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
  26 static int napi_weight = NAPI_POLL_WEIGHT;
  27 module_param(napi_weight, int, 0444);
  28
  29 static bool csum = true, gso = true, napi_tx = true;
  30 module_param(csum, bool, 0444);
  31 module_param(gso, bool, 0444);
  32 module_param(napi_tx, bool, 0644);
  33
  34 /* FIXME: MTU in config. */
  35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
  36 #define GOOD_COPY_LEN   128
  37
  38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
  39
  40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
  41 #define VIRTIO_XDP_HEADROOM 256
  42
  43 /* Separating two types of XDP xmit */
  44 #define VIRTIO_XDP_TX           BIT(0)
  45 #define VIRTIO_XDP_REDIR        BIT(1)
  46
  47 #define VIRTIO_XDP_FLAG BIT(0)
  48
  49 /* RX packet size EWMA. The average packet size is used to determine the packet
  50  * buffer size when refilling RX rings. As the entire RX ring may be refilled
  51  * at once, the weight is chosen so that the EWMA will be insensitive to short-
  52  * term, transient changes in packet size.
  53  */
  54 DECLARE_EWMA(pkt_len, 0, 64)
  55
  56 #define VIRTNET_DRIVER_VERSION "1.0.0"
  57
  58 static const unsigned long guest_offloads[] = {
  59         VIRTIO_NET_F_GUEST_TSO4,
  60         VIRTIO_NET_F_GUEST_TSO6,
  61         VIRTIO_NET_F_GUEST_ECN,
  62         VIRTIO_NET_F_GUEST_UFO,
  63         VIRTIO_NET_F_GUEST_CSUM
  64 };
  65
  66 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
  67                                 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
  68                                 (1ULL << VIRTIO_NET_F_GUEST_ECN)  | \
  69                                 (1ULL << VIRTIO_NET_F_GUEST_UFO))
  70
  71 struct virtnet_stat_desc {
  72         char desc[ETH_GSTRING_LEN];
  73         size_t offset;
  74 };
  75
  76 struct virtnet_sq_stats {
  77         struct u64_stats_sync syncp;
  78         u64 packets;
  79         u64 bytes;
  80         u64 xdp_tx;
  81         u64 xdp_tx_drops;
  82         u64 kicks;
  83         u64 tx_timeouts;
  84 };
  85
  86 struct virtnet_rq_stats {
  87         struct u64_stats_sync syncp;
  88         u64 packets;
  89         u64 bytes;
  90         u64 drops;
  91         u64 xdp_packets;
  92         u64 xdp_tx;
  93         u64 xdp_redirects;
  94         u64 xdp_drops;
  95         u64 kicks;
  96 };
  97
  98 #define VIRTNET_SQ_STAT(m)      offsetof(struct virtnet_sq_stats, m)
  99 #define VIRTNET_RQ_STAT(m)      offsetof(struct virtnet_rq_stats, m)
 100
 101 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
 102         { "packets",            VIRTNET_SQ_STAT(packets) },
 103         { "bytes",              VIRTNET_SQ_STAT(bytes) },
 104         { "xdp_tx",             VIRTNET_SQ_STAT(xdp_tx) },
 105         { "xdp_tx_drops",       VIRTNET_SQ_STAT(xdp_tx_drops) },
 106         { "kicks",              VIRTNET_SQ_STAT(kicks) },
 107         { "tx_timeouts",        VIRTNET_SQ_STAT(tx_timeouts) },
 108 };
 109
 110 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
 111         { "packets",            VIRTNET_RQ_STAT(packets) },
 112         { "bytes",              VIRTNET_RQ_STAT(bytes) },
 113         { "drops",              VIRTNET_RQ_STAT(drops) },
 114         { "xdp_packets",        VIRTNET_RQ_STAT(xdp_packets) },
 115         { "xdp_tx",             VIRTNET_RQ_STAT(xdp_tx) },
 116         { "xdp_redirects",      VIRTNET_RQ_STAT(xdp_redirects) },
 117         { "xdp_drops",          VIRTNET_RQ_STAT(xdp_drops) },
 118         { "kicks",              VIRTNET_RQ_STAT(kicks) },
 119 };
 120
 121 #define VIRTNET_SQ_STATS_LEN    ARRAY_SIZE(virtnet_sq_stats_desc)
 122 #define VIRTNET_RQ_STATS_LEN    ARRAY_SIZE(virtnet_rq_stats_desc)
 123
 124 /* Internal representation of a send virtqueue */
 125 struct send_queue {
 126         /* Virtqueue associated with this send _queue */
 127         struct virtqueue *vq;
 128
 129         /* TX: fragments + linear part + virtio header */
 130         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 131
 132         /* Name of the send queue: output.$index */
 133         char name[40];
 134
 135         struct virtnet_sq_stats stats;
 136
 137         struct napi_struct napi;
 138
 139         /* Record whether sq is in reset state. */
 140         bool reset;
 141 };
 142
 143 /* Internal representation of a receive virtqueue */
 144 struct receive_queue {
 145         /* Virtqueue associated with this receive_queue */
 146         struct virtqueue *vq;
 147
 148         struct napi_struct napi;
 149
 150         struct bpf_prog __rcu *xdp_prog;
 151
 152         struct virtnet_rq_stats stats;
 153
 154         /* Chain pages by the private ptr. */
 155         struct page *pages;
 156
 157         /* Average packet length for mergeable receive buffers. */
 158         struct ewma_pkt_len mrg_avg_pkt_len;
 159
 160         /* Page frag for packet buffer allocation. */
 161         struct page_frag alloc_frag;
 162
 163         /* RX: fragments + linear part + virtio header */
 164         struct scatterlist sg[MAX_SKB_FRAGS + 2];
 165
 166         /* Min single buffer size for mergeable buffers case. */
 167         unsigned int min_buf_len;
 168
 169         /* Name of this receive queue: input.$index */
 170         char name[40];
 171
 172         struct xdp_rxq_info xdp_rxq;
 173 };
 174
 175 /* This structure can contain rss message with maximum settings for indirection table and keysize
 176  * Note, that default structure that describes RSS configuration virtio_net_rss_config
 177  * contains same info but can't handle table values.
 178  * In any case, structure would be passed to virtio hw through sg_buf split by parts
 179  * because table sizes may be differ according to the device configuration.
 180  */
 181 #define VIRTIO_NET_RSS_MAX_KEY_SIZE     40
 182 #define VIRTIO_NET_RSS_MAX_TABLE_LEN    128
 183 struct virtio_net_ctrl_rss {
 184         u32 hash_types;
 185         u16 indirection_table_mask;
 186         u16 unclassified_queue;
 187         u16 indirection_table[VIRTIO_NET_RSS_MAX_TABLE_LEN];
 188         u16 max_tx_vq;
 189         u8 hash_key_length;
 190         u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE];
 191 };
 192
 193 /* Control VQ buffers: protected by the rtnl lock */
 194 struct control_buf {
 195         struct virtio_net_ctrl_hdr hdr;
 196         virtio_net_ctrl_ack status;
 197         struct virtio_net_ctrl_mq mq;
 198         u8 promisc;
 199         u8 allmulti;
 200         __virtio16 vid;
 201         __virtio64 offloads;
 202         struct virtio_net_ctrl_rss rss;
 203 };
 204
 205 struct virtnet_info {
 206         struct virtio_device *vdev;
 207         struct virtqueue *cvq;
 208         struct net_device *dev;
 209         struct send_queue *sq;
 210         struct receive_queue *rq;
 211         unsigned int status;
 212
 213         /* Max # of queue pairs supported by the device */
 214         u16 max_queue_pairs;
 215
 216         /* # of queue pairs currently used by the driver */
 217         u16 curr_queue_pairs;
 218
 219         /* # of XDP queue pairs currently used by the driver */
 220         u16 xdp_queue_pairs;
 221
 222         /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
 223         bool xdp_enabled;
 224
 225         /* I like... big packets and I cannot lie! */
 226         bool big_packets;
 227
 228         /* number of sg entries allocated for big packets */
 229         unsigned int big_packets_num_skbfrags;
 230
 231         /* Host will merge rx buffers for big packets (shake it! shake it!) */
 232         bool mergeable_rx_bufs;
 233
 234         /* Host supports rss and/or hash report */
 235         bool has_rss;
 236         bool has_rss_hash_report;
 237         u8 rss_key_size;
 238         u16 rss_indir_table_size;
 239         u32 rss_hash_types_supported;
 240         u32 rss_hash_types_saved;
 241
 242         /* Has control virtqueue */
 243         bool has_cvq;
 244
 245         /* Host can handle any s/g split between our header and packet data */
 246         bool any_header_sg;
 247
 248         /* Packet virtio header size */
 249         u8 hdr_len;
 250
 251         /* Work struct for delayed refilling if we run low on memory. */
 252         struct delayed_work refill;
 253
 254         /* Is delayed refill enabled? */
 255         bool refill_enabled;
 256
 257         /* The lock to synchronize the access to refill_enabled */
 258         spinlock_t refill_lock;
 259
 260         /* Work struct for config space updates */
 261         struct work_struct config_work;
 262
 263         /* Does the affinity hint is set for virtqueues? */
 264         bool affinity_hint_set;
 265
 266         /* CPU hotplug instances for online & dead */
 267         struct hlist_node node;
 268         struct hlist_node node_dead;
 269
 270         struct control_buf *ctrl;
 271
 272         /* Ethtool settings */
 273         u8 duplex;
 274         u32 speed;
 275
 276         /* Interrupt coalescing settings */
 277         u32 tx_usecs;
 278         u32 rx_usecs;
 279         u32 tx_max_packets;
 280         u32 rx_max_packets;
 281
 282         unsigned long guest_offloads;
 283         unsigned long guest_offloads_capable;
 284
 285         /* failover when STANDBY feature enabled */
 286         struct failover *failover;
 287 };
 288
 289 struct padded_vnet_hdr {
 290         struct virtio_net_hdr_v1_hash hdr;
 291         /*
 292          * hdr is in a separate sg buffer, and data sg buffer shares same page
 293          * with this header sg. This padding makes next sg 16 byte aligned
 294          * after the header.
 295          */
 296         char padding[12];
 297 };
 298
 299 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf);
 300 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf);
 301
 302 static bool is_xdp_frame(void *ptr)
 303 {
 304         return (unsigned long)ptr & VIRTIO_XDP_FLAG;
 305 }
 306
 307 static void *xdp_to_ptr(struct xdp_frame *ptr)
 308 {
 309         return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
 310 }
 311
 312 static struct xdp_frame *ptr_to_xdp(void *ptr)
 313 {
 314         return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 315 }
 316
 317 /* Converting between virtqueue no. and kernel tx/rx queue no.
 318  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
 319  */
 320 static int vq2txq(struct virtqueue *vq)
 321 {
 322         return (vq->index - 1) / 2;
 323 }
 324
 325 static int txq2vq(int txq)
 326 {
 327         return txq * 2 + 1;
 328 }
 329
 330 static int vq2rxq(struct virtqueue *vq)
 331 {
 332         return vq->index / 2;
 333 }
 334
 335 static int rxq2vq(int rxq)
 336 {
 337         return rxq * 2;
 338 }
 339
 340 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
 341 {
 342         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
 343 }
 344
 345 /*
 346  * private is used to chain pages for big packets, put the whole
 347  * most recent used list in the beginning for reuse
 348  */
 349 static void give_pages(struct receive_queue *rq, struct page *page)
 350 {
 351         struct page *end;
 352
 353         /* Find end of list, sew whole thing into vi->rq.pages. */
 354         for (end = page; end->private; end = (struct page *)end->private);
 355         end->private = (unsigned long)rq->pages;
 356         rq->pages = page;
 357 }
 358
 359 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
 360 {
 361         struct page *p = rq->pages;
 362
 363         if (p) {
 364                 rq->pages = (struct page *)p->private;
 365                 /* clear private here, it is used to chain pages */
 366                 p->private = 0;
 367         } else
 368                 p = alloc_page(gfp_mask);
 369         return p;
 370 }
 371
 372 static void enable_delayed_refill(struct virtnet_info *vi)
 373 {
 374         spin_lock_bh(&vi->refill_lock);
 375         vi->refill_enabled = true;
 376         spin_unlock_bh(&vi->refill_lock);
 377 }
 378
 379 static void disable_delayed_refill(struct virtnet_info *vi)
 380 {
 381         spin_lock_bh(&vi->refill_lock);
 382         vi->refill_enabled = false;
 383         spin_unlock_bh(&vi->refill_lock);
 384 }
 385
 386 static void virtqueue_napi_schedule(struct napi_struct *napi,
 387                                     struct virtqueue *vq)
 388 {
 389         if (napi_schedule_prep(napi)) {
 390                 virtqueue_disable_cb(vq);
 391                 __napi_schedule(napi);
 392         }
 393 }
 394
 395 static void virtqueue_napi_complete(struct napi_struct *napi,
 396                                     struct virtqueue *vq, int processed)
 397 {
 398         int opaque;
 399
 400         opaque = virtqueue_enable_cb_prepare(vq);
 401         if (napi_complete_done(napi, processed)) {
 402                 if (unlikely(virtqueue_poll(vq, opaque)))
 403                         virtqueue_napi_schedule(napi, vq);
 404         } else {
 405                 virtqueue_disable_cb(vq);
 406         }
 407 }
 408
 409 static void skb_xmit_done(struct virtqueue *vq)
 410 {
 411         struct virtnet_info *vi = vq->vdev->priv;
 412         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
 413
 414         /* Suppress further interrupts. */
 415         virtqueue_disable_cb(vq);
 416
 417         if (napi->weight)
 418                 virtqueue_napi_schedule(napi, vq);
 419         else
 420                 /* We were probably waiting for more output buffers. */
 421                 netif_wake_subqueue(vi->dev, vq2txq(vq));
 422 }
 423
 424 #define MRG_CTX_HEADER_SHIFT 22
 425 static void *mergeable_len_to_ctx(unsigned int truesize,
 426                                   unsigned int headroom)
 427 {
 428         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
 429 }
 430
 431 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
 432 {
 433         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
 434 }
 435
 436 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
 437 {
 438         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
 439 }
 440
 441 /* Called from bottom half context */
 442 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 443                                    struct receive_queue *rq,
 444                                    struct page *page, unsigned int offset,
 445                                    unsigned int len, unsigned int truesize,
 446                                    bool hdr_valid, unsigned int metasize,
 447                                    unsigned int headroom)
 448 {
 449         struct sk_buff *skb;
 450         struct virtio_net_hdr_mrg_rxbuf *hdr;
 451         unsigned int copy, hdr_len, hdr_padded_len;
 452         struct page *page_to_free = NULL;
 453         int tailroom, shinfo_size;
 454         char *p, *hdr_p, *buf;
 455
 456         p = page_address(page) + offset;
 457         hdr_p = p;
 458
 459         hdr_len = vi->hdr_len;
 460         if (vi->mergeable_rx_bufs)
 461                 hdr_padded_len = hdr_len;
 462         else
 463                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
 464
 465         /* If headroom is not 0, there is an offset between the beginning of the
 466          * data and the allocated space, otherwise the data and the allocated
 467          * space are aligned.
 468          *
 469          * Buffers with headroom use PAGE_SIZE as alloc size, see
 470          * add_recvbuf_mergeable() + get_mergeable_buf_len()
 471          */
 472         truesize = headroom ? PAGE_SIZE : truesize;
 473         tailroom = truesize - headroom;
 474         buf = p - headroom;
 475
 476         len -= hdr_len;
 477         offset += hdr_padded_len;
 478         p += hdr_padded_len;
 479         tailroom -= hdr_padded_len + len;
 480
 481         shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 482
 483         /* copy small packet so we can reuse these pages */
 484         if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
 485                 skb = build_skb(buf, truesize);
 486                 if (unlikely(!skb))
 487                         return NULL;
 488
 489                 skb_reserve(skb, p - buf);
 490                 skb_put(skb, len);
 491
 492                 page = (struct page *)page->private;
 493                 if (page)
 494                         give_pages(rq, page);
 495                 goto ok;
 496         }
 497
 498         /* copy small packet so we can reuse these pages for small data */
 499         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
 500         if (unlikely(!skb))
 501                 return NULL;
 502
 503         /* Copy all frame if it fits skb->head, otherwise
 504          * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
 505          */
 506         if (len <= skb_tailroom(skb))
 507                 copy = len;
 508         else
 509                 copy = ETH_HLEN + metasize;
 510         skb_put_data(skb, p, copy);
 511
 512         len -= copy;
 513         offset += copy;
 514
 515         if (vi->mergeable_rx_bufs) {
 516                 if (len)
 517                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
 518                 else
 519                         page_to_free = page;
 520                 goto ok;
 521         }
 522
 523         /*
 524          * Verify that we can indeed put this data into a skb.
 525          * This is here to handle cases when the device erroneously
 526          * tries to receive more than is possible. This is usually
 527          * the case of a broken device.
 528          */
 529         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
 530                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
 531                 dev_kfree_skb(skb);
 532                 return NULL;
 533         }
 534         BUG_ON(offset >= PAGE_SIZE);
 535         while (len) {
 536                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
 537                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
 538                                 frag_size, truesize);
 539                 len -= frag_size;
 540                 page = (struct page *)page->private;
 541                 offset = 0;
 542         }
 543
 544         if (page)
 545                 give_pages(rq, page);
 546
 547 ok:
 548         /* hdr_valid means no XDP, so we can copy the vnet header */
 549         if (hdr_valid) {
 550                 hdr = skb_vnet_hdr(skb);
 551                 memcpy(hdr, hdr_p, hdr_len);
 552         }
 553         if (page_to_free)
 554                 put_page(page_to_free);
 555
 556         if (metasize) {
 557                 __skb_pull(skb, metasize);
 558                 skb_metadata_set(skb, metasize);
 559         }
 560
 561         return skb;
 562 }
 563
 564 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 565                                    struct send_queue *sq,
 566                                    struct xdp_frame *xdpf)
 567 {
 568         struct virtio_net_hdr_mrg_rxbuf *hdr;
 569         int err;
 570
 571         if (unlikely(xdpf->headroom < vi->hdr_len))
 572                 return -EOVERFLOW;
 573
 574         /* Make room for virtqueue hdr (also change xdpf->headroom?) */
 575         xdpf->data -= vi->hdr_len;
 576         /* Zero header and leave csum up to XDP layers */
 577         hdr = xdpf->data;
 578         memset(hdr, 0, vi->hdr_len);
 579         xdpf->len   += vi->hdr_len;
 580
 581         sg_init_one(sq->sg, xdpf->data, xdpf->len);
 582
 583         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
 584                                    GFP_ATOMIC);
 585         if (unlikely(err))
 586                 return -ENOSPC; /* Caller handle free/refcnt */
 587
 588         return 0;
 589 }
 590
 591 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
 592  * the current cpu, so it does not need to be locked.
 593  *
 594  * Here we use marco instead of inline functions because we have to deal with
 595  * three issues at the same time: 1. the choice of sq. 2. judge and execute the
 596  * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
 597  * functions to perfectly solve these three problems at the same time.
 598  */
 599 #define virtnet_xdp_get_sq(vi) ({                                       \
 600         int cpu = smp_processor_id();                                   \
 601         struct netdev_queue *txq;                                       \
 602         typeof(vi) v = (vi);                                            \
 603         unsigned int qp;                                                \
 604                                                                         \
 605         if (v->curr_queue_pairs > nr_cpu_ids) {                         \
 606                 qp = v->curr_queue_pairs - v->xdp_queue_pairs;          \
 607                 qp += cpu;                                              \
 608                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 609                 __netif_tx_acquire(txq);                                \
 610         } else {                                                        \
 611                 qp = cpu % v->curr_queue_pairs;                         \
 612                 txq = netdev_get_tx_queue(v->dev, qp);                  \
 613                 __netif_tx_lock(txq, cpu);                              \
 614         }                                                               \
 615         v->sq + qp;                                                     \
 616 })
 617
 618 #define virtnet_xdp_put_sq(vi, q) {                                     \
 619         struct netdev_queue *txq;                                       \
 620         typeof(vi) v = (vi);                                            \
 621                                                                         \
 622         txq = netdev_get_tx_queue(v->dev, (q) - v->sq);                 \
 623         if (v->curr_queue_pairs > nr_cpu_ids)                           \
 624                 __netif_tx_release(txq);                                \
 625         else                                                            \
 626                 __netif_tx_unlock(txq);                                 \
 627 }
 628
 629 static int virtnet_xdp_xmit(struct net_device *dev,
 630                             int n, struct xdp_frame **frames, u32 flags)
 631 {
 632         struct virtnet_info *vi = netdev_priv(dev);
 633         struct receive_queue *rq = vi->rq;
 634         struct bpf_prog *xdp_prog;
 635         struct send_queue *sq;
 636         unsigned int len;
 637         int packets = 0;
 638         int bytes = 0;
 639         int nxmit = 0;
 640         int kicks = 0;
 641         void *ptr;
 642         int ret;
 643         int i;
 644
 645         /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
 646          * indicate XDP resources have been successfully allocated.
 647          */
 648         xdp_prog = rcu_access_pointer(rq->xdp_prog);
 649         if (!xdp_prog)
 650                 return -ENXIO;
 651
 652         sq = virtnet_xdp_get_sq(vi);
 653
 654         if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
 655                 ret = -EINVAL;
 656                 goto out;
 657         }
 658
 659         /* Free up any pending old buffers before queueing new ones. */
 660         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
 661                 if (likely(is_xdp_frame(ptr))) {
 662                         struct xdp_frame *frame = ptr_to_xdp(ptr);
 663
 664                         bytes += frame->len;
 665                         xdp_return_frame(frame);
 666                 } else {
 667                         struct sk_buff *skb = ptr;
 668
 669                         bytes += skb->len;
 670                         napi_consume_skb(skb, false);
 671                 }
 672                 packets++;
 673         }
 674
 675         for (i = 0; i < n; i++) {
 676                 struct xdp_frame *xdpf = frames[i];
 677
 678                 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
 679                         break;
 680                 nxmit++;
 681         }
 682         ret = nxmit;
 683
 684         if (flags & XDP_XMIT_FLUSH) {
 685                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
 686                         kicks = 1;
 687         }
 688 out:
 689         u64_stats_update_begin(&sq->stats.syncp);
 690         sq->stats.bytes += bytes;
 691         sq->stats.packets += packets;
 692         sq->stats.xdp_tx += n;
 693         sq->stats.xdp_tx_drops += n - nxmit;
 694         sq->stats.kicks += kicks;
 695         u64_stats_update_end(&sq->stats.syncp);
 696
 697         virtnet_xdp_put_sq(vi, sq);
 698         return ret;
 699 }
 700
 701 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
 702 {
 703         return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
 704 }
 705
 706 /* We copy the packet for XDP in the following cases:
 707  *
 708  * 1) Packet is scattered across multiple rx buffers.
 709  * 2) Headroom space is insufficient.
 710  *
 711  * This is inefficient but it's a temporary condition that
 712  * we hit right after XDP is enabled and until queue is refilled
 713  * with large buffers with sufficient headroom - so it should affect
 714  * at most queue size packets.
 715  * Afterwards, the conditions to enable
 716  * XDP should preclude the underlying device from sending packets
 717  * across multiple buffers (num_buf > 1), and we make sure buffers
 718  * have enough headroom.
 719  */
 720 static struct page *xdp_linearize_page(struct receive_queue *rq,
 721                                        u16 *num_buf,
 722                                        struct page *p,
 723                                        int offset,
 724                                        int page_off,
 725                                        unsigned int *len)
 726 {
 727         struct page *page = alloc_page(GFP_ATOMIC);
 728
 729         if (!page)
 730                 return NULL;
 731
 732         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
 733         page_off += *len;
 734
 735         while (--*num_buf) {
 736                 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 737                 unsigned int buflen;
 738                 void *buf;
 739                 int off;
 740
 741                 buf = virtqueue_get_buf(rq->vq, &buflen);
 742                 if (unlikely(!buf))
 743                         goto err_buf;
 744
 745                 p = virt_to_head_page(buf);
 746                 off = buf - page_address(p);
 747
 748                 /* guard against a misconfigured or uncooperative backend that
 749                  * is sending packet larger than the MTU.
 750                  */
 751                 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
 752                         put_page(p);
 753                         goto err_buf;
 754                 }
 755
 756                 memcpy(page_address(page) + page_off,
 757                        page_address(p) + off, buflen);
 758                 page_off += buflen;
 759                 put_page(p);
 760         }
 761
 762         /* Headroom does not contribute to packet length */
 763         *len = page_off - VIRTIO_XDP_HEADROOM;
 764         return page;
 765 err_buf:
 766         __free_pages(page, 0);
 767         return NULL;
 768 }
 769
 770 static struct sk_buff *receive_small(struct net_device *dev,
 771                                      struct virtnet_info *vi,
 772                                      struct receive_queue *rq,
 773                                      void *buf, void *ctx,
 774                                      unsigned int len,
 775                                      unsigned int *xdp_xmit,
 776                                      struct virtnet_rq_stats *stats)
 777 {
 778         struct sk_buff *skb;
 779         struct bpf_prog *xdp_prog;
 780         unsigned int xdp_headroom = (unsigned long)ctx;
 781         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
 782         unsigned int headroom = vi->hdr_len + header_offset;
 783         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 784                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 785         struct page *page = virt_to_head_page(buf);
 786         unsigned int delta = 0;
 787         struct page *xdp_page;
 788         int err;
 789         unsigned int metasize = 0;
 790
 791         len -= vi->hdr_len;
 792         stats->bytes += len;
 793
 794         if (unlikely(len > GOOD_PACKET_LEN)) {
 795                 pr_debug("%s: rx error: len %u exceeds max size %d\n",
 796                          dev->name, len, GOOD_PACKET_LEN);
 797                 dev->stats.rx_length_errors++;
 798                 goto err;
 799         }
 800
 801         if (likely(!vi->xdp_enabled)) {
 802                 xdp_prog = NULL;
 803                 goto skip_xdp;
 804         }
 805
 806         rcu_read_lock();
 807         xdp_prog = rcu_dereference(rq->xdp_prog);
 808         if (xdp_prog) {
 809                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
 810                 struct xdp_frame *xdpf;
 811                 struct xdp_buff xdp;
 812                 void *orig_data;
 813                 u32 act;
 814
 815                 if (unlikely(hdr->hdr.gso_type))
 816                         goto err_xdp;
 817
 818                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
 819                         int offset = buf - page_address(page) + header_offset;
 820                         unsigned int tlen = len + vi->hdr_len;
 821                         u16 num_buf = 1;
 822
 823                         xdp_headroom = virtnet_get_headroom(vi);
 824                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
 825                         headroom = vi->hdr_len + header_offset;
 826                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
 827                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 828                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
 829                                                       offset, header_offset,
 830                                                       &tlen);
 831                         if (!xdp_page)
 832                                 goto err_xdp;
 833
 834                         buf = page_address(xdp_page);
 835                         put_page(page);
 836                         page = xdp_page;
 837                 }
 838
 839                 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
 840                 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
 841                                  xdp_headroom, len, true);
 842                 orig_data = xdp.data;
 843                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 844                 stats->xdp_packets++;
 845
 846                 switch (act) {
 847                 case XDP_PASS:
 848                         /* Recalculate length in case bpf program changed it */
 849                         delta = orig_data - xdp.data;
 850                         len = xdp.data_end - xdp.data;
 851                         metasize = xdp.data - xdp.data_meta;
 852                         break;
 853                 case XDP_TX:
 854                         stats->xdp_tx++;
 855                         xdpf = xdp_convert_buff_to_frame(&xdp);
 856                         if (unlikely(!xdpf))
 857                                 goto err_xdp;
 858                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
 859                         if (unlikely(!err)) {
 860                                 xdp_return_frame_rx_napi(xdpf);
 861                         } else if (unlikely(err < 0)) {
 862                                 trace_xdp_exception(vi->dev, xdp_prog, act);
 863                                 goto err_xdp;
 864                         }
 865                         *xdp_xmit |= VIRTIO_XDP_TX;
 866                         rcu_read_unlock();
 867                         goto xdp_xmit;
 868                 case XDP_REDIRECT:
 869                         stats->xdp_redirects++;
 870                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
 871                         if (err)
 872                                 goto err_xdp;
 873                         *xdp_xmit |= VIRTIO_XDP_REDIR;
 874                         rcu_read_unlock();
 875                         goto xdp_xmit;
 876                 default:
 877                         bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act);
 878                         fallthrough;
 879                 case XDP_ABORTED:
 880                         trace_xdp_exception(vi->dev, xdp_prog, act);
 881                         goto err_xdp;
 882                 case XDP_DROP:
 883                         goto err_xdp;
 884                 }
 885         }
 886         rcu_read_unlock();
 887
 888 skip_xdp:
 889         skb = build_skb(buf, buflen);
 890         if (!skb)
 891                 goto err;
 892         skb_reserve(skb, headroom - delta);
 893         skb_put(skb, len);
 894         if (!xdp_prog) {
 895                 buf += header_offset;
 896                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 897         } /* keep zeroed vnet hdr since XDP is loaded */
 898
 899         if (metasize)
 900                 skb_metadata_set(skb, metasize);
 901
 902         return skb;
 903
 904 err_xdp:
 905         rcu_read_unlock();
 906         stats->xdp_drops++;
 907 err:
 908         stats->drops++;
 909         put_page(page);
 910 xdp_xmit:
 911         return NULL;
 912 }
 913
 914 static struct sk_buff *receive_big(struct net_device *dev,
 915                                    struct virtnet_info *vi,
 916                                    struct receive_queue *rq,
 917                                    void *buf,
 918                                    unsigned int len,
 919                                    struct virtnet_rq_stats *stats)
 920 {
 921         struct page *page = buf;
 922         struct sk_buff *skb =
 923                 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
 924
 925         stats->bytes += len - vi->hdr_len;
 926         if (unlikely(!skb))
 927                 goto err;
 928
 929         return skb;
 930
 931 err:
 932         stats->drops++;
 933         give_pages(rq, page);
 934         return NULL;
 935 }
 936
 937 static struct sk_buff *receive_mergeable(struct net_device *dev,
 938                                          struct virtnet_info *vi,
 939                                          struct receive_queue *rq,
 940                                          void *buf,
 941                                          void *ctx,
 942                                          unsigned int len,
 943                                          unsigned int *xdp_xmit,
 944                                          struct virtnet_rq_stats *stats)
 945 {
 946         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
 947         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
 948         struct page *page = virt_to_head_page(buf);
 949         int offset = buf - page_address(page);
 950         struct sk_buff *head_skb, *curr_skb;
 951         struct bpf_prog *xdp_prog;
 952         unsigned int truesize = mergeable_ctx_to_truesize(ctx);
 953         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
 954         unsigned int metasize = 0;
 955         unsigned int frame_sz;
 956         int err;
 957
 958         head_skb = NULL;
 959         stats->bytes += len - vi->hdr_len;
 960
 961         if (unlikely(len > truesize)) {
 962                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
 963                          dev->name, len, (unsigned long)ctx);
 964                 dev->stats.rx_length_errors++;
 965                 goto err_skb;
 966         }
 967
 968         if (likely(!vi->xdp_enabled)) {
 969                 xdp_prog = NULL;
 970                 goto skip_xdp;
 971         }
 972
 973         rcu_read_lock();
 974         xdp_prog = rcu_dereference(rq->xdp_prog);
 975         if (xdp_prog) {
 976                 struct xdp_frame *xdpf;
 977                 struct page *xdp_page;
 978                 struct xdp_buff xdp;
 979                 void *data;
 980                 u32 act;
 981
 982                 /* Transient failure which in theory could occur if
 983                  * in-flight packets from before XDP was enabled reach
 984                  * the receive path after XDP is loaded.
 985                  */
 986                 if (unlikely(hdr->hdr.gso_type))
 987                         goto err_xdp;
 988
 989                 /* Buffers with headroom use PAGE_SIZE as alloc size,
 990                  * see add_recvbuf_mergeable() + get_mergeable_buf_len()
 991                  */
 992                 frame_sz = headroom ? PAGE_SIZE : truesize;
 993
 994                 /* This happens when rx buffer size is underestimated
 995                  * or headroom is not enough because of the buffer
 996                  * was refilled before XDP is set. This should only
 997                  * happen for the first several packets, so we don't
 998                  * care much about its performance.
 999                  */
1000                 if (unlikely(num_buf > 1 ||
1001                              headroom < virtnet_get_headroom(vi))) {
1002                         /* linearize data for XDP */
1003                         xdp_page = xdp_linearize_page(rq, &num_buf,
1004                                                       page, offset,
1005                                                       VIRTIO_XDP_HEADROOM,
1006                                                       &len);
1007                         frame_sz = PAGE_SIZE;
1008
1009                         if (!xdp_page)
1010                                 goto err_xdp;
1011                         offset = VIRTIO_XDP_HEADROOM;
1012                 } else {
1013                         xdp_page = page;
1014                 }
1015
1016                 /* Allow consuming headroom but reserve enough space to push
1017                  * the descriptor on if we get an XDP_TX return code.
1018                  */
1019                 data = page_address(xdp_page) + offset;
1020                 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
1021                 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
1022                                  VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
1023
1024                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
1025                 stats->xdp_packets++;
1026
1027                 switch (act) {
1028                 case XDP_PASS:
1029                         metasize = xdp.data - xdp.data_meta;
1030
1031                         /* recalculate offset to account for any header
1032                          * adjustments and minus the metasize to copy the
1033                          * metadata in page_to_skb(). Note other cases do not
1034                          * build an skb and avoid using offset
1035                          */
1036                         offset = xdp.data - page_address(xdp_page) -
1037                                  vi->hdr_len - metasize;
1038
1039                         /* recalculate len if xdp.data, xdp.data_end or
1040                          * xdp.data_meta were adjusted
1041                          */
1042                         len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
1043
1044                         /* recalculate headroom if xdp.data or xdp_data_meta
1045                          * were adjusted, note that offset should always point
1046                          * to the start of the reserved bytes for virtio_net
1047                          * header which are followed by xdp.data, that means
1048                          * that offset is equal to the headroom (when buf is
1049                          * starting at the beginning of the page, otherwise
1050                          * there is a base offset inside the page) but it's used
1051                          * with a different starting point (buf start) than
1052                          * xdp.data (buf start + vnet hdr size). If xdp.data or
1053                          * data_meta were adjusted by the xdp prog then the
1054                          * headroom size has changed and so has the offset, we
1055                          * can use data_hard_start, which points at buf start +
1056                          * vnet hdr size, to calculate the new headroom and use
1057                          * it later to compute buf start in page_to_skb()
1058                          */
1059                         headroom = xdp.data - xdp.data_hard_start - metasize;
1060
1061                         /* We can only create skb based on xdp_page. */
1062                         if (unlikely(xdp_page != page)) {
1063                                 rcu_read_unlock();
1064                                 put_page(page);
1065                                 head_skb = page_to_skb(vi, rq, xdp_page, offset,
1066                                                        len, PAGE_SIZE, false,
1067                                                        metasize,
1068                                                        headroom);
1069                                 return head_skb;
1070                         }
1071                         break;
1072                 case XDP_TX:
1073                         stats->xdp_tx++;
1074                         xdpf = xdp_convert_buff_to_frame(&xdp);
1075                         if (unlikely(!xdpf)) {
1076                                 if (unlikely(xdp_page != page))
1077                                         put_page(xdp_page);
1078                                 goto err_xdp;
1079                         }
1080                         err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
1081                         if (unlikely(!err)) {
1082                                 xdp_return_frame_rx_napi(xdpf);
1083                         } else if (unlikely(err < 0)) {
1084                                 trace_xdp_exception(vi->dev, xdp_prog, act);
1085                                 if (unlikely(xdp_page != page))
1086                                         put_page(xdp_page);
1087                                 goto err_xdp;
1088                         }
1089                         *xdp_xmit |= VIRTIO_XDP_TX;
1090                         if (unlikely(xdp_page != page))
1091                                 put_page(page);
1092                         rcu_read_unlock();
1093                         goto xdp_xmit;
1094                 case XDP_REDIRECT:
1095                         stats->xdp_redirects++;
1096                         err = xdp_do_redirect(dev, &xdp, xdp_prog);
1097                         if (err) {
1098                                 if (unlikely(xdp_page != page))
1099                                         put_page(xdp_page);
1100                                 goto err_xdp;
1101                         }
1102                         *xdp_xmit |= VIRTIO_XDP_REDIR;
1103                         if (unlikely(xdp_page != page))
1104                                 put_page(page);
1105                         rcu_read_unlock();
1106                         goto xdp_xmit;
1107                 default:
1108                         bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act);
1109                         fallthrough;
1110                 case XDP_ABORTED:
1111                         trace_xdp_exception(vi->dev, xdp_prog, act);
1112                         fallthrough;
1113                 case XDP_DROP:
1114                         if (unlikely(xdp_page != page))
1115                                 __free_pages(xdp_page, 0);
1116                         goto err_xdp;
1117                 }
1118         }
1119         rcu_read_unlock();
1120
1121 skip_xdp:
1122         head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1123                                metasize, headroom);
1124         curr_skb = head_skb;
1125
1126         if (unlikely(!curr_skb))
1127                 goto err_skb;
1128         while (--num_buf) {
1129                 int num_skb_frags;
1130
1131                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1132                 if (unlikely(!buf)) {
1133                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
1134                                  dev->name, num_buf,
1135                                  virtio16_to_cpu(vi->vdev,
1136                                                  hdr->num_buffers));
1137                         dev->stats.rx_length_errors++;
1138                         goto err_buf;
1139                 }
1140
1141                 stats->bytes += len;
1142                 page = virt_to_head_page(buf);
1143
1144                 truesize = mergeable_ctx_to_truesize(ctx);
1145                 if (unlikely(len > truesize)) {
1146                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1147                                  dev->name, len, (unsigned long)ctx);
1148                         dev->stats.rx_length_errors++;
1149                         goto err_skb;
1150                 }
1151
1152                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1153                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1154                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1155
1156                         if (unlikely(!nskb))
1157                                 goto err_skb;
1158                         if (curr_skb == head_skb)
1159                                 skb_shinfo(curr_skb)->frag_list = nskb;
1160                         else
1161                                 curr_skb->next = nskb;
1162                         curr_skb = nskb;
1163                         head_skb->truesize += nskb->truesize;
1164                         num_skb_frags = 0;
1165                 }
1166                 if (curr_skb != head_skb) {
1167                         head_skb->data_len += len;
1168                         head_skb->len += len;
1169                         head_skb->truesize += truesize;
1170                 }
1171                 offset = buf - page_address(page);
1172                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1173                         put_page(page);
1174                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1175                                              len, truesize);
1176                 } else {
1177                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
1178                                         offset, len, truesize);
1179                 }
1180         }
1181
1182         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1183         return head_skb;
1184
1185 err_xdp:
1186         rcu_read_unlock();
1187         stats->xdp_drops++;
1188 err_skb:
1189         put_page(page);
1190         while (num_buf-- > 1) {
1191                 buf = virtqueue_get_buf(rq->vq, &len);
1192                 if (unlikely(!buf)) {
1193                         pr_debug("%s: rx error: %d buffers missing\n",
1194                                  dev->name, num_buf);
1195                         dev->stats.rx_length_errors++;
1196                         break;
1197                 }
1198                 stats->bytes += len;
1199                 page = virt_to_head_page(buf);
1200                 put_page(page);
1201         }
1202 err_buf:
1203         stats->drops++;
1204         dev_kfree_skb(head_skb);
1205 xdp_xmit:
1206         return NULL;
1207 }
1208
1209 static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash,
1210                                 struct sk_buff *skb)
1211 {
1212         enum pkt_hash_types rss_hash_type;
1213
1214         if (!hdr_hash || !skb)
1215                 return;
1216
1217         switch (__le16_to_cpu(hdr_hash->hash_report)) {
1218         case VIRTIO_NET_HASH_REPORT_TCPv4:
1219         case VIRTIO_NET_HASH_REPORT_UDPv4:
1220         case VIRTIO_NET_HASH_REPORT_TCPv6:
1221         case VIRTIO_NET_HASH_REPORT_UDPv6:
1222         case VIRTIO_NET_HASH_REPORT_TCPv6_EX:
1223         case VIRTIO_NET_HASH_REPORT_UDPv6_EX:
1224                 rss_hash_type = PKT_HASH_TYPE_L4;
1225                 break;
1226         case VIRTIO_NET_HASH_REPORT_IPv4:
1227         case VIRTIO_NET_HASH_REPORT_IPv6:
1228         case VIRTIO_NET_HASH_REPORT_IPv6_EX:
1229                 rss_hash_type = PKT_HASH_TYPE_L3;
1230                 break;
1231         case VIRTIO_NET_HASH_REPORT_NONE:
1232         default:
1233                 rss_hash_type = PKT_HASH_TYPE_NONE;
1234         }
1235         skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type);
1236 }
1237
1238 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1239                         void *buf, unsigned int len, void **ctx,
1240                         unsigned int *xdp_xmit,
1241                         struct virtnet_rq_stats *stats)
1242 {
1243         struct net_device *dev = vi->dev;
1244         struct sk_buff *skb;
1245         struct virtio_net_hdr_mrg_rxbuf *hdr;
1246
1247         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1248                 pr_debug("%s: short packet %i\n", dev->name, len);
1249                 dev->stats.rx_length_errors++;
1250                 if (vi->mergeable_rx_bufs) {
1251                         put_page(virt_to_head_page(buf));
1252                 } else if (vi->big_packets) {
1253                         give_pages(rq, buf);
1254                 } else {
1255                         put_page(virt_to_head_page(buf));
1256                 }
1257                 return;
1258         }
1259
1260         if (vi->mergeable_rx_bufs)
1261                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1262                                         stats);
1263         else if (vi->big_packets)
1264                 skb = receive_big(dev, vi, rq, buf, len, stats);
1265         else
1266                 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1267
1268         if (unlikely(!skb))
1269                 return;
1270
1271         hdr = skb_vnet_hdr(skb);
1272         if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report)
1273                 virtio_skb_set_hash((const struct virtio_net_hdr_v1_hash *)hdr, skb);
1274
1275         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1276                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1277
1278         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1279                                   virtio_is_little_endian(vi->vdev))) {
1280                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1281                                      dev->name, hdr->hdr.gso_type,
1282                                      hdr->hdr.gso_size);
1283                 goto frame_err;
1284         }
1285
1286         skb_record_rx_queue(skb, vq2rxq(rq->vq));
1287         skb->protocol = eth_type_trans(skb, dev);
1288         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1289                  ntohs(skb->protocol), skb->len, skb->pkt_type);
1290
1291         napi_gro_receive(&rq->napi, skb);
1292         return;
1293
1294 frame_err:
1295         dev->stats.rx_frame_errors++;
1296         dev_kfree_skb(skb);
1297 }
1298
1299 /* Unlike mergeable buffers, all buffers are allocated to the
1300  * same size, except for the headroom. For this reason we do
1301  * not need to use  mergeable_len_to_ctx here - it is enough
1302  * to store the headroom as the context ignoring the truesize.
1303  */
1304 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1305                              gfp_t gfp)
1306 {
1307         struct page_frag *alloc_frag = &rq->alloc_frag;
1308         char *buf;
1309         unsigned int xdp_headroom = virtnet_get_headroom(vi);
1310         void *ctx = (void *)(unsigned long)xdp_headroom;
1311         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1312         int err;
1313
1314         len = SKB_DATA_ALIGN(len) +
1315               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1316         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1317                 return -ENOMEM;
1318
1319         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1320         get_page(alloc_frag->page);
1321         alloc_frag->offset += len;
1322         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1323                     vi->hdr_len + GOOD_PACKET_LEN);
1324         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1325         if (err < 0)
1326                 put_page(virt_to_head_page(buf));
1327         return err;
1328 }
1329
1330 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1331                            gfp_t gfp)
1332 {
1333         struct page *first, *list = NULL;
1334         char *p;
1335         int i, err, offset;
1336
1337         sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2);
1338
1339         /* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */
1340         for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) {
1341                 first = get_a_page(rq, gfp);
1342                 if (!first) {
1343                         if (list)
1344                                 give_pages(rq, list);
1345                         return -ENOMEM;
1346                 }
1347                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1348
1349                 /* chain new page in list head to match sg */
1350                 first->private = (unsigned long)list;
1351                 list = first;
1352         }
1353
1354         first = get_a_page(rq, gfp);
1355         if (!first) {
1356                 give_pages(rq, list);
1357                 return -ENOMEM;
1358         }
1359         p = page_address(first);
1360
1361         /* rq->sg[0], rq->sg[1] share the same page */
1362         /* a separated rq->sg[0] for header - required in case !any_header_sg */
1363         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1364
1365         /* rq->sg[1] for data packet, from offset */
1366         offset = sizeof(struct padded_vnet_hdr);
1367         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1368
1369         /* chain first in list head */
1370         first->private = (unsigned long)list;
1371         err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2,
1372                                   first, gfp);
1373         if (err < 0)
1374                 give_pages(rq, first);
1375
1376         return err;
1377 }
1378
1379 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1380                                           struct ewma_pkt_len *avg_pkt_len,
1381                                           unsigned int room)
1382 {
1383         struct virtnet_info *vi = rq->vq->vdev->priv;
1384         const size_t hdr_len = vi->hdr_len;
1385         unsigned int len;
1386
1387         if (room)
1388                 return PAGE_SIZE - room;
1389
1390         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1391                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
1392
1393         return ALIGN(len, L1_CACHE_BYTES);
1394 }
1395
1396 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1397                                  struct receive_queue *rq, gfp_t gfp)
1398 {
1399         struct page_frag *alloc_frag = &rq->alloc_frag;
1400         unsigned int headroom = virtnet_get_headroom(vi);
1401         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1402         unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1403         char *buf;
1404         void *ctx;
1405         int err;
1406         unsigned int len, hole;
1407
1408         /* Extra tailroom is needed to satisfy XDP's assumption. This
1409          * means rx frags coalescing won't work, but consider we've
1410          * disabled GSO for XDP, it won't be a big issue.
1411          */
1412         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1413         if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1414                 return -ENOMEM;
1415
1416         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1417         buf += headroom; /* advance address leaving hole at front of pkt */
1418         get_page(alloc_frag->page);
1419         alloc_frag->offset += len + room;
1420         hole = alloc_frag->size - alloc_frag->offset;
1421         if (hole < len + room) {
1422                 /* To avoid internal fragmentation, if there is very likely not
1423                  * enough space for another buffer, add the remaining space to
1424                  * the current buffer.
1425                  */
1426                 len += hole;
1427                 alloc_frag->offset += hole;
1428         }
1429
1430         sg_init_one(rq->sg, buf, len);
1431         ctx = mergeable_len_to_ctx(len, headroom);
1432         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1433         if (err < 0)
1434                 put_page(virt_to_head_page(buf));
1435
1436         return err;
1437 }
1438
1439 /*
1440  * Returns false if we couldn't fill entirely (OOM).
1441  *
1442  * Normally run in the receive path, but can also be run from ndo_open
1443  * before we're receiving packets, or from refill_work which is
1444  * careful to disable receiving (using napi_disable).
1445  */
1446 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1447                           gfp_t gfp)
1448 {
1449         int err;
1450         bool oom;
1451
1452         do {
1453                 if (vi->mergeable_rx_bufs)
1454                         err = add_recvbuf_mergeable(vi, rq, gfp);
1455                 else if (vi->big_packets)
1456                         err = add_recvbuf_big(vi, rq, gfp);
1457                 else
1458                         err = add_recvbuf_small(vi, rq, gfp);
1459
1460                 oom = err == -ENOMEM;
1461                 if (err)
1462                         break;
1463         } while (rq->vq->num_free);
1464         if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1465                 unsigned long flags;
1466
1467                 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1468                 rq->stats.kicks++;
1469                 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1470         }
1471
1472         return !oom;
1473 }
1474
1475 static void skb_recv_done(struct virtqueue *rvq)
1476 {
1477         struct virtnet_info *vi = rvq->vdev->priv;
1478         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1479
1480         virtqueue_napi_schedule(&rq->napi, rvq);
1481 }
1482
1483 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1484 {
1485         napi_enable(napi);
1486
1487         /* If all buffers were filled by other side before we napi_enabled, we
1488          * won't get another interrupt, so process any outstanding packets now.
1489          * Call local_bh_enable after to trigger softIRQ processing.
1490          */
1491         local_bh_disable();
1492         virtqueue_napi_schedule(napi, vq);
1493         local_bh_enable();
1494 }
1495
1496 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1497                                    struct virtqueue *vq,
1498                                    struct napi_struct *napi)
1499 {
1500         if (!napi->weight)
1501                 return;
1502
1503         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1504          * enable the feature if this is likely affine with the transmit path.
1505          */
1506         if (!vi->affinity_hint_set) {
1507                 napi->weight = 0;
1508                 return;
1509         }
1510
1511         return virtnet_napi_enable(vq, napi);
1512 }
1513
1514 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1515 {
1516         if (napi->weight)
1517                 napi_disable(napi);
1518 }
1519
1520 static void refill_work(struct work_struct *work)
1521 {
1522         struct virtnet_info *vi =
1523                 container_of(work, struct virtnet_info, refill.work);
1524         bool still_empty;
1525         int i;
1526
1527         for (i = 0; i < vi->curr_queue_pairs; i++) {
1528                 struct receive_queue *rq = &vi->rq[i];
1529
1530                 napi_disable(&rq->napi);
1531                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1532                 virtnet_napi_enable(rq->vq, &rq->napi);
1533
1534                 /* In theory, this can happen: if we don't get any buffers in
1535                  * we will *never* try to fill again.
1536                  */
1537                 if (still_empty)
1538                         schedule_delayed_work(&vi->refill, HZ/2);
1539         }
1540 }
1541
1542 static int virtnet_receive(struct receive_queue *rq, int budget,
1543                            unsigned int *xdp_xmit)
1544 {
1545         struct virtnet_info *vi = rq->vq->vdev->priv;
1546         struct virtnet_rq_stats stats = {};
1547         unsigned int len;
1548         void *buf;
1549         int i;
1550
1551         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1552                 void *ctx;
1553
1554                 while (stats.packets < budget &&
1555                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1556                         receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1557                         stats.packets++;
1558                 }
1559         } else {
1560                 while (stats.packets < budget &&
1561                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1562                         receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1563                         stats.packets++;
1564                 }
1565         }
1566
1567         if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1568                 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
1569                         spin_lock(&vi->refill_lock);
1570                         if (vi->refill_enabled)
1571                                 schedule_delayed_work(&vi->refill, 0);
1572                         spin_unlock(&vi->refill_lock);
1573                 }
1574         }
1575
1576         u64_stats_update_begin(&rq->stats.syncp);
1577         for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1578                 size_t offset = virtnet_rq_stats_desc[i].offset;
1579                 u64 *item;
1580
1581                 item = (u64 *)((u8 *)&rq->stats + offset);
1582                 *item += *(u64 *)((u8 *)&stats + offset);
1583         }
1584         u64_stats_update_end(&rq->stats.syncp);
1585
1586         return stats.packets;
1587 }
1588
1589 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1590 {
1591         unsigned int len;
1592         unsigned int packets = 0;
1593         unsigned int bytes = 0;
1594         void *ptr;
1595
1596         while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1597                 if (likely(!is_xdp_frame(ptr))) {
1598                         struct sk_buff *skb = ptr;
1599
1600                         pr_debug("Sent skb %p\n", skb);
1601
1602                         bytes += skb->len;
1603                         napi_consume_skb(skb, in_napi);
1604                 } else {
1605                         struct xdp_frame *frame = ptr_to_xdp(ptr);
1606
1607                         bytes += frame->len;
1608                         xdp_return_frame(frame);
1609                 }
1610                 packets++;
1611         }
1612
1613         /* Avoid overhead when no packets have been processed
1614          * happens when called speculatively from start_xmit.
1615          */
1616         if (!packets)
1617                 return;
1618
1619         u64_stats_update_begin(&sq->stats.syncp);
1620         sq->stats.bytes += bytes;
1621         sq->stats.packets += packets;
1622         u64_stats_update_end(&sq->stats.syncp);
1623 }
1624
1625 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1626 {
1627         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1628                 return false;
1629         else if (q < vi->curr_queue_pairs)
1630                 return true;
1631         else
1632                 return false;
1633 }
1634
1635 static void virtnet_poll_cleantx(struct receive_queue *rq)
1636 {
1637         struct virtnet_info *vi = rq->vq->vdev->priv;
1638         unsigned int index = vq2rxq(rq->vq);
1639         struct send_queue *sq = &vi->sq[index];
1640         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1641
1642         if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1643                 return;
1644
1645         if (__netif_tx_trylock(txq)) {
1646                 if (sq->reset) {
1647                         __netif_tx_unlock(txq);
1648                         return;
1649                 }
1650
1651                 do {
1652                         virtqueue_disable_cb(sq->vq);
1653                         free_old_xmit_skbs(sq, true);
1654                 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1655
1656                 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1657                         netif_tx_wake_queue(txq);
1658
1659                 __netif_tx_unlock(txq);
1660         }
1661 }
1662
1663 static int virtnet_poll(struct napi_struct *napi, int budget)
1664 {
1665         struct receive_queue *rq =
1666                 container_of(napi, struct receive_queue, napi);
1667         struct virtnet_info *vi = rq->vq->vdev->priv;
1668         struct send_queue *sq;
1669         unsigned int received;
1670         unsigned int xdp_xmit = 0;
1671
1672         virtnet_poll_cleantx(rq);
1673
1674         received = virtnet_receive(rq, budget, &xdp_xmit);
1675
1676         /* Out of packets? */
1677         if (received < budget)
1678                 virtqueue_napi_complete(napi, rq->vq, received);
1679
1680         if (xdp_xmit & VIRTIO_XDP_REDIR)
1681                 xdp_do_flush();
1682
1683         if (xdp_xmit & VIRTIO_XDP_TX) {
1684                 sq = virtnet_xdp_get_sq(vi);
1685                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1686                         u64_stats_update_begin(&sq->stats.syncp);
1687                         sq->stats.kicks++;
1688                         u64_stats_update_end(&sq->stats.syncp);
1689                 }
1690                 virtnet_xdp_put_sq(vi, sq);
1691         }
1692
1693         return received;
1694 }
1695
1696 static int virtnet_open(struct net_device *dev)
1697 {
1698         struct virtnet_info *vi = netdev_priv(dev);
1699         int i, err;
1700
1701         enable_delayed_refill(vi);
1702
1703         for (i = 0; i < vi->max_queue_pairs; i++) {
1704                 if (i < vi->curr_queue_pairs)
1705                         /* Make sure we have some buffers: if oom use wq. */
1706                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1707                                 schedule_delayed_work(&vi->refill, 0);
1708
1709                 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1710                 if (err < 0)
1711                         return err;
1712
1713                 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1714                                                  MEM_TYPE_PAGE_SHARED, NULL);
1715                 if (err < 0) {
1716                         xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1717                         return err;
1718                 }
1719
1720                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1721                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1722         }
1723
1724         return 0;
1725 }
1726
1727 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1728 {
1729         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1730         struct virtnet_info *vi = sq->vq->vdev->priv;
1731         unsigned int index = vq2txq(sq->vq);
1732         struct netdev_queue *txq;
1733         int opaque;
1734         bool done;
1735
1736         if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1737                 /* We don't need to enable cb for XDP */
1738                 napi_complete_done(napi, 0);
1739                 return 0;
1740         }
1741
1742         txq = netdev_get_tx_queue(vi->dev, index);
1743         __netif_tx_lock(txq, raw_smp_processor_id());
1744         virtqueue_disable_cb(sq->vq);
1745         free_old_xmit_skbs(sq, true);
1746
1747         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1748                 netif_tx_wake_queue(txq);
1749
1750         opaque = virtqueue_enable_cb_prepare(sq->vq);
1751
1752         done = napi_complete_done(napi, 0);
1753
1754         if (!done)
1755                 virtqueue_disable_cb(sq->vq);
1756
1757         __netif_tx_unlock(txq);
1758
1759         if (done) {
1760                 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1761                         if (napi_schedule_prep(napi)) {
1762                                 __netif_tx_lock(txq, raw_smp_processor_id());
1763                                 virtqueue_disable_cb(sq->vq);
1764                                 __netif_tx_unlock(txq);
1765                                 __napi_schedule(napi);
1766                         }
1767                 }
1768         }
1769
1770         return 0;
1771 }
1772
1773 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1774 {
1775         struct virtio_net_hdr_mrg_rxbuf *hdr;
1776         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1777         struct virtnet_info *vi = sq->vq->vdev->priv;
1778         int num_sg;
1779         unsigned hdr_len = vi->hdr_len;
1780         bool can_push;
1781
1782         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1783
1784         can_push = vi->any_header_sg &&
1785                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1786                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1787         /* Even if we can, don't push here yet as this would skew
1788          * csum_start offset below. */
1789         if (can_push)
1790                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1791         else
1792                 hdr = skb_vnet_hdr(skb);
1793
1794         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1795                                     virtio_is_little_endian(vi->vdev), false,
1796                                     0))
1797                 return -EPROTO;
1798
1799         if (vi->mergeable_rx_bufs)
1800                 hdr->num_buffers = 0;
1801
1802         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1803         if (can_push) {
1804                 __skb_push(skb, hdr_len);
1805                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1806                 if (unlikely(num_sg < 0))
1807                         return num_sg;
1808                 /* Pull header back to avoid skew in tx bytes calculations. */
1809                 __skb_pull(skb, hdr_len);
1810         } else {
1811                 sg_set_buf(sq->sg, hdr, hdr_len);
1812                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1813                 if (unlikely(num_sg < 0))
1814                         return num_sg;
1815                 num_sg++;
1816         }
1817         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1818 }
1819
1820 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1821 {
1822         struct virtnet_info *vi = netdev_priv(dev);
1823         int qnum = skb_get_queue_mapping(skb);
1824         struct send_queue *sq = &vi->sq[qnum];
1825         int err;
1826         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1827         bool kick = !netdev_xmit_more();
1828         bool use_napi = sq->napi.weight;
1829
1830         /* Free up any pending old buffers before queueing new ones. */
1831         do {
1832                 if (use_napi)
1833                         virtqueue_disable_cb(sq->vq);
1834
1835                 free_old_xmit_skbs(sq, false);
1836
1837         } while (use_napi && kick &&
1838                unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1839
1840         /* timestamp packet in software */
1841         skb_tx_timestamp(skb);
1842
1843         /* Try to transmit */
1844         err = xmit_skb(sq, skb);
1845
1846         /* This should not happen! */
1847         if (unlikely(err)) {
1848                 dev->stats.tx_fifo_errors++;
1849                 if (net_ratelimit())
1850                         dev_warn(&dev->dev,
1851                                  "Unexpected TXQ (%d) queue failure: %d\n",
1852                                  qnum, err);
1853                 dev->stats.tx_dropped++;
1854                 dev_kfree_skb_any(skb);
1855                 return NETDEV_TX_OK;
1856         }
1857
1858         /* Don't wait up for transmitted skbs to be freed. */
1859         if (!use_napi) {
1860                 skb_orphan(skb);
1861                 nf_reset_ct(skb);
1862         }
1863
1864         /* If running out of space, stop queue to avoid getting packets that we
1865          * are then unable to transmit.
1866          * An alternative would be to force queuing layer to requeue the skb by
1867          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1868          * returned in a normal path of operation: it means that driver is not
1869          * maintaining the TX queue stop/start state properly, and causes
1870          * the stack to do a non-trivial amount of useless work.
1871          * Since most packets only take 1 or 2 ring slots, stopping the queue
1872          * early means 16 slots are typically wasted.
1873          */
1874         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1875                 netif_stop_subqueue(dev, qnum);
1876                 if (!use_napi &&
1877                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1878                         /* More just got used, free them then recheck. */
1879                         free_old_xmit_skbs(sq, false);
1880                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1881                                 netif_start_subqueue(dev, qnum);
1882                                 virtqueue_disable_cb(sq->vq);
1883                         }
1884                 }
1885         }
1886
1887         if (kick || netif_xmit_stopped(txq)) {
1888                 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1889                         u64_stats_update_begin(&sq->stats.syncp);
1890                         sq->stats.kicks++;
1891                         u64_stats_update_end(&sq->stats.syncp);
1892                 }
1893         }
1894
1895         return NETDEV_TX_OK;
1896 }
1897
1898 static int virtnet_rx_resize(struct virtnet_info *vi,
1899                              struct receive_queue *rq, u32 ring_num)
1900 {
1901         bool running = netif_running(vi->dev);
1902         int err, qindex;
1903
1904         qindex = rq - vi->rq;
1905
1906         if (running)
1907                 napi_disable(&rq->napi);
1908
1909         err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_free_unused_buf);
1910         if (err)
1911                 netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err);
1912
1913         if (!try_fill_recv(vi, rq, GFP_KERNEL))
1914                 schedule_delayed_work(&vi->refill, 0);
1915
1916         if (running)
1917                 virtnet_napi_enable(rq->vq, &rq->napi);
1918         return err;
1919 }
1920
1921 static int virtnet_tx_resize(struct virtnet_info *vi,
1922                              struct send_queue *sq, u32 ring_num)
1923 {
1924         bool running = netif_running(vi->dev);
1925         struct netdev_queue *txq;
1926         int err, qindex;
1927
1928         qindex = sq - vi->sq;
1929
1930         if (running)
1931                 virtnet_napi_tx_disable(&sq->napi);
1932
1933         txq = netdev_get_tx_queue(vi->dev, qindex);
1934
1935         /* 1. wait all ximt complete
1936          * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue()
1937          */
1938         __netif_tx_lock_bh(txq);
1939
1940         /* Prevent rx poll from accessing sq. */
1941         sq->reset = true;
1942
1943         /* Prevent the upper layer from trying to send packets. */
1944         netif_stop_subqueue(vi->dev, qindex);
1945
1946         __netif_tx_unlock_bh(txq);
1947
1948         err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf);
1949         if (err)
1950                 netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err);
1951
1952         __netif_tx_lock_bh(txq);
1953         sq->reset = false;
1954         netif_tx_wake_queue(txq);
1955         __netif_tx_unlock_bh(txq);
1956
1957         if (running)
1958                 virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
1959         return err;
1960 }
1961
1962 /*
1963  * Send command via the control virtqueue and check status.  Commands
1964  * supported by the hypervisor, as indicated by feature bits, should
1965  * never fail unless improperly formatted.
1966  */
1967 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1968                                  struct scatterlist *out)
1969 {
1970         struct scatterlist *sgs[4], hdr, stat;
1971         unsigned out_num = 0, tmp;
1972         int ret;
1973
1974         /* Caller should know better */
1975         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1976
1977         vi->ctrl->status = ~0;
1978         vi->ctrl->hdr.class = class;
1979         vi->ctrl->hdr.cmd = cmd;
1980         /* Add header */
1981         sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1982         sgs[out_num++] = &hdr;
1983
1984         if (out)
1985                 sgs[out_num++] = out;
1986
1987         /* Add return status. */
1988         sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1989         sgs[out_num] = &stat;
1990
1991         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1992         ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1993         if (ret < 0) {
1994                 dev_warn(&vi->vdev->dev,
1995                          "Failed to add sgs for command vq: %d\n.", ret);
1996                 return false;
1997         }
1998
1999         if (unlikely(!virtqueue_kick(vi->cvq)))
2000                 return vi->ctrl->status == VIRTIO_NET_OK;
2001
2002         /* Spin for a response, the kick causes an ioport write, trapping
2003          * into the hypervisor, so the request should be handled immediately.
2004          */
2005         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
2006                !virtqueue_is_broken(vi->cvq))
2007                 cpu_relax();
2008
2009         return vi->ctrl->status == VIRTIO_NET_OK;
2010 }
2011
2012 static int virtnet_set_mac_address(struct net_device *dev, void *p)
2013 {
2014         struct virtnet_info *vi = netdev_priv(dev);
2015         struct virtio_device *vdev = vi->vdev;
2016         int ret;
2017         struct sockaddr *addr;
2018         struct scatterlist sg;
2019
2020         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2021                 return -EOPNOTSUPP;
2022
2023         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
2024         if (!addr)
2025                 return -ENOMEM;
2026
2027         ret = eth_prepare_mac_addr_change(dev, addr);
2028         if (ret)
2029                 goto out;
2030
2031         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
2032                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
2033                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2034                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
2035                         dev_warn(&vdev->dev,
2036                                  "Failed to set mac address by vq command.\n");
2037                         ret = -EINVAL;
2038                         goto out;
2039                 }
2040         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
2041                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
2042                 unsigned int i;
2043
2044                 /* Naturally, this has an atomicity problem. */
2045                 for (i = 0; i < dev->addr_len; i++)
2046                         virtio_cwrite8(vdev,
2047                                        offsetof(struct virtio_net_config, mac) +
2048                                        i, addr->sa_data[i]);
2049         }
2050
2051         eth_commit_mac_addr_change(dev, p);
2052         ret = 0;
2053
2054 out:
2055         kfree(addr);
2056         return ret;
2057 }
2058
2059 static void virtnet_stats(struct net_device *dev,
2060                           struct rtnl_link_stats64 *tot)
2061 {
2062         struct virtnet_info *vi = netdev_priv(dev);
2063         unsigned int start;
2064         int i;
2065
2066         for (i = 0; i < vi->max_queue_pairs; i++) {
2067                 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
2068                 struct receive_queue *rq = &vi->rq[i];
2069                 struct send_queue *sq = &vi->sq[i];
2070
2071                 do {
2072                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2073                         tpackets = sq->stats.packets;
2074                         tbytes   = sq->stats.bytes;
2075                         terrors  = sq->stats.tx_timeouts;
2076                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2077
2078                 do {
2079                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2080                         rpackets = rq->stats.packets;
2081                         rbytes   = rq->stats.bytes;
2082                         rdrops   = rq->stats.drops;
2083                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2084
2085                 tot->rx_packets += rpackets;
2086                 tot->tx_packets += tpackets;
2087                 tot->rx_bytes   += rbytes;
2088                 tot->tx_bytes   += tbytes;
2089                 tot->rx_dropped += rdrops;
2090                 tot->tx_errors  += terrors;
2091         }
2092
2093         tot->tx_dropped = dev->stats.tx_dropped;
2094         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
2095         tot->rx_length_errors = dev->stats.rx_length_errors;
2096         tot->rx_frame_errors = dev->stats.rx_frame_errors;
2097 }
2098
2099 static void virtnet_ack_link_announce(struct virtnet_info *vi)
2100 {
2101         rtnl_lock();
2102         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
2103                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
2104                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
2105         rtnl_unlock();
2106 }
2107
2108 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2109 {
2110         struct scatterlist sg;
2111         struct net_device *dev = vi->dev;
2112
2113         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
2114                 return 0;
2115
2116         vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
2117         sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
2118
2119         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2120                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
2121                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
2122                          queue_pairs);
2123                 return -EINVAL;
2124         } else {
2125                 vi->curr_queue_pairs = queue_pairs;
2126                 /* virtnet_open() will refill when device is going to up. */
2127                 if (dev->flags & IFF_UP)
2128                         schedule_delayed_work(&vi->refill, 0);
2129         }
2130
2131         return 0;
2132 }
2133
2134 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2135 {
2136         int err;
2137
2138         rtnl_lock();
2139         err = _virtnet_set_queues(vi, queue_pairs);
2140         rtnl_unlock();
2141         return err;
2142 }
2143
2144 static int virtnet_close(struct net_device *dev)
2145 {
2146         struct virtnet_info *vi = netdev_priv(dev);
2147         int i;
2148
2149         /* Make sure NAPI doesn't schedule refill work */
2150         disable_delayed_refill(vi);
2151         /* Make sure refill_work doesn't re-enable napi! */
2152         cancel_delayed_work_sync(&vi->refill);
2153
2154         for (i = 0; i < vi->max_queue_pairs; i++) {
2155                 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
2156                 napi_disable(&vi->rq[i].napi);
2157                 virtnet_napi_tx_disable(&vi->sq[i].napi);
2158         }
2159
2160         return 0;
2161 }
2162
2163 static void virtnet_set_rx_mode(struct net_device *dev)
2164 {
2165         struct virtnet_info *vi = netdev_priv(dev);
2166         struct scatterlist sg[2];
2167         struct virtio_net_ctrl_mac *mac_data;
2168         struct netdev_hw_addr *ha;
2169         int uc_count;
2170         int mc_count;
2171         void *buf;
2172         int i;
2173
2174         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
2175         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
2176                 return;
2177
2178         vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
2179         vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
2180
2181         sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
2182
2183         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2184                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
2185                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
2186                          vi->ctrl->promisc ? "en" : "dis");
2187
2188         sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
2189
2190         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2191                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2192                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2193                          vi->ctrl->allmulti ? "en" : "dis");
2194
2195         uc_count = netdev_uc_count(dev);
2196         mc_count = netdev_mc_count(dev);
2197         /* MAC filter - use one buffer for both lists */
2198         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2199                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2200         mac_data = buf;
2201         if (!buf)
2202                 return;
2203
2204         sg_init_table(sg, 2);
2205
2206         /* Store the unicast list and count in the front of the buffer */
2207         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2208         i = 0;
2209         netdev_for_each_uc_addr(ha, dev)
2210                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2211
2212         sg_set_buf(&sg[0], mac_data,
2213                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2214
2215         /* multicast list and count fill the end */
2216         mac_data = (void *)&mac_data->macs[uc_count][0];
2217
2218         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2219         i = 0;
2220         netdev_for_each_mc_addr(ha, dev)
2221                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2222
2223         sg_set_buf(&sg[1], mac_data,
2224                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2225
2226         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2227                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2228                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2229
2230         kfree(buf);
2231 }
2232
2233 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2234                                    __be16 proto, u16 vid)
2235 {
2236         struct virtnet_info *vi = netdev_priv(dev);
2237         struct scatterlist sg;
2238
2239         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2240         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2241
2242         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2243                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2244                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2245         return 0;
2246 }
2247
2248 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2249                                     __be16 proto, u16 vid)
2250 {
2251         struct virtnet_info *vi = netdev_priv(dev);
2252         struct scatterlist sg;
2253
2254         vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2255         sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2256
2257         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2258                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2259                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2260         return 0;
2261 }
2262
2263 static void virtnet_clean_affinity(struct virtnet_info *vi)
2264 {
2265         int i;
2266
2267         if (vi->affinity_hint_set) {
2268                 for (i = 0; i < vi->max_queue_pairs; i++) {
2269                         virtqueue_set_affinity(vi->rq[i].vq, NULL);
2270                         virtqueue_set_affinity(vi->sq[i].vq, NULL);
2271                 }
2272
2273                 vi->affinity_hint_set = false;
2274         }
2275 }
2276
2277 static void virtnet_set_affinity(struct virtnet_info *vi)
2278 {
2279         cpumask_var_t mask;
2280         int stragglers;
2281         int group_size;
2282         int i, j, cpu;
2283         int num_cpu;
2284         int stride;
2285
2286         if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2287                 virtnet_clean_affinity(vi);
2288                 return;
2289         }
2290
2291         num_cpu = num_online_cpus();
2292         stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2293         stragglers = num_cpu >= vi->curr_queue_pairs ?
2294                         num_cpu % vi->curr_queue_pairs :
2295                         0;
2296         cpu = cpumask_first(cpu_online_mask);
2297
2298         for (i = 0; i < vi->curr_queue_pairs; i++) {
2299                 group_size = stride + (i < stragglers ? 1 : 0);
2300
2301                 for (j = 0; j < group_size; j++) {
2302                         cpumask_set_cpu(cpu, mask);
2303                         cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2304                                                 nr_cpu_ids, false);
2305                 }
2306                 virtqueue_set_affinity(vi->rq[i].vq, mask);
2307                 virtqueue_set_affinity(vi->sq[i].vq, mask);
2308                 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2309                 cpumask_clear(mask);
2310         }
2311
2312         vi->affinity_hint_set = true;
2313         free_cpumask_var(mask);
2314 }
2315
2316 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2317 {
2318         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2319                                                    node);
2320         virtnet_set_affinity(vi);
2321         return 0;
2322 }
2323
2324 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2325 {
2326         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2327                                                    node_dead);
2328         virtnet_set_affinity(vi);
2329         return 0;
2330 }
2331
2332 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2333 {
2334         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2335                                                    node);
2336
2337         virtnet_clean_affinity(vi);
2338         return 0;
2339 }
2340
2341 static enum cpuhp_state virtionet_online;
2342
2343 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2344 {
2345         int ret;
2346
2347         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2348         if (ret)
2349                 return ret;
2350         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2351                                                &vi->node_dead);
2352         if (!ret)
2353                 return ret;
2354         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2355         return ret;
2356 }
2357
2358 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2359 {
2360         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2361         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2362                                             &vi->node_dead);
2363 }
2364
2365 static void virtnet_get_ringparam(struct net_device *dev,
2366                                   struct ethtool_ringparam *ring,
2367                                   struct kernel_ethtool_ringparam *kernel_ring,
2368                                   struct netlink_ext_ack *extack)
2369 {
2370         struct virtnet_info *vi = netdev_priv(dev);
2371
2372         ring->rx_max_pending = vi->rq[0].vq->num_max;
2373         ring->tx_max_pending = vi->sq[0].vq->num_max;
2374         ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2375         ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2376 }
2377
2378 static int virtnet_set_ringparam(struct net_device *dev,
2379                                  struct ethtool_ringparam *ring,
2380                                  struct kernel_ethtool_ringparam *kernel_ring,
2381                                  struct netlink_ext_ack *extack)
2382 {
2383         struct virtnet_info *vi = netdev_priv(dev);
2384         u32 rx_pending, tx_pending;
2385         struct receive_queue *rq;
2386         struct send_queue *sq;
2387         int i, err;
2388
2389         if (ring->rx_mini_pending || ring->rx_jumbo_pending)
2390                 return -EINVAL;
2391
2392         rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2393         tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2394
2395         if (ring->rx_pending == rx_pending &&
2396             ring->tx_pending == tx_pending)
2397                 return 0;
2398
2399         if (ring->rx_pending > vi->rq[0].vq->num_max)
2400                 return -EINVAL;
2401
2402         if (ring->tx_pending > vi->sq[0].vq->num_max)
2403                 return -EINVAL;
2404
2405         for (i = 0; i < vi->max_queue_pairs; i++) {
2406                 rq = vi->rq + i;
2407                 sq = vi->sq + i;
2408
2409                 if (ring->tx_pending != tx_pending) {
2410                         err = virtnet_tx_resize(vi, sq, ring->tx_pending);
2411                         if (err)
2412                                 return err;
2413                 }
2414
2415                 if (ring->rx_pending != rx_pending) {
2416                         err = virtnet_rx_resize(vi, rq, ring->rx_pending);
2417                         if (err)
2418                                 return err;
2419                 }
2420         }
2421
2422         return 0;
2423 }
2424
2425 static bool virtnet_commit_rss_command(struct virtnet_info *vi)
2426 {
2427         struct net_device *dev = vi->dev;
2428         struct scatterlist sgs[4];
2429         unsigned int sg_buf_size;
2430
2431         /* prepare sgs */
2432         sg_init_table(sgs, 4);
2433
2434         sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table);
2435         sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size);
2436
2437         sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1);
2438         sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size);
2439
2440         sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key)
2441                         - offsetof(struct virtio_net_ctrl_rss, max_tx_vq);
2442         sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size);
2443
2444         sg_buf_size = vi->rss_key_size;
2445         sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size);
2446
2447         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2448                                   vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG
2449                                   : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) {
2450                 dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n");
2451                 return false;
2452         }
2453         return true;
2454 }
2455
2456 static void virtnet_init_default_rss(struct virtnet_info *vi)
2457 {
2458         u32 indir_val = 0;
2459         int i = 0;
2460
2461         vi->ctrl->rss.hash_types = vi->rss_hash_types_supported;
2462         vi->rss_hash_types_saved = vi->rss_hash_types_supported;
2463         vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size
2464                                                 ? vi->rss_indir_table_size - 1 : 0;
2465         vi->ctrl->rss.unclassified_queue = 0;
2466
2467         for (; i < vi->rss_indir_table_size; ++i) {
2468                 indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs);
2469                 vi->ctrl->rss.indirection_table[i] = indir_val;
2470         }
2471
2472         vi->ctrl->rss.max_tx_vq = vi->curr_queue_pairs;
2473         vi->ctrl->rss.hash_key_length = vi->rss_key_size;
2474
2475         netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size);
2476 }
2477
2478 static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info)
2479 {
2480         info->data = 0;
2481         switch (info->flow_type) {
2482         case TCP_V4_FLOW:
2483                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) {
2484                         info->data = RXH_IP_SRC | RXH_IP_DST |
2485                                                  RXH_L4_B_0_1 | RXH_L4_B_2_3;
2486                 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
2487                         info->data = RXH_IP_SRC | RXH_IP_DST;
2488                 }
2489                 break;
2490         case TCP_V6_FLOW:
2491                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) {
2492                         info->data = RXH_IP_SRC | RXH_IP_DST |
2493                                                  RXH_L4_B_0_1 | RXH_L4_B_2_3;
2494                 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
2495                         info->data = RXH_IP_SRC | RXH_IP_DST;
2496                 }
2497                 break;
2498         case UDP_V4_FLOW:
2499                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) {
2500                         info->data = RXH_IP_SRC | RXH_IP_DST |
2501                                                  RXH_L4_B_0_1 | RXH_L4_B_2_3;
2502                 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
2503                         info->data = RXH_IP_SRC | RXH_IP_DST;
2504                 }
2505                 break;
2506         case UDP_V6_FLOW:
2507                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) {
2508                         info->data = RXH_IP_SRC | RXH_IP_DST |
2509                                                  RXH_L4_B_0_1 | RXH_L4_B_2_3;
2510                 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
2511                         info->data = RXH_IP_SRC | RXH_IP_DST;
2512                 }
2513                 break;
2514         case IPV4_FLOW:
2515                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4)
2516                         info->data = RXH_IP_SRC | RXH_IP_DST;
2517
2518                 break;
2519         case IPV6_FLOW:
2520                 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6)
2521                         info->data = RXH_IP_SRC | RXH_IP_DST;
2522
2523                 break;
2524         default:
2525                 info->data = 0;
2526                 break;
2527         }
2528 }
2529
2530 static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info)
2531 {
2532         u32 new_hashtypes = vi->rss_hash_types_saved;
2533         bool is_disable = info->data & RXH_DISCARD;
2534         bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3);
2535
2536         /* supports only 'sd', 'sdfn' and 'r' */
2537         if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable))
2538                 return false;
2539
2540         switch (info->flow_type) {
2541         case TCP_V4_FLOW:
2542                 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4);
2543                 if (!is_disable)
2544                         new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
2545                                 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0);
2546                 break;
2547         case UDP_V4_FLOW:
2548                 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4);
2549                 if (!is_disable)
2550                         new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
2551                                 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0);
2552                 break;
2553         case IPV4_FLOW:
2554                 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4;
2555                 if (!is_disable)
2556                         new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4;
2557                 break;
2558         case TCP_V6_FLOW:
2559                 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6);
2560                 if (!is_disable)
2561                         new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
2562                                 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0);
2563                 break;
2564         case UDP_V6_FLOW:
2565                 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6);
2566                 if (!is_disable)
2567                         new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
2568                                 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0);
2569                 break;
2570         case IPV6_FLOW:
2571                 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6;
2572                 if (!is_disable)
2573                         new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6;
2574                 break;
2575         default:
2576                 /* unsupported flow */
2577                 return false;
2578         }
2579
2580         /* if unsupported hashtype was set */
2581         if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported))
2582                 return false;
2583
2584         if (new_hashtypes != vi->rss_hash_types_saved) {
2585                 vi->rss_hash_types_saved = new_hashtypes;
2586                 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
2587                 if (vi->dev->features & NETIF_F_RXHASH)
2588                         return virtnet_commit_rss_command(vi);
2589         }
2590
2591         return true;
2592 }
2593
2594 static void virtnet_get_drvinfo(struct net_device *dev,
2595                                 struct ethtool_drvinfo *info)
2596 {
2597         struct virtnet_info *vi = netdev_priv(dev);
2598         struct virtio_device *vdev = vi->vdev;
2599
2600         strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2601         strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2602         strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2603
2604 }
2605
2606 /* TODO: Eliminate OOO packets during switching */
2607 static int virtnet_set_channels(struct net_device *dev,
2608                                 struct ethtool_channels *channels)
2609 {
2610         struct virtnet_info *vi = netdev_priv(dev);
2611         u16 queue_pairs = channels->combined_count;
2612         int err;
2613
2614         /* We don't support separate rx/tx channels.
2615          * We don't allow setting 'other' channels.
2616          */
2617         if (channels->rx_count || channels->tx_count || channels->other_count)
2618                 return -EINVAL;
2619
2620         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2621                 return -EINVAL;
2622
2623         /* For now we don't support modifying channels while XDP is loaded
2624          * also when XDP is loaded all RX queues have XDP programs so we only
2625          * need to check a single RX queue.
2626          */
2627         if (vi->rq[0].xdp_prog)
2628                 return -EINVAL;
2629
2630         cpus_read_lock();
2631         err = _virtnet_set_queues(vi, queue_pairs);
2632         if (err) {
2633                 cpus_read_unlock();
2634                 goto err;
2635         }
2636         virtnet_set_affinity(vi);
2637         cpus_read_unlock();
2638
2639         netif_set_real_num_tx_queues(dev, queue_pairs);
2640         netif_set_real_num_rx_queues(dev, queue_pairs);
2641  err:
2642         return err;
2643 }
2644
2645 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2646 {
2647         struct virtnet_info *vi = netdev_priv(dev);
2648         unsigned int i, j;
2649         u8 *p = data;
2650
2651         switch (stringset) {
2652         case ETH_SS_STATS:
2653                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2654                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2655                                 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2656                                                 virtnet_rq_stats_desc[j].desc);
2657                 }
2658
2659                 for (i = 0; i < vi->curr_queue_pairs; i++) {
2660                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2661                                 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2662                                                 virtnet_sq_stats_desc[j].desc);
2663                 }
2664                 break;
2665         }
2666 }
2667
2668 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2669 {
2670         struct virtnet_info *vi = netdev_priv(dev);
2671
2672         switch (sset) {
2673         case ETH_SS_STATS:
2674                 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2675                                                VIRTNET_SQ_STATS_LEN);
2676         default:
2677                 return -EOPNOTSUPP;
2678         }
2679 }
2680
2681 static void virtnet_get_ethtool_stats(struct net_device *dev,
2682                                       struct ethtool_stats *stats, u64 *data)
2683 {
2684         struct virtnet_info *vi = netdev_priv(dev);
2685         unsigned int idx = 0, start, i, j;
2686         const u8 *stats_base;
2687         size_t offset;
2688
2689         for (i = 0; i < vi->curr_queue_pairs; i++) {
2690                 struct receive_queue *rq = &vi->rq[i];
2691
2692                 stats_base = (u8 *)&rq->stats;
2693                 do {
2694                         start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2695                         for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2696                                 offset = virtnet_rq_stats_desc[j].offset;
2697                                 data[idx + j] = *(u64 *)(stats_base + offset);
2698                         }
2699                 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2700                 idx += VIRTNET_RQ_STATS_LEN;
2701         }
2702
2703         for (i = 0; i < vi->curr_queue_pairs; i++) {
2704                 struct send_queue *sq = &vi->sq[i];
2705
2706                 stats_base = (u8 *)&sq->stats;
2707                 do {
2708                         start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2709                         for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2710                                 offset = virtnet_sq_stats_desc[j].offset;
2711                                 data[idx + j] = *(u64 *)(stats_base + offset);
2712                         }
2713                 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2714                 idx += VIRTNET_SQ_STATS_LEN;
2715         }
2716 }
2717
2718 static void virtnet_get_channels(struct net_device *dev,
2719                                  struct ethtool_channels *channels)
2720 {
2721         struct virtnet_info *vi = netdev_priv(dev);
2722
2723         channels->combined_count = vi->curr_queue_pairs;
2724         channels->max_combined = vi->max_queue_pairs;
2725         channels->max_other = 0;
2726         channels->rx_count = 0;
2727         channels->tx_count = 0;
2728         channels->other_count = 0;
2729 }
2730
2731 static int virtnet_set_link_ksettings(struct net_device *dev,
2732                                       const struct ethtool_link_ksettings *cmd)
2733 {
2734         struct virtnet_info *vi = netdev_priv(dev);
2735
2736         return ethtool_virtdev_set_link_ksettings(dev, cmd,
2737                                                   &vi->speed, &vi->duplex);
2738 }
2739
2740 static int virtnet_get_link_ksettings(struct net_device *dev,
2741                                       struct ethtool_link_ksettings *cmd)
2742 {
2743         struct virtnet_info *vi = netdev_priv(dev);
2744
2745         cmd->base.speed = vi->speed;
2746         cmd->base.duplex = vi->duplex;
2747         cmd->base.port = PORT_OTHER;
2748
2749         return 0;
2750 }
2751
2752 static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
2753                                        struct ethtool_coalesce *ec)
2754 {
2755         struct scatterlist sgs_tx, sgs_rx;
2756         struct virtio_net_ctrl_coal_tx coal_tx;
2757         struct virtio_net_ctrl_coal_rx coal_rx;
2758
2759         coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
2760         coal_tx.tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames);
2761         sg_init_one(&sgs_tx, &coal_tx, sizeof(coal_tx));
2762
2763         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
2764                                   VIRTIO_NET_CTRL_NOTF_COAL_TX_SET,
2765                                   &sgs_tx))
2766                 return -EINVAL;
2767
2768         /* Save parameters */
2769         vi->tx_usecs = ec->tx_coalesce_usecs;
2770         vi->tx_max_packets = ec->tx_max_coalesced_frames;
2771
2772         coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
2773         coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
2774         sg_init_one(&sgs_rx, &coal_rx, sizeof(coal_rx));
2775
2776         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
2777                                   VIRTIO_NET_CTRL_NOTF_COAL_RX_SET,
2778                                   &sgs_rx))
2779                 return -EINVAL;
2780
2781         /* Save parameters */
2782         vi->rx_usecs = ec->rx_coalesce_usecs;
2783         vi->rx_max_packets = ec->rx_max_coalesced_frames;
2784
2785         return 0;
2786 }
2787
2788 static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
2789 {
2790         /* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL
2791          * feature is negotiated.
2792          */
2793         if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs)
2794                 return -EOPNOTSUPP;
2795
2796         if (ec->tx_max_coalesced_frames > 1 ||
2797             ec->rx_max_coalesced_frames != 1)
2798                 return -EINVAL;
2799
2800         return 0;
2801 }
2802
2803 static int virtnet_set_coalesce(struct net_device *dev,
2804                                 struct ethtool_coalesce *ec,
2805                                 struct kernel_ethtool_coalesce *kernel_coal,
2806                                 struct netlink_ext_ack *extack)
2807 {
2808         struct virtnet_info *vi = netdev_priv(dev);
2809         int ret, i, napi_weight;
2810         bool update_napi = false;
2811
2812         /* Can't change NAPI weight if the link is up */
2813         napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2814         if (napi_weight ^ vi->sq[0].napi.weight) {
2815                 if (dev->flags & IFF_UP)
2816                         return -EBUSY;
2817                 else
2818                         update_napi = true;
2819         }
2820
2821         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL))
2822                 ret = virtnet_send_notf_coal_cmds(vi, ec);
2823         else
2824                 ret = virtnet_coal_params_supported(ec);
2825
2826         if (ret)
2827                 return ret;
2828
2829         if (update_napi) {
2830                 for (i = 0; i < vi->max_queue_pairs; i++)
2831                         vi->sq[i].napi.weight = napi_weight;
2832         }
2833
2834         return ret;
2835 }
2836
2837 static int virtnet_get_coalesce(struct net_device *dev,
2838                                 struct ethtool_coalesce *ec,
2839                                 struct kernel_ethtool_coalesce *kernel_coal,
2840                                 struct netlink_ext_ack *extack)
2841 {
2842         struct virtnet_info *vi = netdev_priv(dev);
2843
2844         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
2845                 ec->rx_coalesce_usecs = vi->rx_usecs;
2846                 ec->tx_coalesce_usecs = vi->tx_usecs;
2847                 ec->tx_max_coalesced_frames = vi->tx_max_packets;
2848                 ec->rx_max_coalesced_frames = vi->rx_max_packets;
2849         } else {
2850                 ec->rx_max_coalesced_frames = 1;
2851
2852                 if (vi->sq[0].napi.weight)
2853                         ec->tx_max_coalesced_frames = 1;
2854         }
2855
2856         return 0;
2857 }
2858
2859 static void virtnet_init_settings(struct net_device *dev)
2860 {
2861         struct virtnet_info *vi = netdev_priv(dev);
2862
2863         vi->speed = SPEED_UNKNOWN;
2864         vi->duplex = DUPLEX_UNKNOWN;
2865 }
2866
2867 static void virtnet_update_settings(struct virtnet_info *vi)
2868 {
2869         u32 speed;
2870         u8 duplex;
2871
2872         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2873                 return;
2874
2875         virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2876
2877         if (ethtool_validate_speed(speed))
2878                 vi->speed = speed;
2879
2880         virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2881
2882         if (ethtool_validate_duplex(duplex))
2883                 vi->duplex = duplex;
2884 }
2885
2886 static u32 virtnet_get_rxfh_key_size(struct net_device *dev)
2887 {
2888         return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size;
2889 }
2890
2891 static u32 virtnet_get_rxfh_indir_size(struct net_device *dev)
2892 {
2893         return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size;
2894 }
2895
2896 static int virtnet_get_rxfh(struct net_device *dev, u32 *indir, u8 *key, u8 *hfunc)
2897 {
2898         struct virtnet_info *vi = netdev_priv(dev);
2899         int i;
2900
2901         if (indir) {
2902                 for (i = 0; i < vi->rss_indir_table_size; ++i)
2903                         indir[i] = vi->ctrl->rss.indirection_table[i];
2904         }
2905
2906         if (key)
2907                 memcpy(key, vi->ctrl->rss.key, vi->rss_key_size);
2908
2909         if (hfunc)
2910                 *hfunc = ETH_RSS_HASH_TOP;
2911
2912         return 0;
2913 }
2914
2915 static int virtnet_set_rxfh(struct net_device *dev, const u32 *indir, const u8 *key, const u8 hfunc)
2916 {
2917         struct virtnet_info *vi = netdev_priv(dev);
2918         int i;
2919
2920         if (hfunc != ETH_RSS_HASH_NO_CHANGE && hfunc != ETH_RSS_HASH_TOP)
2921                 return -EOPNOTSUPP;
2922
2923         if (indir) {
2924                 for (i = 0; i < vi->rss_indir_table_size; ++i)
2925                         vi->ctrl->rss.indirection_table[i] = indir[i];
2926         }
2927         if (key)
2928                 memcpy(vi->ctrl->rss.key, key, vi->rss_key_size);
2929
2930         virtnet_commit_rss_command(vi);
2931
2932         return 0;
2933 }
2934
2935 static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs)
2936 {
2937         struct virtnet_info *vi = netdev_priv(dev);
2938         int rc = 0;
2939
2940         switch (info->cmd) {
2941         case ETHTOOL_GRXRINGS:
2942                 info->data = vi->curr_queue_pairs;
2943                 break;
2944         case ETHTOOL_GRXFH:
2945                 virtnet_get_hashflow(vi, info);
2946                 break;
2947         default:
2948                 rc = -EOPNOTSUPP;
2949         }
2950
2951         return rc;
2952 }
2953
2954 static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info)
2955 {
2956         struct virtnet_info *vi = netdev_priv(dev);
2957         int rc = 0;
2958
2959         switch (info->cmd) {
2960         case ETHTOOL_SRXFH:
2961                 if (!virtnet_set_hashflow(vi, info))
2962                         rc = -EINVAL;
2963
2964                 break;
2965         default:
2966                 rc = -EOPNOTSUPP;
2967         }
2968
2969         return rc;
2970 }
2971
2972 static const struct ethtool_ops virtnet_ethtool_ops = {
2973         .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
2974                 ETHTOOL_COALESCE_USECS,
2975         .get_drvinfo = virtnet_get_drvinfo,
2976         .get_link = ethtool_op_get_link,
2977         .get_ringparam = virtnet_get_ringparam,
2978         .set_ringparam = virtnet_set_ringparam,
2979         .get_strings = virtnet_get_strings,
2980         .get_sset_count = virtnet_get_sset_count,
2981         .get_ethtool_stats = virtnet_get_ethtool_stats,
2982         .set_channels = virtnet_set_channels,
2983         .get_channels = virtnet_get_channels,
2984         .get_ts_info = ethtool_op_get_ts_info,
2985         .get_link_ksettings = virtnet_get_link_ksettings,
2986         .set_link_ksettings = virtnet_set_link_ksettings,
2987         .set_coalesce = virtnet_set_coalesce,
2988         .get_coalesce = virtnet_get_coalesce,
2989         .get_rxfh_key_size = virtnet_get_rxfh_key_size,
2990         .get_rxfh_indir_size = virtnet_get_rxfh_indir_size,
2991         .get_rxfh = virtnet_get_rxfh,
2992         .set_rxfh = virtnet_set_rxfh,
2993         .get_rxnfc = virtnet_get_rxnfc,
2994         .set_rxnfc = virtnet_set_rxnfc,
2995 };
2996
2997 static void virtnet_freeze_down(struct virtio_device *vdev)
2998 {
2999         struct virtnet_info *vi = vdev->priv;
3000
3001         /* Make sure no work handler is accessing the device */
3002         flush_work(&vi->config_work);
3003
3004         netif_tx_lock_bh(vi->dev);
3005         netif_device_detach(vi->dev);
3006         netif_tx_unlock_bh(vi->dev);
3007         if (netif_running(vi->dev))
3008                 virtnet_close(vi->dev);
3009 }
3010
3011 static int init_vqs(struct virtnet_info *vi);
3012
3013 static int virtnet_restore_up(struct virtio_device *vdev)
3014 {
3015         struct virtnet_info *vi = vdev->priv;
3016         int err;
3017
3018         err = init_vqs(vi);
3019         if (err)
3020                 return err;
3021
3022         virtio_device_ready(vdev);
3023
3024         enable_delayed_refill(vi);
3025
3026         if (netif_running(vi->dev)) {
3027                 err = virtnet_open(vi->dev);
3028                 if (err)
3029                         return err;
3030         }
3031
3032         netif_tx_lock_bh(vi->dev);
3033         netif_device_attach(vi->dev);
3034         netif_tx_unlock_bh(vi->dev);
3035         return err;
3036 }
3037
3038 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
3039 {
3040         struct scatterlist sg;
3041         vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
3042
3043         sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
3044
3045         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
3046                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
3047                 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
3048                 return -EINVAL;
3049         }
3050
3051         return 0;
3052 }
3053
3054 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
3055 {
3056         u64 offloads = 0;
3057
3058         if (!vi->guest_offloads)
3059                 return 0;
3060
3061         return virtnet_set_guest_offloads(vi, offloads);
3062 }
3063
3064 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
3065 {
3066         u64 offloads = vi->guest_offloads;
3067
3068         if (!vi->guest_offloads)
3069                 return 0;
3070
3071         return virtnet_set_guest_offloads(vi, offloads);
3072 }
3073
3074 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
3075                            struct netlink_ext_ack *extack)
3076 {
3077         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
3078         struct virtnet_info *vi = netdev_priv(dev);
3079         struct bpf_prog *old_prog;
3080         u16 xdp_qp = 0, curr_qp;
3081         int i, err;
3082
3083         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
3084             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3085                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3086                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3087                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3088                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
3089                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
3090                 return -EOPNOTSUPP;
3091         }
3092
3093         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
3094                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
3095                 return -EINVAL;
3096         }
3097
3098         if (dev->mtu > max_sz) {
3099                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
3100                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
3101                 return -EINVAL;
3102         }
3103
3104         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
3105         if (prog)
3106                 xdp_qp = nr_cpu_ids;
3107
3108         /* XDP requires extra queues for XDP_TX */
3109         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
3110                 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",
3111                                  curr_qp + xdp_qp, vi->max_queue_pairs);
3112                 xdp_qp = 0;
3113         }
3114
3115         old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
3116         if (!prog && !old_prog)
3117                 return 0;
3118
3119         if (prog)
3120                 bpf_prog_add(prog, vi->max_queue_pairs - 1);
3121
3122         /* Make sure NAPI is not using any XDP TX queues for RX. */
3123         if (netif_running(dev)) {
3124                 for (i = 0; i < vi->max_queue_pairs; i++) {
3125                         napi_disable(&vi->rq[i].napi);
3126                         virtnet_napi_tx_disable(&vi->sq[i].napi);
3127                 }
3128         }
3129
3130         if (!prog) {
3131                 for (i = 0; i < vi->max_queue_pairs; i++) {
3132                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3133                         if (i == 0)
3134                                 virtnet_restore_guest_offloads(vi);
3135                 }
3136                 synchronize_net();
3137         }
3138
3139         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
3140         if (err)
3141                 goto err;
3142         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
3143         vi->xdp_queue_pairs = xdp_qp;
3144
3145         if (prog) {
3146                 vi->xdp_enabled = true;
3147                 for (i = 0; i < vi->max_queue_pairs; i++) {
3148                         rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3149                         if (i == 0 && !old_prog)
3150                                 virtnet_clear_guest_offloads(vi);
3151                 }
3152         } else {
3153                 vi->xdp_enabled = false;
3154         }
3155
3156         for (i = 0; i < vi->max_queue_pairs; i++) {
3157                 if (old_prog)
3158                         bpf_prog_put(old_prog);
3159                 if (netif_running(dev)) {
3160                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3161                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3162                                                &vi->sq[i].napi);
3163                 }
3164         }
3165
3166         return 0;
3167
3168 err:
3169         if (!prog) {
3170                 virtnet_clear_guest_offloads(vi);
3171                 for (i = 0; i < vi->max_queue_pairs; i++)
3172                         rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
3173         }
3174
3175         if (netif_running(dev)) {
3176                 for (i = 0; i < vi->max_queue_pairs; i++) {
3177                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3178                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3179                                                &vi->sq[i].napi);
3180                 }
3181         }
3182         if (prog)
3183                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
3184         return err;
3185 }
3186
3187 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
3188 {
3189         switch (xdp->command) {
3190         case XDP_SETUP_PROG:
3191                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
3192         default:
3193                 return -EINVAL;
3194         }
3195 }
3196
3197 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
3198                                       size_t len)
3199 {
3200         struct virtnet_info *vi = netdev_priv(dev);
3201         int ret;
3202
3203         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
3204                 return -EOPNOTSUPP;
3205
3206         ret = snprintf(buf, len, "sby");
3207         if (ret >= len)
3208                 return -EOPNOTSUPP;
3209
3210         return 0;
3211 }
3212
3213 static int virtnet_set_features(struct net_device *dev,
3214                                 netdev_features_t features)
3215 {
3216         struct virtnet_info *vi = netdev_priv(dev);
3217         u64 offloads;
3218         int err;
3219
3220         if ((dev->features ^ features) & NETIF_F_GRO_HW) {
3221                 if (vi->xdp_enabled)
3222                         return -EBUSY;
3223
3224                 if (features & NETIF_F_GRO_HW)
3225                         offloads = vi->guest_offloads_capable;
3226                 else
3227                         offloads = vi->guest_offloads_capable &
3228                                    ~GUEST_OFFLOAD_GRO_HW_MASK;
3229
3230                 err = virtnet_set_guest_offloads(vi, offloads);
3231                 if (err)
3232                         return err;
3233                 vi->guest_offloads = offloads;
3234         }
3235
3236         if ((dev->features ^ features) & NETIF_F_RXHASH) {
3237                 if (features & NETIF_F_RXHASH)
3238                         vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
3239                 else
3240                         vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE;
3241
3242                 if (!virtnet_commit_rss_command(vi))
3243                         return -EINVAL;
3244         }
3245
3246         return 0;
3247 }
3248
3249 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
3250 {
3251         struct virtnet_info *priv = netdev_priv(dev);
3252         struct send_queue *sq = &priv->sq[txqueue];
3253         struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
3254
3255         u64_stats_update_begin(&sq->stats.syncp);
3256         sq->stats.tx_timeouts++;
3257         u64_stats_update_end(&sq->stats.syncp);
3258
3259         netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
3260                    txqueue, sq->name, sq->vq->index, sq->vq->name,
3261                    jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
3262 }
3263
3264 static const struct net_device_ops virtnet_netdev = {
3265         .ndo_open            = virtnet_open,
3266         .ndo_stop            = virtnet_close,
3267         .ndo_start_xmit      = start_xmit,
3268         .ndo_validate_addr   = eth_validate_addr,
3269         .ndo_set_mac_address = virtnet_set_mac_address,
3270         .ndo_set_rx_mode     = virtnet_set_rx_mode,
3271         .ndo_get_stats64     = virtnet_stats,
3272         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
3273         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
3274         .ndo_bpf                = virtnet_xdp,
3275         .ndo_xdp_xmit           = virtnet_xdp_xmit,
3276         .ndo_features_check     = passthru_features_check,
3277         .ndo_get_phys_port_name = virtnet_get_phys_port_name,
3278         .ndo_set_features       = virtnet_set_features,
3279         .ndo_tx_timeout         = virtnet_tx_timeout,
3280 };
3281
3282 static void virtnet_config_changed_work(struct work_struct *work)
3283 {
3284         struct virtnet_info *vi =
3285                 container_of(work, struct virtnet_info, config_work);
3286         u16 v;
3287
3288         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
3289                                  struct virtio_net_config, status, &v) < 0)
3290                 return;
3291
3292         if (v & VIRTIO_NET_S_ANNOUNCE) {
3293                 netdev_notify_peers(vi->dev);
3294                 virtnet_ack_link_announce(vi);
3295         }
3296
3297         /* Ignore unknown (future) status bits */
3298         v &= VIRTIO_NET_S_LINK_UP;
3299
3300         if (vi->status == v)
3301                 return;
3302
3303         vi->status = v;
3304
3305         if (vi->status & VIRTIO_NET_S_LINK_UP) {
3306                 virtnet_update_settings(vi);
3307                 netif_carrier_on(vi->dev);
3308                 netif_tx_wake_all_queues(vi->dev);
3309         } else {
3310                 netif_carrier_off(vi->dev);
3311                 netif_tx_stop_all_queues(vi->dev);
3312         }
3313 }
3314
3315 static void virtnet_config_changed(struct virtio_device *vdev)
3316 {
3317         struct virtnet_info *vi = vdev->priv;
3318
3319         schedule_work(&vi->config_work);
3320 }
3321
3322 static void virtnet_free_queues(struct virtnet_info *vi)
3323 {
3324         int i;
3325
3326         for (i = 0; i < vi->max_queue_pairs; i++) {
3327                 __netif_napi_del(&vi->rq[i].napi);
3328                 __netif_napi_del(&vi->sq[i].napi);
3329         }
3330
3331         /* We called __netif_napi_del(),
3332          * we need to respect an RCU grace period before freeing vi->rq
3333          */
3334         synchronize_net();
3335
3336         kfree(vi->rq);
3337         kfree(vi->sq);
3338         kfree(vi->ctrl);
3339 }
3340
3341 static void _free_receive_bufs(struct virtnet_info *vi)
3342 {
3343         struct bpf_prog *old_prog;
3344         int i;
3345
3346         for (i = 0; i < vi->max_queue_pairs; i++) {
3347                 while (vi->rq[i].pages)
3348                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
3349
3350                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
3351                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
3352                 if (old_prog)
3353                         bpf_prog_put(old_prog);
3354         }
3355 }
3356
3357 static void free_receive_bufs(struct virtnet_info *vi)
3358 {
3359         rtnl_lock();
3360         _free_receive_bufs(vi);
3361         rtnl_unlock();
3362 }
3363
3364 static void free_receive_page_frags(struct virtnet_info *vi)
3365 {
3366         int i;
3367         for (i = 0; i < vi->max_queue_pairs; i++)
3368                 if (vi->rq[i].alloc_frag.page)
3369                         put_page(vi->rq[i].alloc_frag.page);
3370 }
3371
3372 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
3373 {
3374         if (!is_xdp_frame(buf))
3375                 dev_kfree_skb(buf);
3376         else
3377                 xdp_return_frame(ptr_to_xdp(buf));
3378 }
3379
3380 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf)
3381 {
3382         struct virtnet_info *vi = vq->vdev->priv;
3383         int i = vq2rxq(vq);
3384
3385         if (vi->mergeable_rx_bufs)
3386                 put_page(virt_to_head_page(buf));
3387         else if (vi->big_packets)
3388                 give_pages(&vi->rq[i], buf);
3389         else
3390                 put_page(virt_to_head_page(buf));
3391 }
3392
3393 static void free_unused_bufs(struct virtnet_info *vi)
3394 {
3395         void *buf;
3396         int i;
3397
3398         for (i = 0; i < vi->max_queue_pairs; i++) {
3399                 struct virtqueue *vq = vi->sq[i].vq;
3400                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
3401                         virtnet_sq_free_unused_buf(vq, buf);
3402         }
3403
3404         for (i = 0; i < vi->max_queue_pairs; i++) {
3405                 struct virtqueue *vq = vi->rq[i].vq;
3406                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
3407                         virtnet_rq_free_unused_buf(vq, buf);
3408         }
3409 }
3410
3411 static void virtnet_del_vqs(struct virtnet_info *vi)
3412 {
3413         struct virtio_device *vdev = vi->vdev;
3414
3415         virtnet_clean_affinity(vi);
3416
3417         vdev->config->del_vqs(vdev);
3418
3419         virtnet_free_queues(vi);
3420 }
3421
3422 /* How large should a single buffer be so a queue full of these can fit at
3423  * least one full packet?
3424  * Logic below assumes the mergeable buffer header is used.
3425  */
3426 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
3427 {
3428         const unsigned int hdr_len = vi->hdr_len;
3429         unsigned int rq_size = virtqueue_get_vring_size(vq);
3430         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
3431         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
3432         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
3433
3434         return max(max(min_buf_len, hdr_len) - hdr_len,
3435                    (unsigned int)GOOD_PACKET_LEN);
3436 }
3437
3438 static int virtnet_find_vqs(struct virtnet_info *vi)
3439 {
3440         vq_callback_t **callbacks;
3441         struct virtqueue **vqs;
3442         int ret = -ENOMEM;
3443         int i, total_vqs;
3444         const char **names;
3445         bool *ctx;
3446
3447         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
3448          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
3449          * possible control vq.
3450          */
3451         total_vqs = vi->max_queue_pairs * 2 +
3452                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
3453
3454         /* Allocate space for find_vqs parameters */
3455         vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
3456         if (!vqs)
3457                 goto err_vq;
3458         callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
3459         if (!callbacks)
3460                 goto err_callback;
3461         names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
3462         if (!names)
3463                 goto err_names;
3464         if (!vi->big_packets || vi->mergeable_rx_bufs) {
3465                 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
3466                 if (!ctx)
3467                         goto err_ctx;
3468         } else {
3469                 ctx = NULL;
3470         }
3471
3472         /* Parameters for control virtqueue, if any */
3473         if (vi->has_cvq) {
3474                 callbacks[total_vqs - 1] = NULL;
3475                 names[total_vqs - 1] = "control";
3476         }
3477
3478         /* Allocate/initialize parameters for send/receive virtqueues */
3479         for (i = 0; i < vi->max_queue_pairs; i++) {
3480                 callbacks[rxq2vq(i)] = skb_recv_done;
3481                 callbacks[txq2vq(i)] = skb_xmit_done;
3482                 sprintf(vi->rq[i].name, "input.%d", i);
3483                 sprintf(vi->sq[i].name, "output.%d", i);
3484                 names[rxq2vq(i)] = vi->rq[i].name;
3485                 names[txq2vq(i)] = vi->sq[i].name;
3486                 if (ctx)
3487                         ctx[rxq2vq(i)] = true;
3488         }
3489
3490         ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
3491                                   names, ctx, NULL);
3492         if (ret)
3493                 goto err_find;
3494
3495         if (vi->has_cvq) {
3496                 vi->cvq = vqs[total_vqs - 1];
3497                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
3498                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
3499         }
3500
3501         for (i = 0; i < vi->max_queue_pairs; i++) {
3502                 vi->rq[i].vq = vqs[rxq2vq(i)];
3503                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
3504                 vi->sq[i].vq = vqs[txq2vq(i)];
3505         }
3506
3507         /* run here: ret == 0. */
3508
3509
3510 err_find:
3511         kfree(ctx);
3512 err_ctx:
3513         kfree(names);
3514 err_names:
3515         kfree(callbacks);
3516 err_callback:
3517         kfree(vqs);
3518 err_vq:
3519         return ret;
3520 }
3521
3522 static int virtnet_alloc_queues(struct virtnet_info *vi)
3523 {
3524         int i;
3525
3526         if (vi->has_cvq) {
3527                 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
3528                 if (!vi->ctrl)
3529                         goto err_ctrl;
3530         } else {
3531                 vi->ctrl = NULL;
3532         }
3533         vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
3534         if (!vi->sq)
3535                 goto err_sq;
3536         vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
3537         if (!vi->rq)
3538                 goto err_rq;
3539
3540         INIT_DELAYED_WORK(&vi->refill, refill_work);
3541         for (i = 0; i < vi->max_queue_pairs; i++) {
3542                 vi->rq[i].pages = NULL;
3543                 netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll,
3544                                       napi_weight);
3545                 netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi,
3546                                          virtnet_poll_tx,
3547                                          napi_tx ? napi_weight : 0);
3548
3549                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
3550                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
3551                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
3552
3553                 u64_stats_init(&vi->rq[i].stats.syncp);
3554                 u64_stats_init(&vi->sq[i].stats.syncp);
3555         }
3556
3557         return 0;
3558
3559 err_rq:
3560         kfree(vi->sq);
3561 err_sq:
3562         kfree(vi->ctrl);
3563 err_ctrl:
3564         return -ENOMEM;
3565 }
3566
3567 static int init_vqs(struct virtnet_info *vi)
3568 {
3569         int ret;
3570
3571         /* Allocate send & receive queues */
3572         ret = virtnet_alloc_queues(vi);
3573         if (ret)
3574                 goto err;
3575
3576         ret = virtnet_find_vqs(vi);
3577         if (ret)
3578                 goto err_free;
3579
3580         cpus_read_lock();
3581         virtnet_set_affinity(vi);
3582         cpus_read_unlock();
3583
3584         return 0;
3585
3586 err_free:
3587         virtnet_free_queues(vi);
3588 err:
3589         return ret;
3590 }
3591
3592 #ifdef CONFIG_SYSFS
3593 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
3594                 char *buf)
3595 {
3596         struct virtnet_info *vi = netdev_priv(queue->dev);
3597         unsigned int queue_index = get_netdev_rx_queue_index(queue);
3598         unsigned int headroom = virtnet_get_headroom(vi);
3599         unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
3600         struct ewma_pkt_len *avg;
3601
3602         BUG_ON(queue_index >= vi->max_queue_pairs);
3603         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
3604         return sprintf(buf, "%u\n",
3605                        get_mergeable_buf_len(&vi->rq[queue_index], avg,
3606                                        SKB_DATA_ALIGN(headroom + tailroom)));
3607 }
3608
3609 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
3610         __ATTR_RO(mergeable_rx_buffer_size);
3611
3612 static struct attribute *virtio_net_mrg_rx_attrs[] = {
3613         &mergeable_rx_buffer_size_attribute.attr,
3614         NULL
3615 };
3616
3617 static const struct attribute_group virtio_net_mrg_rx_group = {
3618         .name = "virtio_net",
3619         .attrs = virtio_net_mrg_rx_attrs
3620 };
3621 #endif
3622
3623 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3624                                     unsigned int fbit,
3625                                     const char *fname, const char *dname)
3626 {
3627         if (!virtio_has_feature(vdev, fbit))
3628                 return false;
3629
3630         dev_err(&vdev->dev, "device advertises feature %s but not %s",
3631                 fname, dname);
3632
3633         return true;
3634 }
3635
3636 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
3637         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3638
3639 static bool virtnet_validate_features(struct virtio_device *vdev)
3640 {
3641         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3642             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3643                              "VIRTIO_NET_F_CTRL_VQ") ||
3644              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3645                              "VIRTIO_NET_F_CTRL_VQ") ||
3646              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3647                              "VIRTIO_NET_F_CTRL_VQ") ||
3648              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3649              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3650                              "VIRTIO_NET_F_CTRL_VQ") ||
3651              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS,
3652                              "VIRTIO_NET_F_CTRL_VQ") ||
3653              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT,
3654                              "VIRTIO_NET_F_CTRL_VQ") ||
3655              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL,
3656                              "VIRTIO_NET_F_CTRL_VQ"))) {
3657                 return false;
3658         }
3659
3660         return true;
3661 }
3662
3663 #define MIN_MTU ETH_MIN_MTU
3664 #define MAX_MTU ETH_MAX_MTU
3665
3666 static int virtnet_validate(struct virtio_device *vdev)
3667 {
3668         if (!vdev->config->get) {
3669                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3670                         __func__);
3671                 return -EINVAL;
3672         }
3673
3674         if (!virtnet_validate_features(vdev))
3675                 return -EINVAL;
3676
3677         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3678                 int mtu = virtio_cread16(vdev,
3679                                          offsetof(struct virtio_net_config,
3680                                                   mtu));
3681                 if (mtu < MIN_MTU)
3682                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3683         }
3684
3685         return 0;
3686 }
3687
3688 static bool virtnet_check_guest_gso(const struct virtnet_info *vi)
3689 {
3690         return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3691                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3692                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3693                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO);
3694 }
3695
3696 static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu)
3697 {
3698         bool guest_gso = virtnet_check_guest_gso(vi);
3699
3700         /* If device can receive ANY guest GSO packets, regardless of mtu,
3701          * allocate packets of maximum size, otherwise limit it to only
3702          * mtu size worth only.
3703          */
3704         if (mtu > ETH_DATA_LEN || guest_gso) {
3705                 vi->big_packets = true;
3706                 vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE);
3707         }
3708 }
3709
3710 static int virtnet_probe(struct virtio_device *vdev)
3711 {
3712         int i, err = -ENOMEM;
3713         struct net_device *dev;
3714         struct virtnet_info *vi;
3715         u16 max_queue_pairs;
3716         int mtu = 0;
3717
3718         /* Find if host supports multiqueue/rss virtio_net device */
3719         max_queue_pairs = 1;
3720         if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
3721                 max_queue_pairs =
3722                      virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs));
3723
3724         /* We need at least 2 queue's */
3725         if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3726             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3727             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3728                 max_queue_pairs = 1;
3729
3730         /* Allocate ourselves a network device with room for our info */
3731         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3732         if (!dev)
3733                 return -ENOMEM;
3734
3735         /* Set up network device as normal. */
3736         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3737                            IFF_TX_SKB_NO_LINEAR;
3738         dev->netdev_ops = &virtnet_netdev;
3739         dev->features = NETIF_F_HIGHDMA;
3740
3741         dev->ethtool_ops = &virtnet_ethtool_ops;
3742         SET_NETDEV_DEV(dev, &vdev->dev);
3743
3744         /* Do we support "hardware" checksums? */
3745         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3746                 /* This opens up the world of extra features. */
3747                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3748                 if (csum)
3749                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3750
3751                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3752                         dev->hw_features |= NETIF_F_TSO
3753                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3754                 }
3755                 /* Individual feature bits: what can host handle? */
3756                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3757                         dev->hw_features |= NETIF_F_TSO;
3758                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3759                         dev->hw_features |= NETIF_F_TSO6;
3760                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3761                         dev->hw_features |= NETIF_F_TSO_ECN;
3762
3763                 dev->features |= NETIF_F_GSO_ROBUST;
3764
3765                 if (gso)
3766                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3767                 /* (!csum && gso) case will be fixed by register_netdev() */
3768         }
3769         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3770                 dev->features |= NETIF_F_RXCSUM;
3771         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3772             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3773                 dev->features |= NETIF_F_GRO_HW;
3774         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3775                 dev->hw_features |= NETIF_F_GRO_HW;
3776
3777         dev->vlan_features = dev->features;
3778
3779         /* MTU range: 68 - 65535 */
3780         dev->min_mtu = MIN_MTU;
3781         dev->max_mtu = MAX_MTU;
3782
3783         /* Configuration may specify what MAC to use.  Otherwise random. */
3784         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
3785                 u8 addr[ETH_ALEN];
3786
3787                 virtio_cread_bytes(vdev,
3788                                    offsetof(struct virtio_net_config, mac),
3789                                    addr, ETH_ALEN);
3790                 eth_hw_addr_set(dev, addr);
3791         } else {
3792                 eth_hw_addr_random(dev);
3793         }
3794
3795         /* Set up our device-specific information */
3796         vi = netdev_priv(dev);
3797         vi->dev = dev;
3798         vi->vdev = vdev;
3799         vdev->priv = vi;
3800
3801         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3802         spin_lock_init(&vi->refill_lock);
3803
3804         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3805                 vi->mergeable_rx_bufs = true;
3806
3807         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
3808                 vi->rx_usecs = 0;
3809                 vi->tx_usecs = 0;
3810                 vi->tx_max_packets = 0;
3811                 vi->rx_max_packets = 0;
3812         }
3813
3814         if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT))
3815                 vi->has_rss_hash_report = true;
3816
3817         if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
3818                 vi->has_rss = true;
3819
3820         if (vi->has_rss || vi->has_rss_hash_report) {
3821                 vi->rss_indir_table_size =
3822                         virtio_cread16(vdev, offsetof(struct virtio_net_config,
3823                                 rss_max_indirection_table_length));
3824                 vi->rss_key_size =
3825                         virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size));
3826
3827                 vi->rss_hash_types_supported =
3828                     virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types));
3829                 vi->rss_hash_types_supported &=
3830                                 ~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX |
3831                                   VIRTIO_NET_RSS_HASH_TYPE_TCP_EX |
3832                                   VIRTIO_NET_RSS_HASH_TYPE_UDP_EX);
3833
3834                 dev->hw_features |= NETIF_F_RXHASH;
3835         }
3836
3837         if (vi->has_rss_hash_report)
3838                 vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash);
3839         else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3840                  virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3841                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3842         else
3843                 vi->hdr_len = sizeof(struct virtio_net_hdr);
3844
3845         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3846             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3847                 vi->any_header_sg = true;
3848
3849         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3850                 vi->has_cvq = true;
3851
3852         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3853                 mtu = virtio_cread16(vdev,
3854                                      offsetof(struct virtio_net_config,
3855                                               mtu));
3856                 if (mtu < dev->min_mtu) {
3857                         /* Should never trigger: MTU was previously validated
3858                          * in virtnet_validate.
3859                          */
3860                         dev_err(&vdev->dev,
3861                                 "device MTU appears to have changed it is now %d < %d",
3862                                 mtu, dev->min_mtu);
3863                         err = -EINVAL;
3864                         goto free;
3865                 }
3866
3867                 dev->mtu = mtu;
3868                 dev->max_mtu = mtu;
3869         }
3870
3871         virtnet_set_big_packets(vi, mtu);
3872
3873         if (vi->any_header_sg)
3874                 dev->needed_headroom = vi->hdr_len;
3875
3876         /* Enable multiqueue by default */
3877         if (num_online_cpus() >= max_queue_pairs)
3878                 vi->curr_queue_pairs = max_queue_pairs;
3879         else
3880                 vi->curr_queue_pairs = num_online_cpus();
3881         vi->max_queue_pairs = max_queue_pairs;
3882
3883         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3884         err = init_vqs(vi);
3885         if (err)
3886                 goto free;
3887
3888 #ifdef CONFIG_SYSFS
3889         if (vi->mergeable_rx_bufs)
3890                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3891 #endif
3892         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3893         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3894
3895         virtnet_init_settings(dev);
3896
3897         if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3898                 vi->failover = net_failover_create(vi->dev);
3899                 if (IS_ERR(vi->failover)) {
3900                         err = PTR_ERR(vi->failover);
3901                         goto free_vqs;
3902                 }
3903         }
3904
3905         if (vi->has_rss || vi->has_rss_hash_report)
3906                 virtnet_init_default_rss(vi);
3907
3908         /* serialize netdev register + virtio_device_ready() with ndo_open() */
3909         rtnl_lock();
3910
3911         err = register_netdevice(dev);
3912         if (err) {
3913                 pr_debug("virtio_net: registering device failed\n");
3914                 rtnl_unlock();
3915                 goto free_failover;
3916         }
3917
3918         virtio_device_ready(vdev);
3919
3920         rtnl_unlock();
3921
3922         err = virtnet_cpu_notif_add(vi);
3923         if (err) {
3924                 pr_debug("virtio_net: registering cpu notifier failed\n");
3925                 goto free_unregister_netdev;
3926         }
3927
3928         virtnet_set_queues(vi, vi->curr_queue_pairs);
3929
3930         /* Assume link up if device can't report link status,
3931            otherwise get link status from config. */
3932         netif_carrier_off(dev);
3933         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3934                 schedule_work(&vi->config_work);
3935         } else {
3936                 vi->status = VIRTIO_NET_S_LINK_UP;
3937                 virtnet_update_settings(vi);
3938                 netif_carrier_on(dev);
3939         }
3940
3941         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3942                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3943                         set_bit(guest_offloads[i], &vi->guest_offloads);
3944         vi->guest_offloads_capable = vi->guest_offloads;
3945
3946         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3947                  dev->name, max_queue_pairs);
3948
3949         return 0;
3950
3951 free_unregister_netdev:
3952         unregister_netdev(dev);
3953 free_failover:
3954         net_failover_destroy(vi->failover);
3955 free_vqs:
3956         virtio_reset_device(vdev);
3957         cancel_delayed_work_sync(&vi->refill);
3958         free_receive_page_frags(vi);
3959         virtnet_del_vqs(vi);
3960 free:
3961         free_netdev(dev);
3962         return err;
3963 }
3964
3965 static void remove_vq_common(struct virtnet_info *vi)
3966 {
3967         virtio_reset_device(vi->vdev);
3968
3969         /* Free unused buffers in both send and recv, if any. */
3970         free_unused_bufs(vi);
3971
3972         free_receive_bufs(vi);
3973
3974         free_receive_page_frags(vi);
3975
3976         virtnet_del_vqs(vi);
3977 }
3978
3979 static void virtnet_remove(struct virtio_device *vdev)
3980 {
3981         struct virtnet_info *vi = vdev->priv;
3982
3983         virtnet_cpu_notif_remove(vi);
3984
3985         /* Make sure no work handler is accessing the device. */
3986         flush_work(&vi->config_work);
3987
3988         unregister_netdev(vi->dev);
3989
3990         net_failover_destroy(vi->failover);
3991
3992         remove_vq_common(vi);
3993
3994         free_netdev(vi->dev);
3995 }
3996
3997 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3998 {
3999         struct virtnet_info *vi = vdev->priv;
4000
4001         virtnet_cpu_notif_remove(vi);
4002         virtnet_freeze_down(vdev);
4003         remove_vq_common(vi);
4004
4005         return 0;
4006 }
4007
4008 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
4009 {
4010         struct virtnet_info *vi = vdev->priv;
4011         int err;
4012
4013         err = virtnet_restore_up(vdev);
4014         if (err)
4015                 return err;
4016         virtnet_set_queues(vi, vi->curr_queue_pairs);
4017
4018         err = virtnet_cpu_notif_add(vi);
4019         if (err) {
4020                 virtnet_freeze_down(vdev);
4021                 remove_vq_common(vi);
4022                 return err;
4023         }
4024
4025         return 0;
4026 }
4027
4028 static struct virtio_device_id id_table[] = {
4029         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
4030         { 0 },
4031 };
4032
4033 #define VIRTNET_FEATURES \
4034         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
4035         VIRTIO_NET_F_MAC, \
4036         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
4037         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
4038         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
4039         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
4040         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
4041         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
4042         VIRTIO_NET_F_CTRL_MAC_ADDR, \
4043         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
4044         VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \
4045         VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL
4046
4047 static unsigned int features[] = {
4048         VIRTNET_FEATURES,
4049 };
4050
4051 static unsigned int features_legacy[] = {
4052         VIRTNET_FEATURES,
4053         VIRTIO_NET_F_GSO,
4054         VIRTIO_F_ANY_LAYOUT,
4055 };
4056
4057 static struct virtio_driver virtio_net_driver = {
4058         .feature_table = features,
4059         .feature_table_size = ARRAY_SIZE(features),
4060         .feature_table_legacy = features_legacy,
4061         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
4062         .driver.name =  KBUILD_MODNAME,
4063         .driver.owner = THIS_MODULE,
4064         .id_table =     id_table,
4065         .validate =     virtnet_validate,
4066         .probe =        virtnet_probe,
4067         .remove =       virtnet_remove,
4068         .config_changed = virtnet_config_changed,
4069 #ifdef CONFIG_PM_SLEEP
4070         .freeze =       virtnet_freeze,
4071         .restore =      virtnet_restore,
4072 #endif
4073 };
4074
4075 static __init int virtio_net_driver_init(void)
4076 {
4077         int ret;
4078
4079         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
4080                                       virtnet_cpu_online,
4081                                       virtnet_cpu_down_prep);
4082         if (ret < 0)
4083                 goto out;
4084         virtionet_online = ret;
4085         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
4086                                       NULL, virtnet_cpu_dead);
4087         if (ret)
4088                 goto err_dead;
4089         ret = register_virtio_driver(&virtio_net_driver);
4090         if (ret)
4091                 goto err_virtio;
4092         return 0;
4093 err_virtio:
4094         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4095 err_dead:
4096         cpuhp_remove_multi_state(virtionet_online);
4097 out:
4098         return ret;
4099 }
4100 module_init(virtio_net_driver_init);
4101
4102 static __exit void virtio_net_driver_exit(void)
4103 {
4104         unregister_virtio_driver(&virtio_net_driver);
4105         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4106         cpuhp_remove_multi_state(virtionet_online);
4107 }
4108 module_exit(virtio_net_driver_exit);
4109
4110 MODULE_DEVICE_TABLE(virtio, id_table);
4111 MODULE_DESCRIPTION("Virtio network driver");
4112 MODULE_LICENSE("GPL");