 return NULL;
1509
1510         /* Now XDP core assumes frag size is PAGE_SIZE, but buffers
1511          * with headroom may add hole in truesize, which
1512          * make their length exceed PAGE_SIZE. So we disabled the
1513          * hole mechanism for xdp. See add_recvbuf_mergeable().
1514          */
1515         *frame_sz = truesize;
1516
1517         if (likely(headroom >= virtnet_get_headroom(vi) &&
1518                    (*num_buf == 1 || xdp_prog->aux->xdp_has_frags))) {
1519                 return page_address(*page) + offset;
1520         }
1521
1522         /* This happens when headroom is not enough because
1523          * of the buffer was prefilled before XDP is set.
1524          * This should only happen for the first several packets.
1525          * In fact, vq reset can be used here to help us clean up
1526          * the prefilled buffers, but many existing devices do not
1527          * support it, and we don't want to bother users who are
1528          * using xdp normally.
1529          */
1530         if (!xdp_prog->aux->xdp_has_frags) {
1531                 /* linearize data for XDP */
1532                 xdp_page = xdp_linearize_page(rq, num_buf,
1533                                               *page, offset,
1534                                               VIRTIO_XDP_HEADROOM,
1535                                               len);
1536                 if (!xdp_page)
1537                         return NULL;
1538         } else {
1539                 xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
1540                                           sizeof(struct skb_shared_info));
1541                 if (*len + xdp_room > PAGE_SIZE)
1542                         return NULL;
1543
1544                 xdp_page = alloc_page(GFP_ATOMIC);
1545                 if (!xdp_page)
1546                         return NULL;
1547
1548                 memcpy(page_address(xdp_page) + VIRTIO_XDP_HEADROOM,
1549                        page_address(*page) + offset, *len);
1550         }
1551
1552         *frame_sz = PAGE_SIZE;
1553
1554         put_page(*page);
1555
1556         *page = xdp_page;
1557
1558         return page_address(*page) + VIRTIO_XDP_HEADROOM;
1559 }
1560
1561 static struct sk_buff *receive_mergeable_xdp(struct net_device *dev,
1562                                              struct virtnet_info *vi,
1563                                              struct receive_queue *rq,
1564                                              struct bpf_prog *xdp_prog,
1565                                              void *buf,
1566                                              void *ctx,
1567                                              unsigned int len,
1568                                              unsigned int *xdp_xmit,
1569                                              struct virtnet_rq_stats *stats)
1570 {
1571         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1572         int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1573         struct page *page = virt_to_head_page(buf);
1574         int offset = buf - page_address(page);
1575         unsigned int xdp_frags_truesz = 0;
1576         struct sk_buff *head_skb;
1577         unsigned int frame_sz;
1578         struct xdp_buff xdp;
1579         void *data;
1580         u32 act;
1581         int err;
1582
1583         data = mergeable_xdp_get_buf(vi, rq, xdp_prog, ctx, &frame_sz, &num_buf, &page,
1584                                      offset, &len, hdr);
1585         if (unlikely(!data))
1586                 goto err_xdp;
1587
1588         err = virtnet_build_xdp_buff_mrg(dev, vi, rq, &xdp, data, len, frame_sz,
1589                                          &num_buf, &xdp_frags_truesz, stats);
1590         if (unlikely(err))
1591                 goto err_xdp;
1592
1593         act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
1594
1595         switch (act) {
1596         case XDP_PASS:
1597                 head_skb = build_skb_from_xdp_buff(dev, vi, &xdp, xdp_frags_truesz);
1598                 if (unlikely(!head_skb))
1599                         break;
1600                 return head_skb;
1601
1602         case XDP_TX:
1603         case XDP_REDIRECT:
1604                 return NULL;
1605
1606         default:
1607                 break;
1608         }
1609
1610         put_xdp_frags(&xdp);
1611
1612 err_xdp:
1613         put_page(page);
1614         mergeable_buf_free(rq, num_buf, dev, stats);
1615
1616         u64_stats_inc(&stats->xdp_drops);
1617         u64_stats_inc(&stats->drops);
1618         return NULL;
1619 }
1620
1621 static struct sk_buff *receive_mergeable(struct net_device *dev,
1622                                          struct virtnet_info *vi,
1623                                          struct receive_queue *rq,
1624                                          void *buf,
1625                                          void *ctx,
1626                                          unsigned int len,
1627                                          unsigned int *xdp_xmit,
1628                                          struct virtnet_rq_stats *stats)
1629 {
1630         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1631         int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1632         struct page *page = virt_to_head_page(buf);
1633         int offset = buf - page_address(page);
1634         struct sk_buff *head_skb, *curr_skb;
1635         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1636         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1637         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1638         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1639
1640         head_skb = NULL;
1641         u64_stats_add(&stats->bytes, len - vi->hdr_len);
1642
1643         if (unlikely(len > truesize - room)) {
1644                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1645                          dev->name, len, (unsigned long)(truesize - room));
1646                 DEV_STATS_INC(dev, rx_length_errors);
1647                 goto err_skb;
1648         }
1649
1650         if (unlikely(vi->xdp_enabled)) {
1651                 struct bpf_prog *xdp_prog;
1652
1653                 rcu_read_lock();
1654                 xdp_prog = rcu_dereference(rq->xdp_prog);
1655                 if (xdp_prog) {
1656                         head_skb = receive_mergeable_xdp(dev, vi, rq, xdp_prog, buf, ctx,
1657                                                          len, xdp_xmit, stats);
1658                         rcu_read_unlock();
1659                         return head_skb;
1660                 }
1661                 rcu_read_unlock();
1662         }
1663
1664         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, headroom);
1665         curr_skb = head_skb;
1666
1667         if (unlikely(!curr_skb))
1668                 goto err_skb;
1669         while (--num_buf) {
1670                 int num_skb_frags;
1671
1672                 buf = virtnet_rq_get_buf(rq, &len, &ctx);
1673                 if (unlikely(!buf)) {
1674                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1675                                  dev->name, num_buf,
1676                                  virtio16_to_cpu(vi->vdev,
1677                                                  hdr->num_buffers));
1678                         DEV_STATS_INC(dev, rx_length_errors);
1679                         goto err_buf;
1680                 }
1681
1682                 u64_stats_add(&stats->bytes, len);
1683                 page = virt_to_head_page(buf);
1684
1685                 truesize = mergeable_ctx_to_truesize(ctx);
1686                 headroom = mergeable_ctx_to_headroom(ctx);
1687                 tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1688                 room = SKB_DATA_ALIGN(headroom + tailroom);
1689                 if (unlikely(len > truesize - room)) {
1690                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1691                                  dev->name, len, (unsigned long)(truesize - room));
1692                         DEV_STATS_INC(dev, rx_length_errors);
1693                         goto err_skb;
1694                 }
1695
1696                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1697                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1698                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1699
1700                         if (unlikely(!nskb))
1701                                 goto err_skb;
1702                         if (curr_skb == head_skb)
1703                                 skb_shinfo(curr_skb)->frag_list = nskb;
1704                         else
1705                                 curr_skb->next = nskb;
1706                         curr_skb = nskb;
1707                         head_skb->truesize += nskb->truesize;
1708                         num_skb_frags = 0;
1709                 }
1710                 if (curr_skb != head_skb) {
1711                         head_skb->data_len += len;
1712                         head_skb->len += len;
1713                         head_skb->truesize += truesize;
1714                 }
1715                 offset = buf - page_address(page);
1716                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1717                         put_page(page);
1718                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1719                                              len, truesize);
1720                 } else {
1721                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1722                                         offset, len, truesize);
1723                 }
1724         }
1725
1726         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1727         return head_skb;
1728
1729 err_skb:
1730         put_page(page);
1731         mergeable_buf_free(rq, num_buf, dev, stats);
1732
1733 err_buf:
1734         u64_stats_inc(&stats->drops);
1735         dev_kfree_skb(head_skb);
1736         return NULL;
1737 }
1738
1739 static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash,
1740                                 struct sk_buff *skb)
1741 {
1742         enum pkt_hash_types rss_hash_type;
1743
1744         if (!hdr_hash || !skb)
1745                 return;
1746
1747         switch (__le16_to_cpu(hdr_hash->hash_report)) {
1748         case VIRTIO_NET_HASH_REPORT_TCPv4:
1749         case VIRTIO_NET_HASH_REPORT_UDPv4:
1750         case VIRTIO_NET_HASH_REPORT_TCPv6:
1751         case VIRTIO_NET_HASH_REPORT_UDPv6:
1752         case VIRTIO_NET_HASH_REPORT_TCPv6_EX:
1753         case VIRTIO_NET_HASH_REPORT_UDPv6_EX:
1754                 rss_hash_type = PKT_HASH_TYPE_L4;
1755                 break;
1756         case VIRTIO_NET_HASH_REPORT_IPv4:
1757         case VIRTIO_NET_HASH_REPORT_IPv6:
1758         case VIRTIO_NET_HASH_REPORT_IPv6_EX:
1759                 rss_hash_type = PKT_HASH_TYPE_L3;
1760                 break;
1761         case VIRTIO_NET_HASH_REPORT_NONE:
1762         default:
1763                 rss_hash_type = PKT_HASH_TYPE_NONE;
1764         }
1765         skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type);
1766 }
1767
1768 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1769                         void *buf, unsigned int len, void **ctx,
1770                         unsigned int *xdp_xmit,
1771                         struct virtnet_rq_stats *stats)
1772 {
1773         struct net_device *dev = vi->dev;
1774         struct sk_buff *skb;
1775         struct virtio_net_common_hdr *hdr;
1776
1777         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1778                 pr_debug("%s: short packet %i\n", dev->name, len);
1779                 DEV_STATS_INC(dev, rx_length_errors);
1780                 virtnet_rq_free_buf(vi, rq, buf);
1781                 return;
1782         }
1783
1784         if (vi->mergeable_rx_bufs)
1785                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1786                                         stats);
1787         else if (vi->big_packets)
1788                 skb = receive_big(dev, vi, rq, buf, len, stats);
1789         else
1790                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1791
1792         if (unlikely(!skb))
1793                 return;
1794
1795         hdr = skb_vnet_common_hdr(skb);
1796         if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report)
1797                 virtio_skb_set_hash(&hdr->hash_v1_hdr, skb);
1798
1799         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1800                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1801
1802         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1803                                   virtio_is_little_endian(vi->vdev))) {
1804                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1805                                      dev->name, hdr->hdr.gso_type,
1806                                      hdr->hdr.gso_size);
1807                 goto frame_err;
1808         }
1809
1810         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1811         skb->protocol = eth_type_trans(skb, dev);
1812         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1813                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1814
1815         napi_gro_receive(&rq->napi, skb);
1816         return;
1817
1818 frame_err:
1819         DEV_STATS_INC(dev, rx_frame_errors);
1820         dev_kfree_skb(skb);
1821 }
1822
1823 /* Unlike mergeable buffers, all buffers are allocated to the
1824  * same size, except for the headroom. For this reason we do
1825  * not need to use  mergeable_len_to_ctx here - it is enough
1826  * to store the headroom as the context ignoring the truesize.
1827  */
1828 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1829                              gfp_t gfp)
1830 {
1831         char *buf;
1832         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1833         void *ctx = (void *)(unsigned long)xdp_headroom;
1834         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1835         int err;
1836
1837         len = SKB_DATA_ALIGN(len) +
1838               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1839
1840         buf = virtnet_rq_alloc(rq, len, gfp);
1841         if (unlikely(!buf))
1842                 return -ENOMEM;
1843
1844         virtnet_rq_init_one_sg(rq, buf + VIRTNET_RX_PAD + xdp_headroom,
1845                                vi->hdr_len + GOOD_PACKET_LEN);
1846
1847         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1848         if (err < 0) {
1849                 if (rq->do_dma)
1850                         virtnet_rq_unmap(rq, buf, 0);
1851                 put_page(virt_to_head_page(buf));
1852         }
1853
1854         return err;
1855 }
1856
1857 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1858                            gfp_t gfp)
1859 {
1860         struct page *first, *list = NULL;
1861         char *p;
1862         int i, err, offset;
1863
1864         sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2);
1865
1866         /* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */
1867         for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) {
1868                 first = get_a_page(rq, gfp);
1869                 if (!first) {
1870                         if (list)
1871                                 give_pages(rq, list);
1872                         return -ENOMEM;
1873                 }
1874                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1875
1876                 /* chain new page in list head to match sg */
1877                 first->private = (unsigned long)list;
1878                 list = first;
1879         }
1880
1881         first = get_a_page(rq, gfp);
1882         if (!first) {
1883                 give_pages(rq, list);
1884                 return -ENOMEM;
1885         }
1886         p = page_address(first);
1887
1888         /* rq->sg[0], rq->sg[1] share the same page */
1889         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1890         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1891
1892         /* rq->sg[1] for data packet, from offset */
1893         offset = sizeof(struct padded_vnet_hdr);
1894         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1895
1896         /* chain first in list head */
1897         first->private = (unsigned long)list;
1898         err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2,
1899                                   first, gfp);
1900         if (err < 0)
1901                 give_pages(rq, first);
1902
1903         return err;
1904 }
1905
1906 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1907                                           struct ewma_pkt_len *avg_pkt_len,
1908                                           unsigned int room)
1909 {
1910         struct virtnet_info *vi = rq->vq->vdev->priv;
1911         const size_t hdr_len = vi->hdr_len;
1912         unsigned int len;
1913
1914         if (room)
1915                 return PAGE_SIZE - room;
1916
1917         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1918                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1919
1920         return ALIGN(len, L1_CACHE_BYTES);
1921 }
1922
1923 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1924                                  struct receive_queue *rq, gfp_t gfp)
1925 {
1926         struct page_frag *alloc_frag = &rq->alloc_frag;
1927         unsigned int headroom = virtnet_get_headroom(vi);
1928         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1929         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1930         unsigned int len, hole;
1931         void *ctx;
1932         char *buf;
1933         int err;
1934
1935         /* Extra tailroom is needed to satisfy XDP's assumption. This
1936          * means rx frags coalescing won't work, but consider we've
1937          * disabled GSO for XDP, it won't be a big issue.
1938          */
1939         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1940
1941         buf = virtnet_rq_alloc(rq, len + room, gfp);
1942         if (unlikely(!buf))
1943                 return -ENOMEM;
1944
1945         buf += headroom; /* advance address leaving hole at front of pkt */
1946         hole = alloc_frag->size - alloc_frag->offset;
1947         if (hole < len + room) {
1948                 /* To avoid internal fragmentation, if there is very likely not
1949                  * enough space for another buffer, add the remaining space to
1950                  * the current buffer.
1951                  * XDP core assumes that frame_size of xdp_buff and the length
1952                  * of the frag are PAGE_SIZE, so we disable the hole mechanism.
1953                  */
1954                 if (!headroom)
1955                         len += hole;
1956                 alloc_frag->offset += hole;
1957         }
1958
1959         virtnet_rq_init_one_sg(rq, buf, len);
1960
1961         ctx = mergeable_len_to_ctx(len + room, headroom);
1962         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1963         if (err < 0) {
1964                 if (rq->do_dma)
1965                         virtnet_rq_unmap(rq, buf, 0);
1966                 put_page(virt_to_head_page(buf));
1967         }
1968
1969         return err;
1970 }
1971
1972 /*
1973  * Returns false if we couldn't fill entirely (OOM).
1974  *
1975  * Normally run in the receive path, but can also be run from ndo_open
1976  * before we're receiving packets, or from refill_work which is
1977  * careful to disable receiving (using napi_disable).
1978  */
1979 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1980                           gfp_t gfp)
1981 {
1982         int err;
1983         bool oom;
1984
1985         do {
1986                 if (vi->mergeable_rx_bufs)
1987                         err = add_recvbuf_mergeable(vi, rq, gfp);
1988                 else if (vi->big_packets)
1989                         err = add_recvbuf_big(vi, rq, gfp);
1990                 else
1991                         err = add_recvbuf_small(vi, rq, gfp);
1992
1993                 oom = err == -ENOMEM;
1994                 if (err)
1995                         break;
1996         } while (rq->vq->num_free);
1997         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1998                 unsigned long flags;
1999
2000                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
2001                 u64_stats_inc(&rq->stats.kicks);
2002                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
2003         }
2004
2005         return !oom;
2006 }
2007
2008 static void skb_recv_done(struct virtqueue *rvq)
2009 {
2010         struct virtnet_info *vi = rvq->vdev->priv;
2011         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
2012
2013         virtqueue_napi_schedule(&rq->napi, rvq);
2014 }
2015
2016 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
2017 {
2018         napi_enable(napi);
2019
2020         /* If all buffers were filled by other side before we napi_enabled, we
2021          * won't get another interrupt, so process any outstanding packets now.
2022          * Call local_bh_enable after to trigger softIRQ processing.
2023          */
2024         local_bh_disable();
2025         virtqueue_napi_schedule(napi, vq);
2026         local_bh_enable();
2027 }
2028
2029 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
2030                                    struct virtqueue *vq,
2031                                    struct napi_struct *napi)
2032 {
2033         if (!napi->weight)
2034                 return;
2035
2036         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
2037          * enable the feature if this is likely affine with the transmit path.
2038          */
2039         if (!vi->affinity_hint_set) {
2040                 napi->weight = 0;
2041                 return;
2042         }
2043
2044         return virtnet_napi_enable(vq, napi);
2045 }
2046
2047 static void virtnet_napi_tx_disable(struct napi_struct *napi)
2048 {
2049         if (napi->weight)
2050                 napi_disable(napi);
2051 }
2052
2053 static void refill_work(struct work_struct *work)
2054 {
2055         struct virtnet_info *vi =
2056                 container_of(work, struct virtnet_info, refill.work);
2057         bool still_empty;
2058         int i;
2059
2060         for (i = 0; i < vi->curr_queue_pairs; i++) {
2061                 struct receive_queue *rq = &vi->rq[i];
2062
2063                 napi_disable(&rq->napi);
2064                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
2065                 virtnet_napi_enable(rq->vq, &rq->napi);
2066
2067                 /* In theory, this can happen: if we don't get any buffers in
2068                  * we will *never* try to fill again.
2069                  */
2070                 if (still_empty)
2071                         schedule_delayed_work(&vi->refill, HZ/2);
2072         }
2073 }
2074
2075 static int virtnet_receive(struct receive_queue *rq, int budget,
2076                            unsigned int *xdp_xmit)
2077 {
2078         struct virtnet_info *vi = rq->vq->vdev->priv;
2079         struct virtnet_rq_stats stats = {};
2080         unsigned int len;
2081         int packets = 0;
2082         void *buf;
2083         int i;
2084
2085         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2086                 void *ctx;
2087
2088                 while (packets < budget &&
2089                        (buf = virtnet_rq_get_buf(rq, &len, &ctx))) {
2090                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
2091                         packets++;
2092                 }
2093         } else {
2094                 while (packets < budget &&
2095                        (buf = virtnet_rq_get_buf(rq, &len, NULL)) != NULL) {
2096                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
2097                         packets++;
2098                 }
2099         }
2100
2101         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
2102                 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
2103                         spin_lock(&vi->refill_lock);
2104                         if (vi->refill_enabled)
2105                                 schedule_delayed_work(&vi->refill, 0);
2106                         spin_unlock(&vi->refill_lock);
2107                 }
2108         }
2109
2110         u64_stats_set(&stats.packets, packets);
2111         u64_stats_update_begin(&rq->stats.syncp);
2112         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
2113                 size_t offset = virtnet_rq_stats_desc[i].offset;
2114                 u64_stats_t *item, *src;
2115
2116                 item = (u64_stats_t *)((u8 *)&rq->stats + offset);
2117                 src = (u64_stats_t *)((u8 *)&stats + offset);
2118                 u64_stats_add(item, u64_stats_read(src));
2119         }
2120         u64_stats_update_end(&rq->stats.syncp);
2121
2122         return packets;
2123 }
2124
2125 static void virtnet_poll_cleantx(struct receive_queue *rq)
2126 {
2127         struct virtnet_info *vi = rq->vq->vdev->priv;
2128         unsigned int index = vq2rxq(rq->vq);
2129         struct send_queue *sq = &vi->sq[index];
2130         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
2131
2132         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
2133                 return;
2134
2135         if (__netif_tx_trylock(txq)) {
2136                 if (sq->reset) {
2137                         __netif_tx_unlock(txq);
2138                         return;
2139                 }
2140
2141                 do {
2142                         virtqueue_disable_cb(sq->vq);
2143                         free_old_xmit_skbs(sq, true);
2144                 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2145
2146                 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
2147                         netif_tx_wake_queue(txq);
2148
2149                 __netif_tx_unlock(txq);
2150         }
2151 }
2152
2153 static int virtnet_poll(struct napi_struct *napi, int budget)
2154 {
2155         struct receive_queue *rq =
2156                 container_of(napi, struct receive_queue, napi);
2157         struct virtnet_info *vi = rq->vq->vdev->priv;
2158         struct send_queue *sq;
2159         unsigned int received;
2160         unsigned int xdp_xmit = 0;
2161
2162         virtnet_poll_cleantx(rq);
2163
2164         received = virtnet_receive(rq, budget, &xdp_xmit);
2165
2166         if (xdp_xmit & VIRTIO_XDP_REDIR)
2167                 xdp_do_flush();
2168
2169         /* Out of packets? */
2170         if (received < budget)
2171                 virtqueue_napi_complete(napi, rq->vq, received);
2172
2173         if (xdp_xmit & VIRTIO_XDP_TX) {
2174                 sq = virtnet_xdp_get_sq(vi);
2175                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2176                         u64_stats_update_begin(&sq->stats.syncp);
2177                         u64_stats_inc(&sq->stats.kicks);
2178                         u64_stats_update_end(&sq->stats.syncp);
2179                 }
2180                 virtnet_xdp_put_sq(vi, sq);
2181         }
2182
2183         return received;
2184 }
2185
2186 static void virtnet_disable_queue_pair(struct virtnet_info *vi, int qp_index)
2187 {
2188         virtnet_napi_tx_disable(&vi->sq[qp_index].napi);
2189         napi_disable(&vi->rq[qp_index].napi);
2190         xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
2191 }
2192
2193 static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
2194 {
2195         struct net_device *dev = vi->dev;
2196         int err;
2197
2198         err = xdp_rxq_info_reg(&vi->rq[qp_index].xdp_rxq, dev, qp_index,
2199                                vi->rq[qp_index].napi.napi_id);
2200         if (err < 0)
2201                 return err;
2202
2203         err = xdp_rxq_info_reg_mem_model(&vi->rq[qp_index].xdp_rxq,
2204                                          MEM_TYPE_PAGE_SHARED, NULL);
2205         if (err < 0)
2206                 goto err_xdp_reg_mem_model;
2207
2208         virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
2209         virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);
2210
2211         return 0;
2212
2213 err_xdp_reg_mem_model:
2214         xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
2215         return err;
2216 }
2217
2218 static int virtnet_open(struct net_device *dev)
2219 {
2220         struct virtnet_info *vi = netdev_priv(dev);
2221         int i, err;
2222
2223         enable_delayed_refill(vi);
2224
2225         for (i = 0; i < vi->max_queue_pairs; i++) {
2226                 if (i < vi->curr_queue_pairs)
2227                         /* Make sure we have some buffers: if oom use wq. */
2228                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2229                                 schedule_delayed_work(&vi->refill, 0);
2230
2231                 err = virtnet_enable_queue_pair(vi, i);
2232                 if (err < 0)
2233                         goto err_enable_qp;
2234         }
2235
2236         return 0;
2237
2238 err_enable_qp:
2239         disable_delayed_refill(vi);
2240         cancel_delayed_work_sync(&vi->refill);
2241
2242         for (i--; i >= 0; i--)
2243                 virtnet_disable_queue_pair(vi, i);
2244         return err;
2245 }
2246
2247 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
2248 {
2249         struct send_queue *sq = container_of(napi, struct send_queue, napi);
2250         struct virtnet_info *vi = sq->vq->vdev->priv;
2251         unsigned int index = vq2txq(sq->vq);
2252         struct netdev_queue *txq;
2253         int opaque;
2254         bool done;
2255
2256         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
2257                 /* We don't need to enable cb for XDP */
2258                 napi_complete_done(napi, 0);
2259                 return 0;
2260         }
2261
2262         txq = netdev_get_tx_queue(vi->dev, index);
2263         __netif_tx_lock(txq, raw_smp_processor_id());
2264         virtqueue_disable_cb(sq->vq);
2265         free_old_xmit_skbs(sq, true);
2266
2267         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
2268                 netif_tx_wake_queue(txq);
2269
2270         opaque = virtqueue_enable_cb_prepare(sq->vq);
2271
2272         done = napi_complete_done(napi, 0);
2273
2274         if (!done)
2275                 virtqueue_disable_cb(sq->vq);
2276
2277         __netif_tx_unlock(txq);
2278
2279         if (done) {
2280                 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
2281                         if (napi_schedule_prep(napi)) {
2282                                 __netif_tx_lock(txq, raw_smp_processor_id());
2283                                 virtqueue_disable_cb(sq->vq);
2284                                 __netif_tx_unlock(txq);
2285                                 __napi_schedule(napi);
2286                         }
2287                 }
2288         }
2289
2290         return 0;
2291 }
2292
2293 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
2294 {
2295         struct virtio_net_hdr_mrg_rxbuf *hdr;
2296         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
2297         struct virtnet_info *vi = sq->vq->vdev->priv;
2298         int num_sg;
2299         unsigned hdr_len = vi->hdr_len;
2300         bool can_push;
2301
2302         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
2303
2304         can_push = vi->any_header_sg &&
2305                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
2306                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
2307         /* Even if we can, don't push here yet as this would skew
2308          * csum_start offset below. */
2309         if (can_push)
2310                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
2311         else
2312                 hdr = &skb_vnet_common_hdr(skb)->mrg_hdr;
2313
2314         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
2315                                     virtio_is_little_endian(vi->vdev), false,
2316                                     0))
2317                 return -EPROTO;
2318
2319         if (vi->mergeable_rx_bufs)
2320                 hdr->num_buffers = 0;
2321
2322         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
2323         if (can_push) {
2324                 __skb_push(skb, hdr_len);
2325                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
2326                 if (unlikely(num_sg < 0))
2327                         return num_sg;
2328                 /* Pull header back to avoid skew in tx bytes calculations. */
2329                 __skb_pull(skb, hdr_len);
2330         } else {
2331                 sg_set_buf(sq->sg, hdr, hdr_len);
2332                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
2333                 if (unlikely(num_sg < 0))
2334                         return num_sg;
2335                 num_sg++;
2336         }
2337         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
2338 }
2339
2340 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
2341 {
2342         struct virtnet_info *vi = netdev_priv(dev);
2343         int qnum = skb_get_queue_mapping(skb);
2344         struct send_queue *sq = &vi->sq[qnum];
2345         int err;
2346         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
2347         bool kick = !netdev_xmit_more();
2348         bool use_napi = sq->napi.weight;
2349
2350         /* Free up any pending old buffers before queueing new ones. */
2351         do {
2352                 if (use_napi)
2353                         virtqueue_disable_cb(sq->vq);
2354
2355                 free_old_xmit_skbs(sq, false);
2356
2357         } while (use_napi && kick &&
2358                unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2359
2360         /* timestamp packet in software */
2361         skb_tx_timestamp(skb);
2362
2363         /* Try to transmit */
2364         err = xmit_skb(sq, skb);
2365
2366         /* This should not happen! */
2367         if (unlikely(err)) {
2368                 DEV_STATS_INC(dev, tx_fifo_errors);
2369                 if (net_ratelimit())
2370                         dev_warn(&dev->dev,
2371                                  "Unexpected TXQ (%d) queue failure: %d\n",
2372                                  qnum, err);
2373                 DEV_STATS_INC(dev, tx_dropped);
2374                 dev_kfree_skb_any(skb);
2375                 return NETDEV_TX_OK;
2376         }
2377
2378         /* Don't wait up for transmitted skbs to be freed. */
2379         if (!use_napi) {
2380                 skb_orphan(skb);
2381                 nf_reset_ct(skb);
2382         }
2383
2384         check_sq_full_and_disable(vi, dev, sq);
2385
2386         if (kick || netif_xmit_stopped(txq)) {
2387                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2388                         u64_stats_update_begin(&sq->stats.syncp);
2389                         u64_stats_inc(&sq->stats.kicks);
2390                         u64_stats_update_end(&sq->stats.syncp);
2391                 }
2392         }
2393
2394         return NETDEV_TX_OK;
2395 }
2396
2397 static int virtnet_rx_resize(struct virtnet_info *vi,
2398                              struct receive_queue *rq, u32 ring_num)
2399 {
2400         bool running = netif_running(vi->dev);
2401         int err, qindex;
2402
2403         qindex = rq - vi->rq;
2404
2405         if (running)
2406                 napi_disable(&rq->napi);
2407
2408         err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_unmap_free_buf);
2409         if (err)
2410                 netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err);
2411
2412         if (!try_fill_recv(vi, rq, GFP_KERNEL))
2413                 schedule_delayed_work(&vi->refill, 0);
2414
2415         if (running)
2416                 virtnet_napi_enable(rq->vq, &rq->napi);
2417         return err;
2418 }
2419
2420 static int virtnet_tx_resize(struct virtnet_info *vi,
2421                              struct send_queue *sq, u32 ring_num)
2422 {
2423         bool running = netif_running(vi->dev);
2424         struct netdev_queue *txq;
2425         int err, qindex;
2426
2427         qindex = sq - vi->sq;
2428
2429         if (running)
2430                 virtnet_napi_tx_disable(&sq->napi);
2431
2432         txq = netdev_get_tx_queue(vi->dev, qindex);
2433
2434         /* 1. wait all ximt complete
2435          * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue()
2436          */
2437         __netif_tx_lock_bh(txq);
2438
2439         /* Prevent rx poll from accessing sq. */
2440         sq->reset = true;
2441
2442         /* Prevent the upper layer from trying to send packets. */
2443         netif_stop_subqueue(vi->dev, qindex);
2444
2445         __netif_tx_unlock_bh(txq);
2446
2447         err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf);
2448         if (err)
2449                 netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err);
2450
2451         __netif_tx_lock_bh(txq);
2452         sq->reset = false;
2453         netif_tx_wake_queue(txq);
2454         __netif_tx_unlock_bh(txq);
2455
2456         if (running)
2457                 virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
2458         return err;
2459 }
2460
2461 /*
2462  * Send command via the control virtqueue and check status.  Commands
2463  * supported by the hypervisor, as indicated by feature bits, should
2464  * never fail unless improperly formatted.
2465  */
2466 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
2467                                  struct scatterlist *out)
2468 {
2469         struct scatterlist *sgs[4], hdr, stat;
2470         unsigned out_num = 0, tmp;
2471         int ret;
2472
2473         /* Caller should know better */
2474         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
2475
2476         vi->ctrl->status = ~0;
2477         vi->ctrl->hdr.class = class;
2478         vi->ctrl->hdr.cmd = cmd;
2479         /* Add header */
2480         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
2481         sgs[out_num++] = &hdr;
2482
2483         if (out)
2484                 sgs[out_num++] = out;
2485
2486         /* Add return status. */
2487         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
2488         sgs[out_num] = &stat;
2489
2490         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
2491         ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
2492         if (ret < 0) {
2493                 dev_warn(&vi->vdev->dev,
2494                          "Failed to add sgs for command vq: %d\n.", ret);
2495                 return false;
2496         }
2497
2498         if (unlikely(!virtqueue_kick(vi->cvq)))
2499                 return vi->ctrl->status == VIRTIO_NET_OK;
2500
2501         /* Spin for a response, the kick causes an ioport write, trapping
2502          * into the hypervisor, so the request should be handled immediately.
2503          */
2504         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
2505                !virtqueue_is_broken(vi->cvq))
2506                 cpu_relax();
2507
2508         return vi->ctrl->status == VIRTIO_NET_OK;
2509 }
2510
2511 static int virtnet_set_mac_address(struct net_device *dev, void *p)
2512 {
2513         struct virtnet_info *vi = netdev_priv(dev);
2514         struct virtio_device *vdev = vi->vdev;
2515         int ret;
2516         struct sockaddr *addr;
2517         struct scatterlist sg;
2518
2519         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2520                 return -EOPNOTSUPP;
2521
2522         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
2523         if (!addr)
2524                 return -ENOMEM;
2525
2526         ret = eth_prepare_mac_addr_change(dev, addr);
2527         if (ret)
2528                 goto out;
2529
2530         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
2531                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
2532                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2533                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
2534                         dev_warn(&vdev->dev,
2535                                  "Failed to set mac address by vq command.\n");
2536                         ret = -EINVAL;
2537                         goto out;
2538                 }
2539         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
2540                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
2541                 unsigned int i;
2542
2543                 /* Naturally, this has an atomicity problem. */
2544                 for (i = 0; i < dev->addr_len; i++)
2545                         virtio_cwrite8(vdev,
2546                                        offsetof(struct virtio_net_config, mac) +
2547                                        i, addr->sa_data[i]);
2548         }
2549
2550         eth_commit_mac_addr_change(dev, p);
2551         ret = 0;
2552
2553 out:
2554         kfree(addr);
2555         return ret;
2556 }
2557
2558 static void virtnet_stats(struct net_device *dev,
2559                           struct rtnl_link_stats64 *tot)
2560 {
2561         struct virtnet_info *vi = netdev_priv(dev);
2562         unsigned int start;
2563         int i;
2564
2565         for (i = 0; i < vi->max_queue_pairs; i++) {
2566                 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
2567                 struct receive_queue *rq = &vi->rq[i];
2568                 struct send_queue *sq = &vi->sq[i];
2569
2570                 do {
2571                         start = u64_stats_fetch_begin(&sq->stats.syncp);
2572                         tpackets = u64_stats_read(&sq->stats.packets);
2573                         tbytes   = u64_stats_read(&sq->stats.bytes);
2574                         terrors  = u64_stats_read(&sq->stats.tx_timeouts);
2575                 } while (u64_stats_fetch_retry(&sq->stats.syncp, start));
2576
2577                 do {
2578                         start = u64_stats_fetch_begin(&rq->stats.syncp);
2579                         rpackets = u64_stats_read(&rq->stats.packets);
2580                         rbytes   = u64_stats_read(&rq->stats.bytes);
2581                         rdrops   = u64_stats_read(&rq->stats.drops);
2582                 } while (u64_stats_fetch_retry(&rq->stats.syncp, start));
2583
2584                 tot->rx_packets += rpackets;
2585                 tot->tx_packets += tpackets;
2586                 tot->rx_bytes   += rbytes;
2587                 tot->tx_bytes   += tbytes;
2588                 tot->rx_dropped += rdrops;
2589                 tot->tx_errors  += terrors;
2590         }
2591
2592         tot->tx_dropped = DEV_STATS_READ(dev, tx_dropped);
2593         tot->tx_fifo_errors = DEV_STATS_READ(dev, tx_fifo_errors);
2594         tot->rx_length_errors = DEV_STATS_READ(dev, rx_length_errors);
2595         tot->rx_frame_errors = DEV_STATS_READ(dev, rx_frame_errors);
2596 }
2597
2598 static void virtnet_ack_link_announce(struct virtnet_info *vi)
2599 {
2600         rtnl_lock();
2601         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
2602                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
2603                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
2604         rtnl_unlock();
2605 }
2606
2607 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2608 {
2609         struct scatterlist sg;
2610         struct net_device *dev = vi->dev;
2611
2612         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
2613                 return 0;
2614
2615         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
2616         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
2617
2618         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2619                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
2620                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
2621                          queue_pairs);
2622                 return -EINVAL;
2623         } else {
2624                 vi->curr_queue_pairs = queue_pairs;
2625                 /* virtnet_open() will refill when device is going to up. */
2626                 if (dev->flags & IFF_UP)
2627                         schedule_delayed_work(&vi->refill, 0);
2628         }
2629
2630         return 0;
2631 }
2632
2633 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2634 {
2635         int err;
2636
2637         rtnl_lock();
2638         err = _virtnet_set_queues(vi, queue_pairs);
2639         rtnl_unlock();
2640         return err;
2641 }
2642
2643 static int virtnet_close(struct net_device *dev)
2644 {
2645         struct virtnet_info *vi = netdev_priv(dev);
2646         int i;
2647
2648         /* Make sure NAPI doesn't schedule refill work */
2649         disable_delayed_refill(vi);
2650         /* Make sure refill_work doesn't re-enable napi! */
2651         cancel_delayed_work_sync(&vi->refill);
2652
2653         for (i = 0; i < vi->max_queue_pairs; i++)
2654                 virtnet_disable_queue_pair(vi, i);
2655
2656         return 0;
2657 }
2658
2659 static void virtnet_set_rx_mode(struct net_device *dev)
2660 {
2661         struct virtnet_info *vi = netdev_priv(dev);
2662         struct scatterlist sg[2];
2663         struct virtio_net_ctrl_mac *mac_data;
2664         struct netdev_hw_addr *ha;
2665         int uc_count;
2666         int mc_count;
2667         void *buf;
2668         int i;
2669
2670         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
2671         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
2672                 return;
2673
2674         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
2675         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
2676
2677         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
2678
2679         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2680                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
2681                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
2682                          vi->ctrl->promisc ? "en" : "dis");
2683
2684         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
2685
2686         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2687                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2688                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2689                          vi->ctrl->allmulti ? "en" : "dis");
2690
2691         uc_count = netdev_uc_count(dev);
2692         mc_count = netdev_mc_count(dev);
2693         /* MAC filter - use one buffer for both lists */
2694         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2695                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2696         mac_data = buf;
2697         if (!buf)
2698                 return;
2699
2700         sg_init_table(sg, 2);
2701
2702         /* Store the unicast list and count in the front of the buffer */
2703         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2704         i = 0;
2705         netdev_for_each_uc_addr(ha, dev)
2706                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2707
2708         sg_set_buf(&sg[0], mac_data,
2709                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2710
2711         /* multicast list and count fill the end */
2712         mac_data = (void *)&mac_data->macs[uc_count][0];
2713
2714         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2715         i = 0;
2716         netdev_for_each_mc_addr(ha, dev)
2717                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2718
2719         sg_set_buf(&sg[1], mac_data,
2720                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2721
2722         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2723                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2724                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2725
2726         kfree(buf);
2727 }
2728
2729 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2730                                    __be16 proto, u16 vid)
2731 {
2732         struct virtnet_info *vi = netdev_priv(dev);
2733         struct scatterlist sg;
2734
2735         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2736         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2737
2738         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2739                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2740                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2741         return 0;
2742 }
2743
2744 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2745                                     __be16 proto, u16 vid)
2746 {
2747         struct virtnet_info *vi = netdev_priv(dev);
2748         struct scatterlist sg;
2749
2750         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2751         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2752
2753         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2754                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2755                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2756         return 0;
2757 }
2758
2759 static void virtnet_clean_affinity(struct virtnet_info *vi)
2760 {
2761         int i;
2762
2763         if (vi->affinity_hint_set) {
2764                 for (i = 0; i < vi->max_queue_pairs; i++) {
2765                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
2766                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
2767                 }
2768
2769                 vi->affinity_hint_set = false;
2770         }
2771 }
2772
2773 static void virtnet_set_affinity(struct virtnet_info *vi)
2774 {
2775         cpumask_var_t mask;
2776         int stragglers;
2777         int group_size;
2778         int i, j, cpu;
2779         int num_cpu;
2780         int stride;
2781
2782         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2783                 virtnet_clean_affinity(vi);
2784                 return;
2785         }
2786
2787         num_cpu = num_online_cpus();
2788         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2789         stragglers = num_cpu >= vi->curr_queue_pairs ?
2790                         num_cpu % vi->curr_queue_pairs :
2791                         0;
2792         cpu = cpumask_first(cpu_online_mask);
2793
2794         for (i = 0; i < vi->curr_queue_pairs; i++) {
2795                 group_size = stride + (i < stragglers ? 1 : 0);
2796
2797                 for (j = 0; j < group_size; j++) {
2798                         cpumask_set_cpu(cpu, mask);
2799                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2800                                                 nr_cpu_ids, false);
2801                 }
2802                 virtqueue_set_affinity(vi->rq[i].vq, mask);
2803                 virtqueue_set_affinity(vi->sq[i].vq, mask);
2804                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2805                 cpumask_clear(mask);
2806         }
2807
2808         vi->affinity_hint_set = true;
2809         free_cpumask_var(mask);
2810 }
2811
2812 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2813 {
2814         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2815                                                    node);
2816         virtnet_set_affinity(vi);
2817         return 0;
2818 }
2819
2820 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2821 {
2822         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2823                                                    node_dead);
2824         virtnet_set_affinity(vi);
2825         return 0;
2826 }
2827
2828 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2829 {
2830         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2831                                                    node);
2832
2833         virtnet_clean_affinity(vi);
2834         return 0;
2835 }
2836
2837 static enum cpuhp_state virtionet_online;
2838
2839 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2840 {
2841         int ret;
2842
2843         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2844         if (ret)
2845                 return ret;
2846         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2847                                                &vi->node_dead);
2848         if (!ret)
2849                 return ret;
2850         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2851         return ret;
2852 }
2853
2854 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2855 {
2856         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2857         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2858                                             &vi->node_dead);
2859 }
2860
2861 static void virtnet_get_ringparam(struct net_device *dev,
2862                                   struct ethtool_ringparam *ring,
2863                                   struct kernel_ethtool_ringparam *kernel_ring,
2864                                   struct netlink_ext_ack *extack)
2865 {
2866         struct virtnet_info *vi = netdev_priv(dev);
2867
2868         ring->rx_max_pending = vi->rq[0].vq->num_max;
2869         ring->tx_max_pending = vi->sq[0].vq->num_max;
2870         ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2871         ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2872 }
2873
2874 static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
2875                                          u16 vqn, u32 max_usecs, u32 max_packets);
2876
2877 static int virtnet_set_ringparam(struct net_device *dev,
2878                                  struct ethtool_ringparam *ring,
2879                                  struct kernel_ethtool_ringparam *kernel_ring,
2880                                  struct netlink_ext_ack *extack)
2881 {
2882         struct virtnet_info *vi = netdev_priv(dev);
2883         u32 rx_pending, tx_pending;
2884         struct receive_queue *rq;
2885         struct send_queue *sq;
2886         int i, err;
2887
2888         if (ring->rx_mini_pending || ring->rx_jumbo_pending)
2889                 return -EINVAL;
2890
2891         rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2892         tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2893
2894         if (ring->rx_pending == rx_pending &&
2895             ring->tx_pending == tx_pending)
2896                 return 0;
2897
2898         if (ring->rx_pending > vi->rq[0].vq->num_max)
2899                 return -EINVAL;
2900
2901         if (ring->tx_pending > vi->sq[0].vq->num_max)
2902                 return -EINVAL;
2903
2904         for (i = 0; i < vi->max_queue_pairs; i++) {
2905                 rq = vi->rq + i;
2906                 sq = vi->sq + i;
2907
2908                 if (ring->tx_pending != tx_pending) {
2909                         err = virtnet_tx_resize(vi, sq, ring->tx_pending);
2910                         if (err)
2911                                 return err;
2912
2913                         /* Upon disabling and re-enabling a transmit virtqueue, the device must
2914                          * set the coalescing parameters of the virtqueue to those configured
2915                          * through the VIRTIO_NET_CTRL_NOTF_COAL_TX_SET command, or, if the driver
2916                          * did not set any TX coalescing parameters, to 0.
2917                          */
2918                         err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(i),
2919                                                             vi->intr_coal_tx.max_usecs,
2920                                                             vi->intr_coal_tx.max_packets);
2921                         if (err)
2922                                 return err;
2923
2924                         vi->sq[i].intr_coal.max_usecs = vi->intr_coal_tx.max_usecs;
2925                         vi->sq[i].intr_coal.max_packets = vi->intr_coal_tx.max_packets;
2926                 }
2927
2928                 if (ring->rx_pending != rx_pending) {
2929                         err = virtnet_rx_resize(vi, rq, ring->rx_pending);
2930                         if (err)
2931                                 return err;
2932
2933                         /* The reason is same as the transmit virtqueue reset */
2934                         err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(i),
2935                                                             vi->intr_coal_rx.max_usecs,
2936                                                             vi->intr_coal_rx.max_packets);
2937                         if (err)
2938                                 return err;
2939
2940                         vi->rq[i].intr_coal.max_usecs = vi->intr_coal_rx.max_usecs;
2941                         vi->rq[i].intr_coal.max_packets = vi->intr_coal_rx.max_packets;
2942                 }
2943         }
2944
2945         return 0;
2946 }
2947
2948 static bool virtnet_commit_rss_command(struct virtnet_info *vi)
2949 {
2950         struct net_device *dev = vi->dev;
2951         struct scatterlist sgs[4];
2952         unsigned int sg_buf_size;
2953
2954         /* prepare sgs */
2955         sg_init_table(sgs, 4);
2956
2957         sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table);
2958         sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size);
2959
2960         sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1);
2961         sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size);
2962
2963         sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key)
2964                         - offsetof(struct virtio_net_ctrl_rss, max_tx_vq);
2965         sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size);
2966
2967         sg_buf_size = vi->rss_key_size;
2968         sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size);
2969
2970         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2971                                   vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG
2972                                   : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) {
2973                 dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n");
2974                 return false;
2975         }
2976         return true;
2977 }
2978
2979 static void virtnet_init_default_rss(struct virtnet_info *vi)
2980 {
2981         u32 indir_val = 0;
2982         int i = 0;
2983
2984         vi->ctrl->rss.hash_types = vi->rss_hash_types_supported;
2985         vi->rss_hash_types_saved = vi->rss_hash_types_supported;
2986         vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size
2987                                                 ? vi->rss_indir_table_size - 1 : 0;
2988         vi->ctrl->rss.unclassified_queue = 0;
2989
2990         for (; i < vi->rss_indir_table_size; ++i) {
2991                 indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs);
2992                 vi->ctrl->rss.indirection_table[i] = indir_val;
2993         }
2994
2995         vi->ctrl->rss.max_tx_vq = vi->has_rss ? vi->curr_queue_pairs : 0;
2996         vi->ctrl->rss.hash_key_length = vi->rss_key_size;
2997
2998         netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size);
2999 }
3000
3001 static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info)
3002 {
3003         info->data = 0;
3004         switch (info->flow_type) {
3005         case TCP_V4_FLOW:
3006                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) {
3007                         info->data = RXH_IP_SRC | RXH_IP_DST |
3008                                                  RXH_L4_B_0_1 | RXH_L4_B_2_3;
3009                 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
3010                         info->data = RXH_IP_SRC | RXH_IP_DST;
3011                 }
3012                 break;
3013         case TCP_V6_FLOW:
3014                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) {
3015                         info->data = RXH_IP_SRC | RXH_IP_DST |
3016                                                  RXH_L4_B_0_1 | RXH_L4_B_2_3;
3017                 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
3018                         info->data = RXH_IP_SRC | RXH_IP_DST;
3019                 }
3020                 break;
3021         case UDP_V4_FLOW:
3022                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) {
3023                         info->data = RXH_IP_SRC | RXH_IP_DST |
3024                                                  RXH_L4_B_0_1 | RXH_L4_B_2_3;
3025                 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
3026                         info->data = RXH_IP_SRC | RXH_IP_DST;
3027                 }
3028                 break;
3029         case UDP_V6_FLOW:
3030                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) {
3031                         info->data = RXH_IP_SRC | RXH_IP_DST |
3032                                                  RXH_L4_B_0_1 | RXH_L4_B_2_3;
3033                 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
3034                         info->data = RXH_IP_SRC | RXH_IP_DST;
3035                 }
3036                 break;
3037         case IPV4_FLOW:
3038                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4)
3039                         info->data = RXH_IP_SRC | RXH_IP_DST;
3040
3041                 break;
3042         case IPV6_FLOW:
3043                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6)
3044                         info->data = RXH_IP_SRC | RXH_IP_DST;
3045
3046                 break;
3047         default:
3048                 info->data = 0;
3049                 break;
3050         }
3051 }
3052
3053 static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info)
3054 {
3055         u32 new_hashtypes = vi->rss_hash_types_saved;
3056         bool is_disable = info->data & RXH_DISCARD;
3057         bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3);
3058
3059         /* supports only 'sd', 'sdfn' and 'r' */
3060         if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable))
3061                 return false;
3062
3063         switch (info->flow_type) {
3064         case TCP_V4_FLOW:
3065                 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4);
3066                 if (!is_disable)
3067                         new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
3068                                 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0);
3069                 break;
3070         case UDP_V4_FLOW:
3071                 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4);
3072                 if (!is_disable)
3073                         new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
3074                                 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0);
3075                 break;
3076         case IPV4_FLOW:
3077                 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4;
3078                 if (!is_disable)
3079                         new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4;
3080                 break;
3081         case TCP_V6_FLOW:
3082                 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6);
3083                 if (!is_disable)
3084                         new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
3085                                 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0);
3086                 break;
3087         case UDP_V6_FLOW:
3088                 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6);
3089                 if (!is_disable)
3090                         new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
3091                                 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0);
3092                 break;
3093         case IPV6_FLOW:
3094                 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6;
3095                 if (!is_disable)
3096                         new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6;
3097                 break;
3098         default:
3099                 /* unsupported flow */
3100                 return false;
3101         }
3102
3103         /* if unsupported hashtype was set */
3104         if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported))
3105                 return false;
3106
3107         if (new_hashtypes != vi->rss_hash_types_saved) {
3108                 vi->rss_hash_types_saved = new_hashtypes;
3109                 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
3110                 if (vi->dev->features & NETIF_F_RXHASH)
3111                         return virtnet_commit_rss_command(vi);
3112         }
3113
3114         return true;
3115 }
3116
3117 static void virtnet_get_drvinfo(struct net_device *dev,
3118                                 struct ethtool_drvinfo *info)
3119 {
3120         struct virtnet_info *vi = netdev_priv(dev);
3121         struct virtio_device *vdev = vi->vdev;
3122
3123         strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
3124         strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
3125         strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
3126
3127 }
3128
3129 /* TODO: Eliminate OOO packets during switching */
3130 static int virtnet_set_channels(struct net_device *dev,
3131                                 struct ethtool_channels *channels)
3132 {
3133         struct virtnet_info *vi = netdev_priv(dev);
3134         u16 queue_pairs = channels->combined_count;
3135         int err;
3136
3137         /* We don't support separate rx/tx channels.
3138          * We don't allow setting 'other' channels.
3139          */
3140         if (channels->rx_count || channels->tx_count || channels->other_count)
3141                 return -EINVAL;
3142
3143         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
3144                 return -EINVAL;
3145
3146         /* For now we don't support modifying channels while XDP is loaded
3147          * also when XDP is loaded all RX queues have XDP programs so we only
3148          * need to check a single RX queue.
3149          */
3150         if (vi->rq[0].xdp_prog)
3151                 return -EINVAL;
3152
3153         cpus_read_lock();
3154         err = _virtnet_set_queues(vi, queue_pairs);
3155         if (err) {
3156                 cpus_read_unlock();
3157                 goto err;
3158         }
3159         virtnet_set_affinity(vi);
3160         cpus_read_unlock();
3161
3162         netif_set_real_num_tx_queues(dev, queue_pairs);
3163         netif_set_real_num_rx_queues(dev, queue_pairs);
3164  err:
3165         return err;
3166 }
3167
3168 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
3169 {
3170         struct virtnet_info *vi = netdev_priv(dev);
3171         unsigned int i, j;
3172         u8 *p = data;
3173
3174         switch (stringset) {
3175         case ETH_SS_STATS:
3176                 for (i = 0; i < vi->curr_queue_pairs; i++) {
3177                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
3178                                 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
3179                                                 virtnet_rq_stats_desc[j].desc);
3180                 }
3181
3182                 for (i = 0; i < vi->curr_queue_pairs; i++) {
3183                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
3184                                 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
3185                                                 virtnet_sq_stats_desc[j].desc);
3186                 }
3187                 break;
3188         }
3189 }
3190
3191 static int virtnet_get_sset_count(struct net_device *dev, int sset)
3192 {
3193         struct virtnet_info *vi = netdev_priv(dev);
3194
3195         switch (sset) {
3196         case ETH_SS_STATS:
3197                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
3198                                                VIRTNET_SQ_STATS_LEN);
3199         default:
3200                 return -EOPNOTSUPP;
3201         }
3202 }
3203
3204 static void virtnet_get_ethtool_stats(struct net_device *dev,
3205                                       struct ethtool_stats *stats, u64 *data)
3206 {
3207         struct virtnet_info *vi = netdev_priv(dev);
3208         unsigned int idx = 0, start, i, j;
3209         const u8 *stats_base;
3210         const u64_stats_t *p;
3211         size_t offset;
3212
3213         for (i = 0; i < vi->curr_queue_pairs; i++) {
3214                 struct receive_queue *rq = &vi->rq[i];
3215
3216                 stats_base = (const u8 *)&rq->stats;
3217                 do {
3218                         start = u64_stats_fetch_begin(&rq->stats.syncp);
3219                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
3220                                 offset = virtnet_rq_stats_desc[j].offset;
3221                                 p = (const u64_stats_t *)(stats_base + offset);
3222                                 data[idx + j] = u64_stats_read(p);
3223                         }
3224                 } while (u64_stats_fetch_retry(&rq->stats.syncp, start));
3225                 idx += VIRTNET_RQ_STATS_LEN;
3226         }
3227
3228         for (i = 0; i < vi->curr_queue_pairs; i++) {
3229                 struct send_queue *sq = &vi->sq[i];
3230
3231                 stats_base = (const u8 *)&sq->stats;
3232                 do {
3233                         start = u64_stats_fetch_begin(&sq->stats.syncp);
3234                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
3235                                 offset = virtnet_sq_stats_desc[j].offset;
3236                                 p = (const u64_stats_t *)(stats_base + offset);
3237                                 data[idx + j] = u64_stats_read(p);
3238                         }
3239                 } while (u64_stats_fetch_retry(&sq->stats.syncp, start));
3240                 idx += VIRTNET_SQ_STATS_LEN;
3241         }
3242 }
3243
3244 static void virtnet_get_channels(struct net_device *dev,
3245                                  struct ethtool_channels *channels)
3246 {
3247         struct virtnet_info *vi = netdev_priv(dev);
3248
3249         channels->combined_count = vi->curr_queue_pairs;
3250         channels->max_combined = vi->max_queue_pairs;
3251         channels->max_other = 0;
3252         channels->rx_count = 0;
3253         channels->tx_count = 0;
3254         channels->other_count = 0;
3255 }
3256
3257 static int virtnet_set_link_ksettings(struct net_device *dev,
3258                                       const struct ethtool_link_ksettings *cmd)
3259 {
3260         struct virtnet_info *vi = netdev_priv(dev);
3261
3262         return ethtool_virtdev_set_link_ksettings(dev, cmd,
3263                                                   &vi->speed, &vi->duplex);
3264 }
3265
3266 static int virtnet_get_link_ksettings(struct net_device *dev,
3267                                       struct ethtool_link_ksettings *cmd)
3268 {
3269         struct virtnet_info *vi = netdev_priv(dev);
3270
3271         cmd->base.speed = vi->speed;
3272         cmd->base.duplex = vi->duplex;
3273         cmd->base.port = PORT_OTHER;
3274
3275         return 0;
3276 }
3277
3278 static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
3279                                        struct ethtool_coalesce *ec)
3280 {
3281         struct scatterlist sgs_tx, sgs_rx;
3282         int i;
3283
3284         vi->ctrl->coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
3285         vi->ctrl->coal_tx.tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames);
3286         sg_init_one(&sgs_tx, &vi->ctrl->coal_tx, sizeof(vi->ctrl->coal_tx));
3287
3288         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3289                                   VIRTIO_NET_CTRL_NOTF_COAL_TX_SET,
3290                                   &sgs_tx))
3291                 return -EINVAL;
3292
3293         /* Save parameters */
3294         vi->intr_coal_tx.max_usecs = ec->tx_coalesce_usecs;
3295         vi->intr_coal_tx.max_packets = ec->tx_max_coalesced_frames;
3296         for (i = 0; i < vi->max_queue_pairs; i++) {
3297                 vi->sq[i].intr_coal.max_usecs = ec->tx_coalesce_usecs;
3298                 vi->sq[i].intr_coal.max_packets = ec->tx_max_coalesced_frames;
3299         }
3300
3301         vi->ctrl->coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
3302         vi->ctrl->coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
3303         sg_init_one(&sgs_rx, &vi->ctrl->coal_rx, sizeof(vi->ctrl->coal_rx));
3304
3305         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3306                                   VIRTIO_NET_CTRL_NOTF_COAL_RX_SET,
3307                                   &sgs_rx))
3308                 return -EINVAL;
3309
3310         /* Save parameters */
3311         vi->intr_coal_rx.max_usecs = ec->rx_coalesce_usecs;
3312         vi->intr_coal_rx.max_packets = ec->rx_max_coalesced_frames;
3313         for (i = 0; i < vi->max_queue_pairs; i++) {
3314                 vi->rq[i].intr_coal.max_usecs = ec->rx_coalesce_usecs;
3315                 vi->rq[i].intr_coal.max_packets = ec->rx_max_coalesced_frames;
3316         }
3317
3318         return 0;
3319 }
3320
3321 static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
3322                                          u16 vqn, u32 max_usecs, u32 max_packets)
3323 {
3324         struct scatterlist sgs;
3325
3326         vi->ctrl->coal_vq.vqn = cpu_to_le16(vqn);
3327         vi->ctrl->coal_vq.coal.max_usecs = cpu_to_le32(max_usecs);
3328         vi->ctrl->coal_vq.coal.max_packets = cpu_to_le32(max_packets);
3329         sg_init_one(&sgs, &vi->ctrl->coal_vq, sizeof(vi->ctrl->coal_vq));
3330
3331         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3332                                   VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET,
3333                                   &sgs))
3334                 return -EINVAL;
3335
3336         return 0;
3337 }
3338
3339 static int virtnet_send_notf_coal_vq_cmds(struct virtnet_info *vi,
3340                                           struct ethtool_coalesce *ec,
3341                                           u16 queue)
3342 {
3343         int err;
3344
3345         err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(queue),
3346                                             ec->rx_coalesce_usecs,
3347                                             ec->rx_max_coalesced_frames);
3348         if (err)
3349                 return err;
3350
3351         vi->rq[queue].intr_coal.max_usecs = ec->rx_coalesce_usecs;
3352         vi->rq[queue].intr_coal.max_packets = ec->rx_max_coalesced_frames;
3353
3354         err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(queue),
3355                                             ec->tx_coalesce_usecs,
3356                                             ec->tx_max_coalesced_frames);
3357         if (err)
3358                 return err;
3359
3360         vi->sq[queue].intr_coal.max_usecs = ec->tx_coalesce_usecs;
3361         vi->sq[queue].intr_coal.max_packets = ec->tx_max_coalesced_frames;
3362
3363         return 0;
3364 }
3365
3366 static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
3367 {
3368         /* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL
3369          * feature is negotiated.
3370          */
3371         if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs)
3372                 return -EOPNOTSUPP;
3373
3374         if (ec->tx_max_coalesced_frames > 1 ||
3375             ec->rx_max_coalesced_frames != 1)
3376                 return -EINVAL;
3377
3378         return 0;
3379 }
3380
3381 static int virtnet_should_update_vq_weight(int dev_flags, int weight,
3382                                            int vq_weight, bool *should_update)
3383 {
3384         if (weight ^ vq_weight) {
3385                 if (dev_flags & IFF_UP)
3386                         return -EBUSY;
3387                 *should_update = true;
3388         }
3389
3390         return 0;
3391 }
3392
3393 static int virtnet_set_coalesce(struct net_device *dev,
3394                                 struct ethtool_coalesce *ec,
3395                                 struct kernel_ethtool_coalesce *kernel_coal,
3396                                 struct netlink_ext_ack *extack)
3397 {
3398         struct virtnet_info *vi = netdev_priv(dev);
3399         int ret, queue_number, napi_weight;
3400         bool update_napi = false;
3401
3402         /* Can't change NAPI weight if the link is up */
3403         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
3404         for (queue_number = 0; queue_number < vi->max_queue_pairs; queue_number++) {
3405                 ret = virtnet_should_update_vq_weight(dev->flags, napi_weight,
3406                                                       vi->sq[queue_number].napi.weight,
3407                                                       &update_napi);
3408                 if (ret)
3409                         return ret;
3410
3411                 if (update_napi) {
3412                         /* All queues that belong to [queue_number, vi->max_queue_pairs] will be
3413                          * updated for the sake of simplicity, which might not be necessary
3414                          */
3415                         break;
3416                 }
3417         }
3418
3419         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL))
3420                 ret = virtnet_send_notf_coal_cmds(vi, ec);
3421         else
3422                 ret = virtnet_coal_params_supported(ec);
3423
3424         if (ret)
3425                 return ret;
3426
3427         if (update_napi) {
3428                 for (; queue_number < vi->max_queue_pairs; queue_number++)
3429                         vi->sq[queue_number].napi.weight = napi_weight;
3430         }
3431
3432         return ret;
3433 }
3434
3435 static int virtnet_get_coalesce(struct net_device *dev,
3436                                 struct ethtool_coalesce *ec,
3437                                 struct kernel_ethtool_coalesce *kernel_coal,
3438                                 struct netlink_ext_ack *extack)
3439 {
3440         struct virtnet_info *vi = netdev_priv(dev);
3441
3442         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
3443                 ec->rx_coalesce_usecs = vi->intr_coal_rx.max_usecs;
3444                 ec->tx_coalesce_usecs = vi->intr_coal_tx.max_usecs;
3445                 ec->tx_max_coalesced_frames = vi->intr_coal_tx.max_packets;
3446                 ec->rx_max_coalesced_frames = vi->intr_coal_rx.max_packets;
3447         } else {
3448                 ec->rx_max_coalesced_frames = 1;
3449
3450                 if (vi->sq[0].napi.weight)
3451                         ec->tx_max_coalesced_frames = 1;
3452         }
3453
3454         return 0;
3455 }
3456
3457 static int virtnet_set_per_queue_coalesce(struct net_device *dev,
3458                                           u32 queue,
3459                                           struct ethtool_coalesce *ec)
3460 {
3461         struct virtnet_info *vi = netdev_priv(dev);
3462         int ret, napi_weight;
3463         bool update_napi = false;
3464
3465         if (queue >= vi->max_queue_pairs)
3466                 return -EINVAL;
3467
3468         /* Can't change NAPI weight if the link is up */
3469         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
3470         ret = virtnet_should_update_vq_weight(dev->flags, napi_weight,
3471                                               vi->sq[queue].napi.weight,
3472                                               &update_napi);
3473         if (ret)
3474                 return ret;
3475
3476         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
3477                 ret = virtnet_send_notf_coal_vq_cmds(vi, ec, queue);
3478         else
3479                 ret = virtnet_coal_params_supported(ec);
3480
3481         if (ret)
3482                 return ret;
3483
3484         if (update_napi)
3485                 vi->sq[queue].napi.weight = napi_weight;
3486
3487         return 0;
3488 }
3489
3490 static int virtnet_get_per_queue_coalesce(struct net_device *dev,
3491                                           u32 queue,
3492                                           struct ethtool_coalesce *ec)
3493 {
3494         struct virtnet_info *vi = netdev_priv(dev);
3495
3496         if (queue >= vi->max_queue_pairs)
3497                 return -EINVAL;
3498
3499         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL)) {
3500                 ec->rx_coalesce_usecs = vi->rq[queue].intr_coal.max_usecs;
3501                 ec->tx_coalesce_usecs = vi->sq[queue].intr_coal.max_usecs;
3502                 ec->tx_max_coalesced_frames = vi->sq[queue].intr_coal.max_packets;
3503                 ec->rx_max_coalesced_frames = vi->rq[queue].intr_coal.max_packets;
3504         } else {
3505                 ec->rx_max_coalesced_frames = 1;
3506
3507                 if (vi->sq[queue].napi.weight)
3508                         ec->tx_max_coalesced_frames = 1;
3509         }
3510
3511         return 0;
3512 }
3513
3514 static void virtnet_init_settings(struct net_device *dev)
3515 {
3516         struct virtnet_info *vi = netdev_priv(dev);
3517
3518         vi->speed = SPEED_UNKNOWN;
3519         vi->duplex = DUPLEX_UNKNOWN;
3520 }
3521
3522 static void virtnet_update_settings(struct virtnet_info *vi)
3523 {
3524         u32 speed;
3525         u8 duplex;
3526
3527         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
3528                 return;
3529
3530         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
3531
3532         if (ethtool_validate_speed(speed))
3533                 vi->speed = speed;
3534
3535         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
3536
3537         if (ethtool_validate_duplex(duplex))
3538                 vi->duplex = duplex;
3539 }
3540
3541 static u32 virtnet_get_rxfh_key_size(struct net_device *dev)
3542 {
3543         return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size;
3544 }
3545
3546 static u32 virtnet_get_rxfh_indir_size(struct net_device *dev)
3547 {
3548         return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size;
3549 }
3550
3551 static int virtnet_get_rxfh(struct net_device *dev, u32 *indir, u8 *key, u8 *hfunc)
3552 {
3553         struct virtnet_info *vi = netdev_priv(dev);
3554         int i;
3555
3556         if (indir) {
3557                 for (i = 0; i < vi->rss_indir_table_size; ++i)
3558                         indir[i] = vi->ctrl->rss.indirection_table[i];
3559         }
3560
3561         if (key)
3562                 memcpy(key, vi->ctrl->rss.key, vi->rss_key_size);
3563
3564         if (hfunc)
3565                 *hfunc = ETH_RSS_HASH_TOP;
3566
3567         return 0;
3568 }
3569
3570 static int virtnet_set_rxfh(struct net_device *dev, const u32 *indir, const u8 *key, const u8 hfunc)
3571 {
3572         struct virtnet_info *vi = netdev_priv(dev);
3573         int i;
3574
3575         if (hfunc != ETH_RSS_HASH_NO_CHANGE && hfunc != ETH_RSS_HASH_TOP)
3576                 return -EOPNOTSUPP;
3577
3578         if (indir) {
3579                 for (i = 0; i < vi->rss_indir_table_size; ++i)
3580                         vi->ctrl->rss.indirection_table[i] = indir[i];
3581         }
3582         if (key)
3583                 memcpy(vi->ctrl->rss.key, key, vi->rss_key_size);
3584
3585         virtnet_commit_rss_command(vi);
3586
3587         return 0;
3588 }
3589
3590 static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs)
3591 {
3592         struct virtnet_info *vi = netdev_priv(dev);
3593         int rc = 0;
3594
3595         switch (info->cmd) {
3596         case ETHTOOL_GRXRINGS:
3597                 info->data = vi->curr_queue_pairs;
3598                 break;
3599         case ETHTOOL_GRXFH:
3600                 virtnet_get_hashflow(vi, info);
3601                 break;
3602         default:
3603                 rc = -EOPNOTSUPP;
3604         }
3605
3606         return rc;
3607 }
3608
3609 static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info)
3610 {
3611         struct virtnet_info *vi = netdev_priv(dev);
3612         int rc = 0;
3613
3614         switch (info->cmd) {
3615         case ETHTOOL_SRXFH:
3616                 if (!virtnet_set_hashflow(vi, info))
3617                         rc = -EINVAL;
3618
3619                 break;
3620         default:
3621                 rc = -EOPNOTSUPP;
3622         }
3623
3624         return rc;
3625 }
3626
3627 static const struct ethtool_ops virtnet_ethtool_ops = {
3628         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
3629                 ETHTOOL_COALESCE_USECS,
3630         .get_drvinfo = virtnet_get_drvinfo,
3631         .get_link = ethtool_op_get_link,
3632         .get_ringparam = virtnet_get_ringparam,
3633         .set_ringparam = virtnet_set_ringparam,
3634         .get_strings = virtnet_get_strings,
3635         .get_sset_count = virtnet_get_sset_count,
3636         .get_ethtool_stats = virtnet_get_ethtool_stats,
3637         .set_channels = virtnet_set_channels,
3638         .get_channels = virtnet_get_channels,
3639         .get_ts_info = ethtool_op_get_ts_info,
3640         .get_link_ksettings = virtnet_get_link_ksettings,
3641         .set_link_ksettings = virtnet_set_link_ksettings,
3642         .set_coalesce = virtnet_set_coalesce,
3643         .get_coalesce = virtnet_get_coalesce,
3644         .set_per_queue_coalesce = virtnet_set_per_queue_coalesce,
3645         .get_per_queue_coalesce = virtnet_get_per_queue_coalesce,
3646         .get_rxfh_key_size = virtnet_get_rxfh_key_size,
3647         .get_rxfh_indir_size = virtnet_get_rxfh_indir_size,
3648         .get_rxfh = virtnet_get_rxfh,
3649         .set_rxfh = virtnet_set_rxfh,
3650         .get_rxnfc = virtnet_get_rxnfc,
3651         .set_rxnfc = virtnet_set_rxnfc,
3652 };
3653
3654 static void virtnet_freeze_down(struct virtio_device *vdev)
3655 {
3656         struct virtnet_info *vi = vdev->priv;
3657
3658         /* Make sure no work handler is accessing the device */
3659         flush_work(&vi->config_work);
3660
3661         netif_tx_lock_bh(vi->dev);
3662         netif_device_detach(vi->dev);
3663         netif_tx_unlock_bh(vi->dev);
3664         if (netif_running(vi->dev))
3665                 virtnet_close(vi->dev);
3666 }
3667
3668 static int init_vqs(struct virtnet_info *vi);
3669
3670 static int virtnet_restore_up(struct virtio_device *vdev)
3671 {
3672         struct virtnet_info *vi = vdev->priv;
3673         int err;
3674
3675         err = init_vqs(vi);
3676         if (err)
3677                 return err;
3678
3679         virtio_device_ready(vdev);
3680
3681         enable_delayed_refill(vi);
3682
3683         if (netif_running(vi->dev)) {
3684                 err = virtnet_open(vi->dev);
3685                 if (err)
3686                         return err;
3687         }
3688
3689         netif_tx_lock_bh(vi->dev);
3690         netif_device_attach(vi->dev);
3691         netif_tx_unlock_bh(vi->dev);
3692         return err;
3693 }
3694
3695 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
3696 {
3697         struct scatterlist sg;
3698         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
3699
3700         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
3701
3702         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
3703                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
3704                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
3705                 return -EINVAL;
3706         }
3707
3708         return 0;
3709 }
3710
3711 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
3712 {
3713         u64 offloads = 0;
3714
3715         if (!vi->guest_offloads)
3716                 return 0;
3717
3718         return virtnet_set_guest_offloads(vi, offloads);
3719 }
3720
3721 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
3722 {
3723         u64 offloads = vi->guest_offloads;
3724
3725         if (!vi->guest_offloads)
3726                 return 0;
3727
3728         return virtnet_set_guest_offloads(vi, offloads);
3729 }
3730
3731 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
3732                            struct netlink_ext_ack *extack)
3733 {
3734         unsigned int room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
3735                                            sizeof(struct skb_shared_info));
3736         unsigned int max_sz = PAGE_SIZE - room - ETH_HLEN;
3737         struct virtnet_info *vi = netdev_priv(dev);
3738         struct bpf_prog *old_prog;
3739         u16 xdp_qp = 0, curr_qp;
3740         int i, err;
3741
3742         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
3743             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3744                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3745                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3746                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3747                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM) ||
3748                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) ||
3749                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6))) {
3750                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
3751                 return -EOPNOTSUPP;
3752         }
3753
3754         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
3755                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
3756                 return -EINVAL;
3757         }
3758
3759         if (prog && !prog->aux->xdp_has_frags && dev->mtu > max_sz) {
3760                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP without frags");
3761                 netdev_warn(dev, "single-buffer XDP requires MTU less than %u\n", max_sz);
3762                 return -EINVAL;
3763         }
3764
3765         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
3766         if (prog)
3767                 xdp_qp = nr_cpu_ids;
3768
3769         /* XDP requires extra queues for XDP_TX */
3770         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
3771                 netdev_warn_once(dev, "XDP request %i queues but max is %i. XDP_TX and XDP_REDIRECT will operate in a slower locked tx mode.\n",
3772                                  curr_qp + xdp_qp, vi->max_queue_pairs);
3773                 xdp_qp = 0;
3774         }
3775
3776         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
3777         if (!prog && !old_prog)
3778                 return 0;
3779
3780         if (prog)
3781                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
3782
3783         /* Make sure NAPI is not using any XDP TX queues for RX. */
3784         if (netif_running(dev)) {
3785                 for (i = 0; i < vi->max_queue_pairs; i++) {
3786                         napi_disable(&vi->rq[i].napi);
3787                         virtnet_napi_tx_disable(&vi->sq[i].napi);
3788                 }
3789         }
3790
3791         if (!prog) {
3792                 for (i = 0; i < vi->max_queue_pairs; i++) {
3793                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3794                         if (i == 0)
3795                                 virtnet_restore_guest_offloads(vi);
3796                 }
3797                 synchronize_net();
3798         }
3799
3800         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
3801         if (err)
3802                 goto err;
3803         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
3804         vi->xdp_queue_pairs = xdp_qp;
3805
3806         if (prog) {
3807                 vi->xdp_enabled = true;
3808                 for (i = 0; i < vi->max_queue_pairs; i++) {
3809                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3810                         if (i == 0 && !old_prog)
3811                                 virtnet_clear_guest_offloads(vi);
3812                 }
3813                 if (!old_prog)
3814                         xdp_features_set_redirect_target(dev, true);
3815         } else {
3816                 xdp_features_clear_redirect_target(dev);
3817                 vi->xdp_enabled = false;
3818         }
3819
3820         for (i = 0; i < vi->max_queue_pairs; i++) {
3821                 if (old_prog)
3822                         bpf_prog_put(old_prog);
3823                 if (netif_running(dev)) {
3824                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3825                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3826                                                &vi->sq[i].napi);
3827                 }
3828         }
3829
3830         return 0;
3831
3832 err:
3833         if (!prog) {
3834                 virtnet_clear_guest_offloads(vi);
3835                 for (i = 0; i < vi->max_queue_pairs; i++)
3836                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
3837         }
3838
3839         if (netif_running(dev)) {
3840                 for (i = 0; i < vi->max_queue_pairs; i++) {
3841                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3842                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3843                                                &vi->sq[i].napi);
3844                 }
3845         }
3846         if (prog)
3847                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
3848         return err;
3849 }
3850
3851 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
3852 {
3853         switch (xdp->command) {
3854         case XDP_SETUP_PROG:
3855                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
3856         default:
3857                 return -EINVAL;
3858         }
3859 }
3860
3861 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
3862                                       size_t len)
3863 {
3864         struct virtnet_info *vi = netdev_priv(dev);
3865         int ret;
3866
3867         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
3868                 return -EOPNOTSUPP;
3869
3870         ret = snprintf(buf, len, "sby");
3871         if (ret >= len)
3872                 return -EOPNOTSUPP;
3873
3874         return 0;
3875 }
3876
3877 static int virtnet_set_features(struct net_device *dev,
3878                                 netdev_features_t features)
3879 {
3880         struct virtnet_info *vi = netdev_priv(dev);
3881         u64 offloads;
3882         int err;
3883
3884         if ((dev->features ^ features) & NETIF_F_GRO_HW) {
3885                 if (vi->xdp_enabled)
3886                         return -EBUSY;
3887
3888                 if (features & NETIF_F_GRO_HW)
3889                         offloads = vi->guest_offloads_capable;
3890                 else
3891                         offloads = vi->guest_offloads_capable &
3892                                    ~GUEST_OFFLOAD_GRO_HW_MASK;
3893
3894                 err = virtnet_set_guest_offloads(vi, offloads);
3895                 if (err)
3896                         return err;
3897                 vi->guest_offloads = offloads;
3898         }
3899
3900         if ((dev->features ^ features) & NETIF_F_RXHASH) {
3901                 if (features & NETIF_F_RXHASH)
3902                         vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
3903                 else
3904                         vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE;
3905
3906                 if (!virtnet_commit_rss_command(vi))
3907                         return -EINVAL;
3908         }
3909
3910         return 0;
3911 }
3912
3913 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
3914 {
3915         struct virtnet_info *priv = netdev_priv(dev);
3916         struct send_queue *sq = &priv->sq[txqueue];
3917         struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
3918
3919         u64_stats_update_begin(&sq->stats.syncp);
3920         u64_stats_inc(&sq->stats.tx_timeouts);
3921         u64_stats_update_end(&sq->stats.syncp);
3922
3923         netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
3924                    txqueue, sq->name, sq->vq->index, sq->vq->name,
3925                    jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
3926 }
3927
3928 static const struct net_device_ops virtnet_netdev = {
3929         .ndo_open            = virtnet_open,
3930         .ndo_stop            = virtnet_close,
3931         .ndo_start_xmit      = start_xmit,
3932         .ndo_validate_addr   = eth_validate_addr,
3933         .ndo_set_mac_address = virtnet_set_mac_address,
3934         .ndo_set_rx_mode     = virtnet_set_rx_mode,
3935         .ndo_get_stats64     = virtnet_stats,
3936         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
3937         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
3938         .ndo_bpf                = virtnet_xdp,
3939         .ndo_xdp_xmit           = virtnet_xdp_xmit,
3940         .ndo_features_check     = passthru_features_check,
3941         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
3942         .ndo_set_features       = virtnet_set_features,
3943         .ndo_tx_timeout         = virtnet_tx_timeout,
3944 };
3945
3946 static void virtnet_config_changed_work(struct work_struct *work)
3947 {
3948         struct virtnet_info *vi =
3949                 container_of(work, struct virtnet_info, config_work);
3950         u16 v;
3951
3952         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
3953                                  struct virtio_net_config, status, &v) < 0)
3954                 return;
3955
3956         if (v & VIRTIO_NET_S_ANNOUNCE) {
3957                 netdev_notify_peers(vi->dev);
3958                 virtnet_ack_link_announce(vi);
3959         }
3960
3961         /* Ignore unknown (future) status bits */
3962         v &= VIRTIO_NET_S_LINK_UP;
3963
3964         if (vi->status == v)
3965                 return;
3966
3967         vi->status = v;
3968
3969         if (vi->status & VIRTIO_NET_S_LINK_UP) {
3970                 virtnet_update_settings(vi);
3971                 netif_carrier_on(vi->dev);
3972                 netif_tx_wake_all_queues(vi->dev);
3973         } else {
3974                 netif_carrier_off(vi->dev);
3975                 netif_tx_stop_all_queues(vi->dev);
3976         }
3977 }
3978
3979 static void virtnet_config_changed(struct virtio_device *vdev)
3980 {
3981         struct virtnet_info *vi = vdev->priv;
3982
3983         schedule_work(&vi->config_work);
3984 }
3985
3986 static void virtnet_free_queues(struct virtnet_info *vi)
3987 {
3988         int i;
3989
3990         for (i = 0; i < vi->max_queue_pairs; i++) {
3991                 __netif_napi_del(&vi->rq[i].napi);
3992                 __netif_napi_del(&vi->sq[i].napi);
3993         }
3994
3995         /* We called __netif_napi_del(),
3996          * we need to respect an RCU grace period before freeing vi->rq
3997          */
3998         synchronize_net();
3999
4000         kfree(vi->rq);
4001         kfree(vi->sq);
4002         kfree(vi->ctrl);
4003 }
4004
4005 static void _free_receive_bufs(struct virtnet_info *vi)
4006 {
4007         struct bpf_prog *old_prog;
4008         int i;
4009
4010         for (i = 0; i < vi->max_queue_pairs; i++) {
4011                 while (vi->rq[i].pages)
4012                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
4013
4014                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
4015                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
4016                 if (old_prog)
4017                         bpf_prog_put(old_prog);
4018         }
4019 }
4020
4021 static void free_receive_bufs(struct virtnet_info *vi)
4022 {
4023         rtnl_lock();
4024         _free_receive_bufs(vi);
4025         rtnl_unlock();
4026 }
4027
4028 static void free_receive_page_frags(struct virtnet_info *vi)
4029 {
4030         int i;
4031         for (i = 0; i < vi->max_queue_pairs; i++)
4032                 if (vi->rq[i].alloc_frag.page) {
4033                         if (vi->rq[i].do_dma && vi->rq[i].last_dma)
4034                                 virtnet_rq_unmap(&vi->rq[i], vi->rq[i].last_dma, 0);
4035                         put_page(vi->rq[i].alloc_frag.page);
4036                 }
4037 }
4038
4039 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
4040 {
4041         if (!is_xdp_frame(buf))
4042                 dev_kfree_skb(buf);
4043         else
4044                 xdp_return_frame(ptr_to_xdp(buf));
4045 }
4046
4047 static void free_unused_bufs(struct virtnet_info *vi)
4048 {
4049         void *buf;
4050         int i;
4051
4052         for (i = 0; i < vi->max_queue_pairs; i++) {
4053                 struct virtqueue *vq = vi->sq[i].vq;
4054                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
4055                         virtnet_sq_free_unused_buf(vq, buf);
4056                 cond_resched();
4057         }
4058
4059         for (i = 0; i < vi->max_queue_pairs; i++) {
4060                 struct virtqueue *vq = vi->rq[i].vq;
4061
4062                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
4063                         virtnet_rq_unmap_free_buf(vq, buf);
4064                 cond_resched();
4065         }
4066 }
4067
4068 static void virtnet_del_vqs(struct virtnet_info *vi)
4069 {
4070         struct virtio_device *vdev = vi->vdev;
4071
4072         virtnet_clean_affinity(vi);
4073
4074         vdev->config->del_vqs(vdev);
4075
4076         virtnet_free_queues(vi);
4077 }
4078
4079 /* How large should a single buffer be so a queue full of these can fit at
4080  * least one full packet?
4081  * Logic below assumes the mergeable buffer header is used.
4082  */
4083 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
4084 {
4085         const unsigned int hdr_len = vi->hdr_len;
4086         unsigned int rq_size = virtqueue_get_vring_size(vq);
4087         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
4088         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
4089         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
4090
4091         return max(max(min_buf_len, hdr_len) - hdr_len,
4092                    (unsigned int)GOOD_PACKET_LEN);
4093 }
4094
4095 static int virtnet_find_vqs(struct virtnet_info *vi)
4096 {
4097         vq_callback_t **callbacks;
4098         struct virtqueue **vqs;
4099         const char **names;
4100         int ret = -ENOMEM;
4101         int total_vqs;
4102         bool *ctx;
4103         u16 i;
4104
4105         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
4106          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
4107          * possible control vq.
4108          */
4109         total_vqs = vi->max_queue_pairs * 2 +
4110                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
4111
4112         /* Allocate space for find_vqs parameters */
4113         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
4114         if (!vqs)
4115                 goto err_vq;
4116         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
4117         if (!callbacks)
4118                 goto err_callback;
4119         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
4120         if (!names)
4121                 goto err_names;
4122         if (!vi->big_packets || vi->mergeable_rx_bufs) {
4123                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
4124                 if (!ctx)
4125                         goto err_ctx;
4126         } else {
4127                 ctx = NULL;
4128         }
4129
4130         /* Parameters for control virtqueue, if any */
4131         if (vi->has_cvq) {
4132                 callbacks[total_vqs - 1] = NULL;
4133                 names[total_vqs - 1] = "control";
4134         }
4135
4136         /* Allocate/initialize parameters for send/receive virtqueues */
4137         for (i = 0; i < vi->max_queue_pairs; i++) {
4138                 callbacks[rxq2vq(i)] = skb_recv_done;
4139                 callbacks[txq2vq(i)] = skb_xmit_done;
4140                 sprintf(vi->rq[i].name, "input.%u", i);
4141                 sprintf(vi->sq[i].name, "output.%u", i);
4142                 names[rxq2vq(i)] = vi->rq[i].name;
4143                 names[txq2vq(i)] = vi->sq[i].name;
4144                 if (ctx)
4145                         ctx[rxq2vq(i)] = true;
4146         }
4147
4148         ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
4149                                   names, ctx, NULL);
4150         if (ret)
4151                 goto err_find;
4152
4153         if (vi->has_cvq) {
4154                 vi->cvq = vqs[total_vqs - 1];
4155                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
4156                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
4157         }
4158
4159         for (i = 0; i < vi->max_queue_pairs; i++) {
4160                 vi->rq[i].vq = vqs[rxq2vq(i)];
4161                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
4162                 vi->sq[i].vq = vqs[txq2vq(i)];
4163         }
4164
4165         /* run here: ret == 0. */
4166
4167
4168 err_find:
4169         kfree(ctx);
4170 err_ctx:
4171         kfree(names);
4172 err_names:
4173         kfree(callbacks);
4174 err_callback:
4175         kfree(vqs);
4176 err_vq:
4177         return ret;
4178 }
4179
4180 static int virtnet_alloc_queues(struct virtnet_info *vi)
4181 {
4182         int i;
4183
4184         if (vi->has_cvq) {
4185                 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
4186                 if (!vi->ctrl)
4187                         goto err_ctrl;
4188         } else {
4189                 vi->ctrl = NULL;
4190         }
4191         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
4192         if (!vi->sq)
4193                 goto err_sq;
4194         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
4195         if (!vi->rq)
4196                 goto err_rq;
4197
4198         INIT_DELAYED_WORK(&vi->refill, refill_work);
4199         for (i = 0; i < vi->max_queue_pairs; i++) {
4200                 vi->rq[i].pages = NULL;
4201                 netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll,
4202                                       napi_weight);
4203                 netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi,
4204                                          virtnet_poll_tx,
4205                                          napi_tx ? napi_weight : 0);
4206
4207                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
4208                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
4209                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
4210
4211                 u64_stats_init(&vi->rq[i].stats.syncp);
4212                 u64_stats_init(&vi->sq[i].stats.syncp);
4213         }
4214
4215         return 0;
4216
4217 err_rq:
4218         kfree(vi->sq);
4219 err_sq:
4220         kfree(vi->ctrl);
4221 err_ctrl:
4222         return -ENOMEM;
4223 }
4224
4225 static int init_vqs(struct virtnet_info *vi)
4226 {
4227         int ret;
4228
4229         /* Allocate send & receive queues */
4230         ret = virtnet_alloc_queues(vi);
4231         if (ret)
4232                 goto err;
4233
4234         ret = virtnet_find_vqs(vi);
4235         if (ret)
4236                 goto err_free;
4237
4238         virtnet_rq_set_premapped(vi);
4239
4240         cpus_read_lock();
4241         virtnet_set_affinity(vi);
4242         cpus_read_unlock();
4243
4244         return 0;
4245
4246 err_free:
4247         virtnet_free_queues(vi);
4248 err:
4249         return ret;
4250 }
4251
4252 #ifdef CONFIG_SYSFS
4253 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
4254                 char *buf)
4255 {
4256         struct virtnet_info *vi = netdev_priv(queue->dev);
4257         unsigned int queue_index = get_netdev_rx_queue_index(queue);
4258         unsigned int headroom = virtnet_get_headroom(vi);
4259         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
4260         struct ewma_pkt_len *avg;
4261
4262         BUG_ON(queue_index >= vi->max_queue_pairs);
4263         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
4264         return sprintf(buf, "%u\n",
4265                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
4266                                        SKB_DATA_ALIGN(headroom + tailroom)));
4267 }
4268
4269 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
4270         __ATTR_RO(mergeable_rx_buffer_size);
4271
4272 static struct attribute *virtio_net_mrg_rx_attrs[] = {
4273         &mergeable_rx_buffer_size_attribute.attr,
4274         NULL
4275 };
4276
4277 static const struct attribute_group virtio_net_mrg_rx_group = {
4278         .name = "virtio_net",
4279         .attrs = virtio_net_mrg_rx_attrs
4280 };
4281 #endif
4282
4283 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
4284                                     unsigned int fbit,
4285                                     const char *fname, const char *dname)
4286 {
4287         if (!virtio_has_feature(vdev, fbit))
4288                 return false;
4289
4290         dev_err(&vdev->dev, "device advertises feature %s but not %s",
4291                 fname, dname);
4292
4293         return true;
4294 }
4295
4296 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
4297         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
4298
4299 static bool virtnet_validate_features(struct virtio_device *vdev)
4300 {
4301         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
4302             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
4303                              "VIRTIO_NET_F_CTRL_VQ") ||
4304              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
4305                              "VIRTIO_NET_F_CTRL_VQ") ||
4306              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
4307                              "VIRTIO_NET_F_CTRL_VQ") ||
4308              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
4309              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
4310                              "VIRTIO_NET_F_CTRL_VQ") ||
4311              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS,
4312                              "VIRTIO_NET_F_CTRL_VQ") ||
4313              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT,
4314                              "VIRTIO_NET_F_CTRL_VQ") ||
4315              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL,
4316                              "VIRTIO_NET_F_CTRL_VQ") ||
4317              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_VQ_NOTF_COAL,
4318                              "VIRTIO_NET_F_CTRL_VQ"))) {
4319                 return false;
4320         }
4321
4322         return true;
4323 }
4324
4325 #define MIN_MTU ETH_MIN_MTU
4326 #define MAX_MTU ETH_MAX_MTU
4327
4328 static int virtnet_validate(struct virtio_device *vdev)
4329 {
4330         if (!vdev->config->get) {
4331                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
4332                         __func__);
4333                 return -EINVAL;
4334         }
4335
4336         if (!virtnet_validate_features(vdev))
4337                 return -EINVAL;
4338
4339         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4340                 int mtu = virtio_cread16(vdev,
4341                                          offsetof(struct virtio_net_config,
4342                                                   mtu));
4343                 if (mtu < MIN_MTU)
4344                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
4345         }
4346
4347         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY) &&
4348             !virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
4349                 dev_warn(&vdev->dev, "device advertises feature VIRTIO_NET_F_STANDBY but not VIRTIO_NET_F_MAC, disabling standby");
4350                 __virtio_clear_bit(vdev, VIRTIO_NET_F_STANDBY);
4351         }
4352
4353         return 0;
4354 }
4355
4356 static bool virtnet_check_guest_gso(const struct virtnet_info *vi)
4357 {
4358         return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
4359                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
4360                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
4361                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
4362                 (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) &&
4363                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6));
4364 }
4365
4366 static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu)
4367 {
4368         bool guest_gso = virtnet_check_guest_gso(vi);
4369
4370         /* If device can receive ANY guest GSO packets, regardless of mtu,
4371          * allocate packets of maximum size, otherwise limit it to only
4372          * mtu size worth only.
4373          */
4374         if (mtu > ETH_DATA_LEN || guest_gso) {
4375                 vi->big_packets = true;
4376                 vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE);
4377         }
4378 }
4379
4380 static int virtnet_probe(struct virtio_device *vdev)
4381 {
4382         int i, err = -ENOMEM;
4383         struct net_device *dev;
4384         struct virtnet_info *vi;
4385         u16 max_queue_pairs;
4386         int mtu = 0;
4387
4388         /* Find if host supports multiqueue/rss virtio_net device */
4389         max_queue_pairs = 1;
4390         if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
4391                 max_queue_pairs =
4392                      virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs));
4393
4394         /* We need at least 2 queue's */
4395         if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
4396             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
4397             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4398                 max_queue_pairs = 1;
4399
4400         /* Allocate ourselves a network device with room for our info */
4401         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
4402         if (!dev)
4403                 return -ENOMEM;
4404
4405         /* Set up network device as normal. */
4406         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
4407                            IFF_TX_SKB_NO_LINEAR;
4408         dev->netdev_ops = &virtnet_netdev;
4409         dev->features = NETIF_F_HIGHDMA;
4410
4411         dev->ethtool_ops = &virtnet_ethtool_ops;
4412         SET_NETDEV_DEV(dev, &vdev->dev);
4413
4414         /* Do we support "hardware" checksums? */
4415         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
4416                 /* This opens up the world of extra features. */
4417                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
4418                 if (csum)
4419                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
4420
4421                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
4422                         dev->hw_features |= NETIF_F_TSO
4423                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
4424                 }
4425                 /* Individual feature bits: what can host handle? */
4426                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
4427                         dev->hw_features |= NETIF_F_TSO;
4428                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
4429                         dev->hw_features |= NETIF_F_TSO6;
4430                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
4431                         dev->hw_features |= NETIF_F_TSO_ECN;
4432                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_USO))
4433                         dev->hw_features |= NETIF_F_GSO_UDP_L4;
4434
4435                 dev->features |= NETIF_F_GSO_ROBUST;
4436
4437                 if (gso)
4438                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
4439                 /* (!csum && gso) case will be fixed by register_netdev() */
4440         }
4441         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
4442                 dev->features |= NETIF_F_RXCSUM;
4443         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
4444             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
4445                 dev->features |= NETIF_F_GRO_HW;
4446         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
4447                 dev->hw_features |= NETIF_F_GRO_HW;
4448
4449         dev->vlan_features = dev->features;
4450         dev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT;
4451
4452         /* MTU range: 68 - 65535 */
4453         dev->min_mtu = MIN_MTU;
4454         dev->max_mtu = MAX_MTU;
4455
4456         /* Configuration may specify what MAC to use.  Otherwise random. */
4457         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
4458                 u8 addr[ETH_ALEN];
4459
4460                 virtio_cread_bytes(vdev,
4461                                    offsetof(struct virtio_net_config, mac),
4462                                    addr, ETH_ALEN);
4463                 eth_hw_addr_set(dev, addr);
4464         } else {
4465                 eth_hw_addr_random(dev);
4466                 dev_info(&vdev->dev, "Assigned random MAC address %pM\n",
4467                          dev->dev_addr);
4468         }
4469
4470         /* Set up our device-specific information */
4471         vi = netdev_priv(dev);
4472         vi->dev = dev;
4473         vi->vdev = vdev;
4474         vdev->priv = vi;
4475
4476         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
4477         spin_lock_init(&vi->refill_lock);
4478
4479         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) {
4480                 vi->mergeable_rx_bufs = true;
4481                 dev->xdp_features |= NETDEV_XDP_ACT_RX_SG;
4482         }
4483
4484         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
4485                 vi->intr_coal_rx.max_usecs = 0;
4486                 vi->intr_coal_tx.max_usecs = 0;
4487                 vi->intr_coal_tx.max_packets = 0;
4488                 vi->intr_coal_rx.max_packets = 0;
4489         }
4490
4491         if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT))
4492                 vi->has_rss_hash_report = true;
4493
4494         if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
4495                 vi->has_rss = true;
4496
4497         if (vi->has_rss || vi->has_rss_hash_report) {
4498                 vi->rss_indir_table_size =
4499                         virtio_cread16(vdev, offsetof(struct virtio_net_config,
4500                                 rss_max_indirection_table_length));
4501                 vi->rss_key_size =
4502                         virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size));
4503
4504                 vi->rss_hash_types_supported =
4505                     virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types));
4506                 vi->rss_hash_types_supported &=
4507                                 ~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX |
4508                                   VIRTIO_NET_RSS_HASH_TYPE_TCP_EX |
4509                                   VIRTIO_NET_RSS_HASH_TYPE_UDP_EX);
4510
4511                 dev->hw_features |= NETIF_F_RXHASH;
4512         }
4513
4514         if (vi->has_rss_hash_report)
4515                 vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash);
4516         else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
4517                  virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4518                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
4519         else
4520                 vi->hdr_len = sizeof(struct virtio_net_hdr);
4521
4522         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
4523             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4524                 vi->any_header_sg = true;
4525
4526         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4527                 vi->has_cvq = true;
4528
4529         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4530                 mtu = virtio_cread16(vdev,
4531                                      offsetof(struct virtio_net_config,
4532                                               mtu));
4533                 if (mtu < dev->min_mtu) {
4534                         /* Should never trigger: MTU was previously validated
4535                          * in virtnet_validate.
4536                          */
4537                         dev_err(&vdev->dev,
4538                                 "device MTU appears to have changed it is now %d < %d",
4539                                 mtu, dev->min_mtu);
4540                         err = -EINVAL;
4541                         goto free;
4542                 }
4543
4544                 dev->mtu = mtu;
4545                 dev->max_mtu = mtu;
4546         }
4547
4548         virtnet_set_big_packets(vi, mtu);
4549
4550         if (vi->any_header_sg)
4551                 dev->needed_headroom = vi->hdr_len;
4552
4553         /* Enable multiqueue by default */
4554         if (num_online_cpus() >= max_queue_pairs)
4555                 vi->curr_queue_pairs = max_queue_pairs;
4556         else
4557                 vi->curr_queue_pairs = num_online_cpus();
4558         vi->max_queue_pairs = max_queue_pairs;
4559
4560         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
4561         err = init_vqs(vi);
4562         if (err)
4563                 goto free;
4564
4565 #ifdef CONFIG_SYSFS
4566         if (vi->mergeable_rx_bufs)
4567                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
4568 #endif
4569         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
4570         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
4571
4572         virtnet_init_settings(dev);
4573
4574         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
4575                 vi->failover = net_failover_create(vi->dev);
4576                 if (IS_ERR(vi->failover)) {
4577                         err = PTR_ERR(vi->failover);
4578                         goto free_vqs;
4579                 }
4580         }
4581
4582         if (vi->has_rss || vi->has_rss_hash_report)
4583                 virtnet_init_default_rss(vi);
4584
4585         /* serialize netdev register + virtio_device_ready() with ndo_open() */
4586         rtnl_lock();
4587
4588         err = register_netdevice(dev);
4589         if (err) {
4590                 pr_debug("virtio_net: registering device failed\n");
4591                 rtnl_unlock();
4592                 goto free_failover;
4593         }
4594
4595         virtio_device_ready(vdev);
4596
4597         _virtnet_set_queues(vi, vi->curr_queue_pairs);
4598
4599         /* a random MAC address has been assigned, notify the device.
4600          * We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not there
4601          * because many devices work fine without getting MAC explicitly
4602          */
4603         if (!virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
4604             virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
4605                 struct scatterlist sg;
4606
4607                 sg_init_one(&sg, dev->dev_addr, dev->addr_len);
4608                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
4609                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
4610                         pr_debug("virtio_net: setting MAC address failed\n");
4611                         rtnl_unlock();
4612                         err = -EINVAL;
4613                         goto free_unregister_netdev;
4614                 }
4615         }
4616
4617         rtnl_unlock();
4618
4619         err = virtnet_cpu_notif_add(vi);
4620         if (err) {
4621                 pr_debug("virtio_net: registering cpu notifier failed\n");
4622                 goto free_unregister_netdev;
4623         }
4624
4625         /* Assume link up if device can't report link status,
4626            otherwise get link status from config. */
4627         netif_carrier_off(dev);
4628         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
4629                 schedule_work(&vi->config_work);
4630         } else {
4631                 vi->status = VIRTIO_NET_S_LINK_UP;
4632                 virtnet_update_settings(vi);
4633                 netif_carrier_on(dev);
4634         }
4635
4636         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
4637                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
4638                         set_bit(guest_offloads[i], &vi->guest_offloads);
4639         vi->guest_offloads_capable = vi->guest_offloads;
4640
4641         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
4642                  dev->name, max_queue_pairs);
4643
4644         return 0;
4645
4646 free_unregister_netdev:
4647         unregister_netdev(dev);
4648 free_failover:
4649         net_failover_destroy(vi->failover);
4650 free_vqs:
4651         virtio_reset_device(vdev);
4652         cancel_delayed_work_sync(&vi->refill);
4653         free_receive_page_frags(vi);
4654         virtnet_del_vqs(vi);
4655 free:
4656         free_netdev(dev);
4657         return err;
4658 }
4659
4660 static void remove_vq_common(struct virtnet_info *vi)
4661 {
4662         virtio_reset_device(vi->vdev);
4663
4664         /* Free unused buffers in both send and recv, if any. */
4665         free_unused_bufs(vi);
4666
4667         free_receive_bufs(vi);
4668
4669         free_receive_page_frags(vi);
4670
4671         virtnet_del_vqs(vi);
4672 }
4673
4674 static void virtnet_remove(struct virtio_device *vdev)
4675 {
4676         struct virtnet_info *vi = vdev->priv;
4677
4678         virtnet_cpu_notif_remove(vi);
4679
4680         /* Make sure no work handler is accessing the device. */
4681         flush_work(&vi->config_work);
4682
4683         unregister_netdev(vi->dev);
4684
4685         net_failover_destroy(vi->failover);
4686
4687         remove_vq_common(vi);
4688
4689         free_netdev(vi->dev);
4690 }
4691
4692 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
4693 {
4694         struct virtnet_info *vi = vdev->priv;
4695
4696         virtnet_cpu_notif_remove(vi);
4697         virtnet_freeze_down(vdev);
4698         remove_vq_common(vi);
4699
4700         return 0;
4701 }
4702
4703 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
4704 {
4705         struct virtnet_info *vi = vdev->priv;
4706         int err;
4707
4708         err = virtnet_restore_up(vdev);
4709         if (err)
4710                 return err;
4711         virtnet_set_queues(vi, vi->curr_queue_pairs);
4712
4713         err = virtnet_cpu_notif_add(vi);
4714         if (err) {
4715                 virtnet_freeze_down(vdev);
4716                 remove_vq_common(vi);
4717                 return err;
4718         }
4719
4720         return 0;
4721 }
4722
4723 static struct virtio_device_id id_table[] = {
4724         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
4725         { 0 },
4726 };
4727
4728 #define VIRTNET_FEATURES \
4729         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
4730         VIRTIO_NET_F_MAC, \
4731         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
4732         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
4733         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
4734         VIRTIO_NET_F_HOST_USO, VIRTIO_NET_F_GUEST_USO4, VIRTIO_NET_F_GUEST_USO6, \
4735         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
4736         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
4737         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
4738         VIRTIO_NET_F_CTRL_MAC_ADDR, \
4739         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
4740         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \
4741         VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL, \
4742         VIRTIO_NET_F_VQ_NOTF_COAL, \
4743         VIRTIO_NET_F_GUEST_HDRLEN
4744
4745 static unsigned int features[] = {
4746         VIRTNET_FEATURES,
4747 };
4748
4749 static unsigned int features_legacy[] = {
4750         VIRTNET_FEATURES,
4751         VIRTIO_NET_F_GSO,
4752         VIRTIO_F_ANY_LAYOUT,
4753 };
4754
4755 static struct virtio_driver virtio_net_driver = {
4756         .feature_table = features,
4757         .feature_table_size = ARRAY_SIZE(features),
4758         .feature_table_legacy = features_legacy,
4759         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
4760         .driver.name =  KBUILD_MODNAME,
4761         .driver.owner = THIS_MODULE,
4762         .id_table =     id_table,
4763         .validate =     virtnet_validate,
4764         .probe =        virtnet_probe,
4765         .remove =       virtnet_remove,
4766         .config_changed = virtnet_config_changed,
4767 #ifdef CONFIG_PM_SLEEP
4768         .freeze =       virtnet_freeze,
4769         .restore =      virtnet_restore,
4770 #endif
4771 };
4772
4773 static __init int virtio_net_driver_init(void)
4774 {
4775         int ret;
4776
4777         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
4778                                       virtnet_cpu_online,
4779                                       virtnet_cpu_down_prep);
4780         if (ret < 0)
4781                 goto out;
4782         virtionet_online = ret;
4783         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
4784                                       NULL, virtnet_cpu_dead);
4785         if (ret)
4786                 goto err_dead;
4787         ret = register_virtio_driver(&virtio_net_driver);
4788         if (ret)
4789                 goto err_virtio;
4790         return 0;
4791 err_virtio:
4792         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4793 err_dead:
4794         cpuhp_remove_multi_state(virtionet_online);
4795 out:
4796         return ret;
4797 }
4798 module_init(virtio_net_driver_init);
4799
4800 static __exit void virtio_net_driver_exit(void)
4801 {
4802         unregister_virtio_driver(&virtio_net_driver);
4803         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4804         cpuhp_remove_multi_state(virtionet_online);
4805 }
4806 module_exit(virtio_net_driver_exit);
4807
4808 MODULE_DEVICE_TABLE(virtio, id_table);
4809 MODULE_DESCRIPTION("Virtio network driver");
4810 MODULE_LICENSE("GPL");