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