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