1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* A network driver using virtio. 3 * 4 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation 5 */ 6 //#define DEBUG 7 #include <linux/netdevice.h> 8 #include <linux/etherdevice.h> 9 #include <linux/ethtool.h> 10 #include <linux/module.h> 11 #include <linux/virtio.h> 12 #include <linux/virtio_net.h> 13 #include <linux/bpf.h> 14 #include <linux/bpf_trace.h> 15 #include <linux/scatterlist.h> 16 #include <linux/if_vlan.h> 17 #include <linux/slab.h> 18 #include <linux/cpu.h> 19 #include <linux/average.h> 20 #include <linux/filter.h> 21 #include <linux/kernel.h> 22 #include <net/route.h> 23 #include <net/xdp.h> 24 #include <net/net_failover.h> 25 26 static int napi_weight = NAPI_POLL_WEIGHT; 27 module_param(napi_weight, int, 0444); 28 29 static bool csum = true, gso = true, napi_tx = true; 30 module_param(csum, bool, 0444); 31 module_param(gso, bool, 0444); 32 module_param(napi_tx, bool, 0644); 33 34 /* FIXME: MTU in config. */ 35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN) 36 #define GOOD_COPY_LEN 128 37 38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD) 39 40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */ 41 #define VIRTIO_XDP_HEADROOM 256 42 43 /* Separating two types of XDP xmit */ 44 #define VIRTIO_XDP_TX BIT(0) 45 #define VIRTIO_XDP_REDIR BIT(1) 46 47 #define VIRTIO_XDP_FLAG BIT(0) 48 49 /* RX packet size EWMA. The average packet size is used to determine the packet 50 * buffer size when refilling RX rings. As the entire RX ring may be refilled 51 * at once, the weight is chosen so that the EWMA will be insensitive to short- 52 * term, transient changes in packet size. 53 */ 54 DECLARE_EWMA(pkt_len, 0, 64) 55 56 #define VIRTNET_DRIVER_VERSION "1.0.0" 57 58 static const unsigned long guest_offloads[] = { 59 VIRTIO_NET_F_GUEST_TSO4, 60 VIRTIO_NET_F_GUEST_TSO6, 61 VIRTIO_NET_F_GUEST_ECN, 62 VIRTIO_NET_F_GUEST_UFO, 63 VIRTIO_NET_F_GUEST_CSUM 64 }; 65 66 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \ 67 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \ 68 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \ 69 (1ULL << VIRTIO_NET_F_GUEST_UFO)) 70 71 struct virtnet_stat_desc { 72 char desc[ETH_GSTRING_LEN]; 73 size_t offset; 74 }; 75 76 struct virtnet_sq_stats { 77 struct u64_stats_sync syncp; 78 u64 packets; 79 u64 bytes; 80 u64 xdp_tx; 81 u64 xdp_tx_drops; 82 u64 kicks; 83 u64 tx_timeouts; 84 }; 85 86 struct virtnet_rq_stats { 87 struct u64_stats_sync syncp; 88 u64 packets; 89 u64 bytes; 90 u64 drops; 91 u64 xdp_packets; 92 u64 xdp_tx; 93 u64 xdp_redirects; 94 u64 xdp_drops; 95 u64 kicks; 96 }; 97 98 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m) 99 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m) 100 101 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = { 102 { "packets", VIRTNET_SQ_STAT(packets) }, 103 { "bytes", VIRTNET_SQ_STAT(bytes) }, 104 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx) }, 105 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops) }, 106 { "kicks", VIRTNET_SQ_STAT(kicks) }, 107 { "tx_timeouts", VIRTNET_SQ_STAT(tx_timeouts) }, 108 }; 109 110 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = { 111 { "packets", VIRTNET_RQ_STAT(packets) }, 112 { "bytes", VIRTNET_RQ_STAT(bytes) }, 113 { "drops", VIRTNET_RQ_STAT(drops) }, 114 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets) }, 115 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx) }, 116 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects) }, 117 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops) }, 118 { "kicks", VIRTNET_RQ_STAT(kicks) }, 119 }; 120 121 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc) 122 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc) 123 124 /* Internal representation of a send virtqueue */ 125 struct send_queue { 126 /* Virtqueue associated with this send _queue */ 127 struct virtqueue *vq; 128 129 /* TX: fragments + linear part + virtio header */ 130 struct scatterlist sg[MAX_SKB_FRAGS + 2]; 131 132 /* Name of the send queue: output.$index */ 133 char name[40]; 134 135 struct virtnet_sq_stats stats; 136 137 struct napi_struct napi; 138 }; 139 140 /* Internal representation of a receive virtqueue */ 141 struct receive_queue { 142 /* Virtqueue associated with this receive_queue */ 143 struct virtqueue *vq; 144 145 struct napi_struct napi; 146 147 struct bpf_prog __rcu *xdp_prog; 148 149 struct virtnet_rq_stats stats; 150 151 /* Chain pages by the private ptr. */ 152 struct page *pages; 153 154 /* Average packet length for mergeable receive buffers. */ 155 struct ewma_pkt_len mrg_avg_pkt_len; 156 157 /* Page frag for packet buffer allocation. */ 158 struct page_frag alloc_frag; 159 160 /* RX: fragments + linear part + virtio header */ 161 struct scatterlist sg[MAX_SKB_FRAGS + 2]; 162 163 /* Min single buffer size for mergeable buffers case. */ 164 unsigned int min_buf_len; 165 166 /* Name of this receive queue: input.$index */ 167 char name[40]; 168 169 struct xdp_rxq_info xdp_rxq; 170 }; 171 172 /* This structure can contain rss message with maximum settings for indirection table and keysize 173 * Note, that default structure that describes RSS configuration virtio_net_rss_config 174 * contains same info but can't handle table values. 175 * In any case, structure would be passed to virtio hw through sg_buf split by parts 176 * because table sizes may be differ according to the device configuration. 177 */ 178 #define VIRTIO_NET_RSS_MAX_KEY_SIZE 40 179 #define VIRTIO_NET_RSS_MAX_TABLE_LEN 128 180 struct virtio_net_ctrl_rss { 181 u32 hash_types; 182 u16 indirection_table_mask; 183 u16 unclassified_queue; 184 u16 indirection_table[VIRTIO_NET_RSS_MAX_TABLE_LEN]; 185 u16 max_tx_vq; 186 u8 hash_key_length; 187 u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE]; 188 }; 189 190 /* Control VQ buffers: protected by the rtnl lock */ 191 struct control_buf { 192 struct virtio_net_ctrl_hdr hdr; 193 virtio_net_ctrl_ack status; 194 struct virtio_net_ctrl_mq mq; 195 u8 promisc; 196 u8 allmulti; 197 __virtio16 vid; 198 __virtio64 offloads; 199 struct virtio_net_ctrl_rss rss; 200 }; 201 202 struct virtnet_info { 203 struct virtio_device *vdev; 204 struct virtqueue *cvq; 205 struct net_device *dev; 206 struct send_queue *sq; 207 struct receive_queue *rq; 208 unsigned int status; 209 210 /* Max # of queue pairs supported by the device */ 211 u16 max_queue_pairs; 212 213 /* # of queue pairs currently used by the driver */ 214 u16 curr_queue_pairs; 215 216 /* # of XDP queue pairs currently used by the driver */ 217 u16 xdp_queue_pairs; 218 219 /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */ 220 bool xdp_enabled; 221 222 /* I like... big packets and I cannot lie! */ 223 bool big_packets; 224 225 /* Host will merge rx buffers for big packets (shake it! shake it!) */ 226 bool mergeable_rx_bufs; 227 228 /* Host supports rss and/or hash report */ 229 bool has_rss; 230 bool has_rss_hash_report; 231 u8 rss_key_size; 232 u16 rss_indir_table_size; 233 u32 rss_hash_types_supported; 234 u32 rss_hash_types_saved; 235 236 /* Has control virtqueue */ 237 bool has_cvq; 238 239 /* Host can handle any s/g split between our header and packet data */ 240 bool any_header_sg; 241 242 /* Packet virtio header size */ 243 u8 hdr_len; 244 245 /* Work struct for delayed refilling if we run low on memory. */ 246 struct delayed_work refill; 247 248 /* Is delayed refill enabled? */ 249 bool refill_enabled; 250 251 /* The lock to synchronize the access to refill_enabled */ 252 spinlock_t refill_lock; 253 254 /* Work struct for config space updates */ 255 struct work_struct config_work; 256 257 /* Does the affinity hint is set for virtqueues? */ 258 bool affinity_hint_set; 259 260 /* CPU hotplug instances for online & dead */ 261 struct hlist_node node; 262 struct hlist_node node_dead; 263 264 struct control_buf *ctrl; 265 266 /* Ethtool settings */ 267 u8 duplex; 268 u32 speed; 269 270 unsigned long guest_offloads; 271 unsigned long guest_offloads_capable; 272 273 /* failover when STANDBY feature enabled */ 274 struct failover *failover; 275 }; 276 277 struct padded_vnet_hdr { 278 struct virtio_net_hdr_v1_hash hdr; 279 /* 280 * hdr is in a separate sg buffer, and data sg buffer shares same page 281 * with this header sg. This padding makes next sg 16 byte aligned 282 * after the header. 283 */ 284 char padding[12]; 285 }; 286 287 static bool is_xdp_frame(void *ptr) 288 { 289 return (unsigned long)ptr & VIRTIO_XDP_FLAG; 290 } 291 292 static void *xdp_to_ptr(struct xdp_frame *ptr) 293 { 294 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG); 295 } 296 297 static struct xdp_frame *ptr_to_xdp(void *ptr) 298 { 299 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG); 300 } 301 302 /* Converting between virtqueue no. and kernel tx/rx queue no. 303 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq 304 */ 305 static int vq2txq(struct virtqueue *vq) 306 { 307 return (vq->index - 1) / 2; 308 } 309 310 static int txq2vq(int txq) 311 { 312 return txq * 2 + 1; 313 } 314 315 static int vq2rxq(struct virtqueue *vq) 316 { 317 return vq->index / 2; 318 } 319 320 static int rxq2vq(int rxq) 321 { 322 return rxq * 2; 323 } 324 325 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb) 326 { 327 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb; 328 } 329 330 /* 331 * private is used to chain pages for big packets, put the whole 332 * most recent used list in the beginning for reuse 333 */ 334 static void give_pages(struct receive_queue *rq, struct page *page) 335 { 336 struct page *end; 337 338 /* Find end of list, sew whole thing into vi->rq.pages. */ 339 for (end = page; end->private; end = (struct page *)end->private); 340 end->private = (unsigned long)rq->pages; 341 rq->pages = page; 342 } 343 344 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask) 345 { 346 struct page *p = rq->pages; 347 348 if (p) { 349 rq->pages = (struct page *)p->private; 350 /* clear private here, it is used to chain pages */ 351 p->private = 0; 352 } else 353 p = alloc_page(gfp_mask); 354 return p; 355 } 356 357 static void enable_delayed_refill(struct virtnet_info *vi) 358 { 359 spin_lock_bh(&vi->refill_lock); 360 vi->refill_enabled = true; 361 spin_unlock_bh(&vi->refill_lock); 362 } 363 364 static void disable_delayed_refill(struct virtnet_info *vi) 365 { 366 spin_lock_bh(&vi->refill_lock); 367 vi->refill_enabled = false; 368 spin_unlock_bh(&vi->refill_lock); 369 } 370 371 static void virtqueue_napi_schedule(struct napi_struct *napi, 372 struct virtqueue *vq) 373 { 374 if (napi_schedule_prep(napi)) { 375 virtqueue_disable_cb(vq); 376 __napi_schedule(napi); 377 } 378 } 379 380 static void virtqueue_napi_complete(struct napi_struct *napi, 381 struct virtqueue *vq, int processed) 382 { 383 int opaque; 384 385 opaque = virtqueue_enable_cb_prepare(vq); 386 if (napi_complete_done(napi, processed)) { 387 if (unlikely(virtqueue_poll(vq, opaque))) 388 virtqueue_napi_schedule(napi, vq); 389 } else { 390 virtqueue_disable_cb(vq); 391 } 392 } 393 394 static void skb_xmit_done(struct virtqueue *vq) 395 { 396 struct virtnet_info *vi = vq->vdev->priv; 397 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi; 398 399 /* Suppress further interrupts. */ 400 virtqueue_disable_cb(vq); 401 402 if (napi->weight) 403 virtqueue_napi_schedule(napi, vq); 404 else 405 /* We were probably waiting for more output buffers. */ 406 netif_wake_subqueue(vi->dev, vq2txq(vq)); 407 } 408 409 #define MRG_CTX_HEADER_SHIFT 22 410 static void *mergeable_len_to_ctx(unsigned int truesize, 411 unsigned int headroom) 412 { 413 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize); 414 } 415 416 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx) 417 { 418 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT; 419 } 420 421 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx) 422 { 423 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1); 424 } 425 426 /* Called from bottom half context */ 427 static struct sk_buff *page_to_skb(struct virtnet_info *vi, 428 struct receive_queue *rq, 429 struct page *page, unsigned int offset, 430 unsigned int len, unsigned int truesize, 431 bool hdr_valid, unsigned int metasize, 432 unsigned int headroom) 433 { 434 struct sk_buff *skb; 435 struct virtio_net_hdr_mrg_rxbuf *hdr; 436 unsigned int copy, hdr_len, hdr_padded_len; 437 struct page *page_to_free = NULL; 438 int tailroom, shinfo_size; 439 char *p, *hdr_p, *buf; 440 441 p = page_address(page) + offset; 442 hdr_p = p; 443 444 hdr_len = vi->hdr_len; 445 if (vi->mergeable_rx_bufs) 446 hdr_padded_len = hdr_len; 447 else 448 hdr_padded_len = sizeof(struct padded_vnet_hdr); 449 450 /* If headroom is not 0, there is an offset between the beginning of the 451 * data and the allocated space, otherwise the data and the allocated 452 * space are aligned. 453 * 454 * Buffers with headroom use PAGE_SIZE as alloc size, see 455 * add_recvbuf_mergeable() + get_mergeable_buf_len() 456 */ 457 truesize = headroom ? PAGE_SIZE : truesize; 458 tailroom = truesize - headroom; 459 buf = p - headroom; 460 461 len -= hdr_len; 462 offset += hdr_padded_len; 463 p += hdr_padded_len; 464 tailroom -= hdr_padded_len + len; 465 466 shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 467 468 /* copy small packet so we can reuse these pages */ 469 if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) { 470 skb = build_skb(buf, truesize); 471 if (unlikely(!skb)) 472 return NULL; 473 474 skb_reserve(skb, p - buf); 475 skb_put(skb, len); 476 477 page = (struct page *)page->private; 478 if (page) 479 give_pages(rq, page); 480 goto ok; 481 } 482 483 /* copy small packet so we can reuse these pages for small data */ 484 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN); 485 if (unlikely(!skb)) 486 return NULL; 487 488 /* Copy all frame if it fits skb->head, otherwise 489 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed. 490 */ 491 if (len <= skb_tailroom(skb)) 492 copy = len; 493 else 494 copy = ETH_HLEN + metasize; 495 skb_put_data(skb, p, copy); 496 497 len -= copy; 498 offset += copy; 499 500 if (vi->mergeable_rx_bufs) { 501 if (len) 502 skb_add_rx_frag(skb, 0, page, offset, len, truesize); 503 else 504 page_to_free = page; 505 goto ok; 506 } 507 508 /* 509 * Verify that we can indeed put this data into a skb. 510 * This is here to handle cases when the device erroneously 511 * tries to receive more than is possible. This is usually 512 * the case of a broken device. 513 */ 514 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) { 515 net_dbg_ratelimited("%s: too much data\n", skb->dev->name); 516 dev_kfree_skb(skb); 517 return NULL; 518 } 519 BUG_ON(offset >= PAGE_SIZE); 520 while (len) { 521 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len); 522 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset, 523 frag_size, truesize); 524 len -= frag_size; 525 page = (struct page *)page->private; 526 offset = 0; 527 } 528 529 if (page) 530 give_pages(rq, page); 531 532 ok: 533 /* hdr_valid means no XDP, so we can copy the vnet header */ 534 if (hdr_valid) { 535 hdr = skb_vnet_hdr(skb); 536 memcpy(hdr, hdr_p, hdr_len); 537 } 538 if (page_to_free) 539 put_page(page_to_free); 540 541 if (metasize) { 542 __skb_pull(skb, metasize); 543 skb_metadata_set(skb, metasize); 544 } 545 546 return skb; 547 } 548 549 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi, 550 struct send_queue *sq, 551 struct xdp_frame *xdpf) 552 { 553 struct virtio_net_hdr_mrg_rxbuf *hdr; 554 int err; 555 556 if (unlikely(xdpf->headroom < vi->hdr_len)) 557 return -EOVERFLOW; 558 559 /* Make room for virtqueue hdr (also change xdpf->headroom?) */ 560 xdpf->data -= vi->hdr_len; 561 /* Zero header and leave csum up to XDP layers */ 562 hdr = xdpf->data; 563 memset(hdr, 0, vi->hdr_len); 564 xdpf->len += vi->hdr_len; 565 566 sg_init_one(sq->sg, xdpf->data, xdpf->len); 567 568 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf), 569 GFP_ATOMIC); 570 if (unlikely(err)) 571 return -ENOSPC; /* Caller handle free/refcnt */ 572 573 return 0; 574 } 575 576 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on 577 * the current cpu, so it does not need to be locked. 578 * 579 * Here we use marco instead of inline functions because we have to deal with 580 * three issues at the same time: 1. the choice of sq. 2. judge and execute the 581 * lock/unlock of txq 3. make sparse happy. It is difficult for two inline 582 * functions to perfectly solve these three problems at the same time. 583 */ 584 #define virtnet_xdp_get_sq(vi) ({ \ 585 int cpu = smp_processor_id(); \ 586 struct netdev_queue *txq; \ 587 typeof(vi) v = (vi); \ 588 unsigned int qp; \ 589 \ 590 if (v->curr_queue_pairs > nr_cpu_ids) { \ 591 qp = v->curr_queue_pairs - v->xdp_queue_pairs; \ 592 qp += cpu; \ 593 txq = netdev_get_tx_queue(v->dev, qp); \ 594 __netif_tx_acquire(txq); \ 595 } else { \ 596 qp = cpu % v->curr_queue_pairs; \ 597 txq = netdev_get_tx_queue(v->dev, qp); \ 598 __netif_tx_lock(txq, cpu); \ 599 } \ 600 v->sq + qp; \ 601 }) 602 603 #define virtnet_xdp_put_sq(vi, q) { \ 604 struct netdev_queue *txq; \ 605 typeof(vi) v = (vi); \ 606 \ 607 txq = netdev_get_tx_queue(v->dev, (q) - v->sq); \ 608 if (v->curr_queue_pairs > nr_cpu_ids) \ 609 __netif_tx_release(txq); \ 610 else \ 611 __netif_tx_unlock(txq); \ 612 } 613 614 static int virtnet_xdp_xmit(struct net_device *dev, 615 int n, struct xdp_frame **frames, u32 flags) 616 { 617 struct virtnet_info *vi = netdev_priv(dev); 618 struct receive_queue *rq = vi->rq; 619 struct bpf_prog *xdp_prog; 620 struct send_queue *sq; 621 unsigned int len; 622 int packets = 0; 623 int bytes = 0; 624 int nxmit = 0; 625 int kicks = 0; 626 void *ptr; 627 int ret; 628 int i; 629 630 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this 631 * indicate XDP resources have been successfully allocated. 632 */ 633 xdp_prog = rcu_access_pointer(rq->xdp_prog); 634 if (!xdp_prog) 635 return -ENXIO; 636 637 sq = virtnet_xdp_get_sq(vi); 638 639 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) { 640 ret = -EINVAL; 641 goto out; 642 } 643 644 /* Free up any pending old buffers before queueing new ones. */ 645 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) { 646 if (likely(is_xdp_frame(ptr))) { 647 struct xdp_frame *frame = ptr_to_xdp(ptr); 648 649 bytes += frame->len; 650 xdp_return_frame(frame); 651 } else { 652 struct sk_buff *skb = ptr; 653 654 bytes += skb->len; 655 napi_consume_skb(skb, false); 656 } 657 packets++; 658 } 659 660 for (i = 0; i < n; i++) { 661 struct xdp_frame *xdpf = frames[i]; 662 663 if (__virtnet_xdp_xmit_one(vi, sq, xdpf)) 664 break; 665 nxmit++; 666 } 667 ret = nxmit; 668 669 if (flags & XDP_XMIT_FLUSH) { 670 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) 671 kicks = 1; 672 } 673 out: 674 u64_stats_update_begin(&sq->stats.syncp); 675 sq->stats.bytes += bytes; 676 sq->stats.packets += packets; 677 sq->stats.xdp_tx += n; 678 sq->stats.xdp_tx_drops += n - nxmit; 679 sq->stats.kicks += kicks; 680 u64_stats_update_end(&sq->stats.syncp); 681 682 virtnet_xdp_put_sq(vi, sq); 683 return ret; 684 } 685 686 static unsigned int virtnet_get_headroom(struct virtnet_info *vi) 687 { 688 return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0; 689 } 690 691 /* We copy the packet for XDP in the following cases: 692 * 693 * 1) Packet is scattered across multiple rx buffers. 694 * 2) Headroom space is insufficient. 695 * 696 * This is inefficient but it's a temporary condition that 697 * we hit right after XDP is enabled and until queue is refilled 698 * with large buffers with sufficient headroom - so it should affect 699 * at most queue size packets. 700 * Afterwards, the conditions to enable 701 * XDP should preclude the underlying device from sending packets 702 * across multiple buffers (num_buf > 1), and we make sure buffers 703 * have enough headroom. 704 */ 705 static struct page *xdp_linearize_page(struct receive_queue *rq, 706 u16 *num_buf, 707 struct page *p, 708 int offset, 709 int page_off, 710 unsigned int *len) 711 { 712 struct page *page = alloc_page(GFP_ATOMIC); 713 714 if (!page) 715 return NULL; 716 717 memcpy(page_address(page) + page_off, page_address(p) + offset, *len); 718 page_off += *len; 719 720 while (--*num_buf) { 721 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 722 unsigned int buflen; 723 void *buf; 724 int off; 725 726 buf = virtqueue_get_buf(rq->vq, &buflen); 727 if (unlikely(!buf)) 728 goto err_buf; 729 730 p = virt_to_head_page(buf); 731 off = buf - page_address(p); 732 733 /* guard against a misconfigured or uncooperative backend that 734 * is sending packet larger than the MTU. 735 */ 736 if ((page_off + buflen + tailroom) > PAGE_SIZE) { 737 put_page(p); 738 goto err_buf; 739 } 740 741 memcpy(page_address(page) + page_off, 742 page_address(p) + off, buflen); 743 page_off += buflen; 744 put_page(p); 745 } 746 747 /* Headroom does not contribute to packet length */ 748 *len = page_off - VIRTIO_XDP_HEADROOM; 749 return page; 750 err_buf: 751 __free_pages(page, 0); 752 return NULL; 753 } 754 755 static struct sk_buff *receive_small(struct net_device *dev, 756 struct virtnet_info *vi, 757 struct receive_queue *rq, 758 void *buf, void *ctx, 759 unsigned int len, 760 unsigned int *xdp_xmit, 761 struct virtnet_rq_stats *stats) 762 { 763 struct sk_buff *skb; 764 struct bpf_prog *xdp_prog; 765 unsigned int xdp_headroom = (unsigned long)ctx; 766 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom; 767 unsigned int headroom = vi->hdr_len + header_offset; 768 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 769 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 770 struct page *page = virt_to_head_page(buf); 771 unsigned int delta = 0; 772 struct page *xdp_page; 773 int err; 774 unsigned int metasize = 0; 775 776 len -= vi->hdr_len; 777 stats->bytes += len; 778 779 if (unlikely(len > GOOD_PACKET_LEN)) { 780 pr_debug("%s: rx error: len %u exceeds max size %d\n", 781 dev->name, len, GOOD_PACKET_LEN); 782 dev->stats.rx_length_errors++; 783 goto err; 784 } 785 786 if (likely(!vi->xdp_enabled)) { 787 xdp_prog = NULL; 788 goto skip_xdp; 789 } 790 791 rcu_read_lock(); 792 xdp_prog = rcu_dereference(rq->xdp_prog); 793 if (xdp_prog) { 794 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset; 795 struct xdp_frame *xdpf; 796 struct xdp_buff xdp; 797 void *orig_data; 798 u32 act; 799 800 if (unlikely(hdr->hdr.gso_type)) 801 goto err_xdp; 802 803 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) { 804 int offset = buf - page_address(page) + header_offset; 805 unsigned int tlen = len + vi->hdr_len; 806 u16 num_buf = 1; 807 808 xdp_headroom = virtnet_get_headroom(vi); 809 header_offset = VIRTNET_RX_PAD + xdp_headroom; 810 headroom = vi->hdr_len + header_offset; 811 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 812 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 813 xdp_page = xdp_linearize_page(rq, &num_buf, page, 814 offset, header_offset, 815 &tlen); 816 if (!xdp_page) 817 goto err_xdp; 818 819 buf = page_address(xdp_page); 820 put_page(page); 821 page = xdp_page; 822 } 823 824 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq); 825 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len, 826 xdp_headroom, len, true); 827 orig_data = xdp.data; 828 act = bpf_prog_run_xdp(xdp_prog, &xdp); 829 stats->xdp_packets++; 830 831 switch (act) { 832 case XDP_PASS: 833 /* Recalculate length in case bpf program changed it */ 834 delta = orig_data - xdp.data; 835 len = xdp.data_end - xdp.data; 836 metasize = xdp.data - xdp.data_meta; 837 break; 838 case XDP_TX: 839 stats->xdp_tx++; 840 xdpf = xdp_convert_buff_to_frame(&xdp); 841 if (unlikely(!xdpf)) 842 goto err_xdp; 843 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0); 844 if (unlikely(!err)) { 845 xdp_return_frame_rx_napi(xdpf); 846 } else if (unlikely(err < 0)) { 847 trace_xdp_exception(vi->dev, xdp_prog, act); 848 goto err_xdp; 849 } 850 *xdp_xmit |= VIRTIO_XDP_TX; 851 rcu_read_unlock(); 852 goto xdp_xmit; 853 case XDP_REDIRECT: 854 stats->xdp_redirects++; 855 err = xdp_do_redirect(dev, &xdp, xdp_prog); 856 if (err) 857 goto err_xdp; 858 *xdp_xmit |= VIRTIO_XDP_REDIR; 859 rcu_read_unlock(); 860 goto xdp_xmit; 861 default: 862 bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act); 863 fallthrough; 864 case XDP_ABORTED: 865 trace_xdp_exception(vi->dev, xdp_prog, act); 866 goto err_xdp; 867 case XDP_DROP: 868 goto err_xdp; 869 } 870 } 871 rcu_read_unlock(); 872 873 skip_xdp: 874 skb = build_skb(buf, buflen); 875 if (!skb) 876 goto err; 877 skb_reserve(skb, headroom - delta); 878 skb_put(skb, len); 879 if (!xdp_prog) { 880 buf += header_offset; 881 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len); 882 } /* keep zeroed vnet hdr since XDP is loaded */ 883 884 if (metasize) 885 skb_metadata_set(skb, metasize); 886 887 return skb; 888 889 err_xdp: 890 rcu_read_unlock(); 891 stats->xdp_drops++; 892 err: 893 stats->drops++; 894 put_page(page); 895 xdp_xmit: 896 return NULL; 897 } 898 899 static struct sk_buff *receive_big(struct net_device *dev, 900 struct virtnet_info *vi, 901 struct receive_queue *rq, 902 void *buf, 903 unsigned int len, 904 struct virtnet_rq_stats *stats) 905 { 906 struct page *page = buf; 907 struct sk_buff *skb = 908 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0); 909 910 stats->bytes += len - vi->hdr_len; 911 if (unlikely(!skb)) 912 goto err; 913 914 return skb; 915 916 err: 917 stats->drops++; 918 give_pages(rq, page); 919 return NULL; 920 } 921 922 static struct sk_buff *receive_mergeable(struct net_device *dev, 923 struct virtnet_info *vi, 924 struct receive_queue *rq, 925 void *buf, 926 void *ctx, 927 unsigned int len, 928 unsigned int *xdp_xmit, 929 struct virtnet_rq_stats *stats) 930 { 931 struct virtio_net_hdr_mrg_rxbuf *hdr = buf; 932 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers); 933 struct page *page = virt_to_head_page(buf); 934 int offset = buf - page_address(page); 935 struct sk_buff *head_skb, *curr_skb; 936 struct bpf_prog *xdp_prog; 937 unsigned int truesize = mergeable_ctx_to_truesize(ctx); 938 unsigned int headroom = mergeable_ctx_to_headroom(ctx); 939 unsigned int metasize = 0; 940 unsigned int frame_sz; 941 int err; 942 943 head_skb = NULL; 944 stats->bytes += len - vi->hdr_len; 945 946 if (unlikely(len > truesize)) { 947 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 948 dev->name, len, (unsigned long)ctx); 949 dev->stats.rx_length_errors++; 950 goto err_skb; 951 } 952 953 if (likely(!vi->xdp_enabled)) { 954 xdp_prog = NULL; 955 goto skip_xdp; 956 } 957 958 rcu_read_lock(); 959 xdp_prog = rcu_dereference(rq->xdp_prog); 960 if (xdp_prog) { 961 struct xdp_frame *xdpf; 962 struct page *xdp_page; 963 struct xdp_buff xdp; 964 void *data; 965 u32 act; 966 967 /* Transient failure which in theory could occur if 968 * in-flight packets from before XDP was enabled reach 969 * the receive path after XDP is loaded. 970 */ 971 if (unlikely(hdr->hdr.gso_type)) 972 goto err_xdp; 973 974 /* Buffers with headroom use PAGE_SIZE as alloc size, 975 * see add_recvbuf_mergeable() + get_mergeable_buf_len() 976 */ 977 frame_sz = headroom ? PAGE_SIZE : truesize; 978 979 /* This happens when rx buffer size is underestimated 980 * or headroom is not enough because of the buffer 981 * was refilled before XDP is set. This should only 982 * happen for the first several packets, so we don't 983 * care much about its performance. 984 */ 985 if (unlikely(num_buf > 1 || 986 headroom < virtnet_get_headroom(vi))) { 987 /* linearize data for XDP */ 988 xdp_page = xdp_linearize_page(rq, &num_buf, 989 page, offset, 990 VIRTIO_XDP_HEADROOM, 991 &len); 992 frame_sz = PAGE_SIZE; 993 994 if (!xdp_page) 995 goto err_xdp; 996 offset = VIRTIO_XDP_HEADROOM; 997 } else { 998 xdp_page = page; 999 } 1000 1001 /* Allow consuming headroom but reserve enough space to push 1002 * the descriptor on if we get an XDP_TX return code. 1003 */ 1004 data = page_address(xdp_page) + offset; 1005 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq); 1006 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len, 1007 VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true); 1008 1009 act = bpf_prog_run_xdp(xdp_prog, &xdp); 1010 stats->xdp_packets++; 1011 1012 switch (act) { 1013 case XDP_PASS: 1014 metasize = xdp.data - xdp.data_meta; 1015 1016 /* recalculate offset to account for any header 1017 * adjustments and minus the metasize to copy the 1018 * metadata in page_to_skb(). Note other cases do not 1019 * build an skb and avoid using offset 1020 */ 1021 offset = xdp.data - page_address(xdp_page) - 1022 vi->hdr_len - metasize; 1023 1024 /* recalculate len if xdp.data, xdp.data_end or 1025 * xdp.data_meta were adjusted 1026 */ 1027 len = xdp.data_end - xdp.data + vi->hdr_len + metasize; 1028 1029 /* recalculate headroom if xdp.data or xdp_data_meta 1030 * were adjusted, note that offset should always point 1031 * to the start of the reserved bytes for virtio_net 1032 * header which are followed by xdp.data, that means 1033 * that offset is equal to the headroom (when buf is 1034 * starting at the beginning of the page, otherwise 1035 * there is a base offset inside the page) but it's used 1036 * with a different starting point (buf start) than 1037 * xdp.data (buf start + vnet hdr size). If xdp.data or 1038 * data_meta were adjusted by the xdp prog then the 1039 * headroom size has changed and so has the offset, we 1040 * can use data_hard_start, which points at buf start + 1041 * vnet hdr size, to calculate the new headroom and use 1042 * it later to compute buf start in page_to_skb() 1043 */ 1044 headroom = xdp.data - xdp.data_hard_start - metasize; 1045 1046 /* We can only create skb based on xdp_page. */ 1047 if (unlikely(xdp_page != page)) { 1048 rcu_read_unlock(); 1049 put_page(page); 1050 head_skb = page_to_skb(vi, rq, xdp_page, offset, 1051 len, PAGE_SIZE, false, 1052 metasize, 1053 headroom); 1054 return head_skb; 1055 } 1056 break; 1057 case XDP_TX: 1058 stats->xdp_tx++; 1059 xdpf = xdp_convert_buff_to_frame(&xdp); 1060 if (unlikely(!xdpf)) { 1061 if (unlikely(xdp_page != page)) 1062 put_page(xdp_page); 1063 goto err_xdp; 1064 } 1065 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0); 1066 if (unlikely(!err)) { 1067 xdp_return_frame_rx_napi(xdpf); 1068 } else if (unlikely(err < 0)) { 1069 trace_xdp_exception(vi->dev, xdp_prog, act); 1070 if (unlikely(xdp_page != page)) 1071 put_page(xdp_page); 1072 goto err_xdp; 1073 } 1074 *xdp_xmit |= VIRTIO_XDP_TX; 1075 if (unlikely(xdp_page != page)) 1076 put_page(page); 1077 rcu_read_unlock(); 1078 goto xdp_xmit; 1079 case XDP_REDIRECT: 1080 stats->xdp_redirects++; 1081 err = xdp_do_redirect(dev, &xdp, xdp_prog); 1082 if (err) { 1083 if (unlikely(xdp_page != page)) 1084 put_page(xdp_page); 1085 goto err_xdp; 1086 } 1087 *xdp_xmit |= VIRTIO_XDP_REDIR; 1088 if (unlikely(xdp_page != page)) 1089 put_page(page); 1090 rcu_read_unlock(); 1091 goto xdp_xmit; 1092 default: 1093 bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act); 1094 fallthrough; 1095 case XDP_ABORTED: 1096 trace_xdp_exception(vi->dev, xdp_prog, act); 1097 fallthrough; 1098 case XDP_DROP: 1099 if (unlikely(xdp_page != page)) 1100 __free_pages(xdp_page, 0); 1101 goto err_xdp; 1102 } 1103 } 1104 rcu_read_unlock(); 1105 1106 skip_xdp: 1107 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog, 1108 metasize, headroom); 1109 curr_skb = head_skb; 1110 1111 if (unlikely(!curr_skb)) 1112 goto err_skb; 1113 while (--num_buf) { 1114 int num_skb_frags; 1115 1116 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx); 1117 if (unlikely(!buf)) { 1118 pr_debug("%s: rx error: %d buffers out of %d missing\n", 1119 dev->name, num_buf, 1120 virtio16_to_cpu(vi->vdev, 1121 hdr->num_buffers)); 1122 dev->stats.rx_length_errors++; 1123 goto err_buf; 1124 } 1125 1126 stats->bytes += len; 1127 page = virt_to_head_page(buf); 1128 1129 truesize = mergeable_ctx_to_truesize(ctx); 1130 if (unlikely(len > truesize)) { 1131 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 1132 dev->name, len, (unsigned long)ctx); 1133 dev->stats.rx_length_errors++; 1134 goto err_skb; 1135 } 1136 1137 num_skb_frags = skb_shinfo(curr_skb)->nr_frags; 1138 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) { 1139 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC); 1140 1141 if (unlikely(!nskb)) 1142 goto err_skb; 1143 if (curr_skb == head_skb) 1144 skb_shinfo(curr_skb)->frag_list = nskb; 1145 else 1146 curr_skb->next = nskb; 1147 curr_skb = nskb; 1148 head_skb->truesize += nskb->truesize; 1149 num_skb_frags = 0; 1150 } 1151 if (curr_skb != head_skb) { 1152 head_skb->data_len += len; 1153 head_skb->len += len; 1154 head_skb->truesize += truesize; 1155 } 1156 offset = buf - page_address(page); 1157 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) { 1158 put_page(page); 1159 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1, 1160 len, truesize); 1161 } else { 1162 skb_add_rx_frag(curr_skb, num_skb_frags, page, 1163 offset, len, truesize); 1164 } 1165 } 1166 1167 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len); 1168 return head_skb; 1169 1170 err_xdp: 1171 rcu_read_unlock(); 1172 stats->xdp_drops++; 1173 err_skb: 1174 put_page(page); 1175 while (num_buf-- > 1) { 1176 buf = virtqueue_get_buf(rq->vq, &len); 1177 if (unlikely(!buf)) { 1178 pr_debug("%s: rx error: %d buffers missing\n", 1179 dev->name, num_buf); 1180 dev->stats.rx_length_errors++; 1181 break; 1182 } 1183 stats->bytes += len; 1184 page = virt_to_head_page(buf); 1185 put_page(page); 1186 } 1187 err_buf: 1188 stats->drops++; 1189 dev_kfree_skb(head_skb); 1190 xdp_xmit: 1191 return NULL; 1192 } 1193 1194 static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash, 1195 struct sk_buff *skb) 1196 { 1197 enum pkt_hash_types rss_hash_type; 1198 1199 if (!hdr_hash || !skb) 1200 return; 1201 1202 switch (__le16_to_cpu(hdr_hash->hash_report)) { 1203 case VIRTIO_NET_HASH_REPORT_TCPv4: 1204 case VIRTIO_NET_HASH_REPORT_UDPv4: 1205 case VIRTIO_NET_HASH_REPORT_TCPv6: 1206 case VIRTIO_NET_HASH_REPORT_UDPv6: 1207 case VIRTIO_NET_HASH_REPORT_TCPv6_EX: 1208 case VIRTIO_NET_HASH_REPORT_UDPv6_EX: 1209 rss_hash_type = PKT_HASH_TYPE_L4; 1210 break; 1211 case VIRTIO_NET_HASH_REPORT_IPv4: 1212 case VIRTIO_NET_HASH_REPORT_IPv6: 1213 case VIRTIO_NET_HASH_REPORT_IPv6_EX: 1214 rss_hash_type = PKT_HASH_TYPE_L3; 1215 break; 1216 case VIRTIO_NET_HASH_REPORT_NONE: 1217 default: 1218 rss_hash_type = PKT_HASH_TYPE_NONE; 1219 } 1220 skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type); 1221 } 1222 1223 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq, 1224 void *buf, unsigned int len, void **ctx, 1225 unsigned int *xdp_xmit, 1226 struct virtnet_rq_stats *stats) 1227 { 1228 struct net_device *dev = vi->dev; 1229 struct sk_buff *skb; 1230 struct virtio_net_hdr_mrg_rxbuf *hdr; 1231 1232 if (unlikely(len < vi->hdr_len + ETH_HLEN)) { 1233 pr_debug("%s: short packet %i\n", dev->name, len); 1234 dev->stats.rx_length_errors++; 1235 if (vi->mergeable_rx_bufs) { 1236 put_page(virt_to_head_page(buf)); 1237 } else if (vi->big_packets) { 1238 give_pages(rq, buf); 1239 } else { 1240 put_page(virt_to_head_page(buf)); 1241 } 1242 return; 1243 } 1244 1245 if (vi->mergeable_rx_bufs) 1246 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit, 1247 stats); 1248 else if (vi->big_packets) 1249 skb = receive_big(dev, vi, rq, buf, len, stats); 1250 else 1251 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats); 1252 1253 if (unlikely(!skb)) 1254 return; 1255 1256 hdr = skb_vnet_hdr(skb); 1257 if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report) 1258 virtio_skb_set_hash((const struct virtio_net_hdr_v1_hash *)hdr, skb); 1259 1260 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) 1261 skb->ip_summed = CHECKSUM_UNNECESSARY; 1262 1263 if (virtio_net_hdr_to_skb(skb, &hdr->hdr, 1264 virtio_is_little_endian(vi->vdev))) { 1265 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n", 1266 dev->name, hdr->hdr.gso_type, 1267 hdr->hdr.gso_size); 1268 goto frame_err; 1269 } 1270 1271 skb_record_rx_queue(skb, vq2rxq(rq->vq)); 1272 skb->protocol = eth_type_trans(skb, dev); 1273 pr_debug("Receiving skb proto 0x%04x len %i type %i\n", 1274 ntohs(skb->protocol), skb->len, skb->pkt_type); 1275 1276 napi_gro_receive(&rq->napi, skb); 1277 return; 1278 1279 frame_err: 1280 dev->stats.rx_frame_errors++; 1281 dev_kfree_skb(skb); 1282 } 1283 1284 /* Unlike mergeable buffers, all buffers are allocated to the 1285 * same size, except for the headroom. For this reason we do 1286 * not need to use mergeable_len_to_ctx here - it is enough 1287 * to store the headroom as the context ignoring the truesize. 1288 */ 1289 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq, 1290 gfp_t gfp) 1291 { 1292 struct page_frag *alloc_frag = &rq->alloc_frag; 1293 char *buf; 1294 unsigned int xdp_headroom = virtnet_get_headroom(vi); 1295 void *ctx = (void *)(unsigned long)xdp_headroom; 1296 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom; 1297 int err; 1298 1299 len = SKB_DATA_ALIGN(len) + 1300 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 1301 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp))) 1302 return -ENOMEM; 1303 1304 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1305 get_page(alloc_frag->page); 1306 alloc_frag->offset += len; 1307 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom, 1308 vi->hdr_len + GOOD_PACKET_LEN); 1309 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1310 if (err < 0) 1311 put_page(virt_to_head_page(buf)); 1312 return err; 1313 } 1314 1315 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq, 1316 gfp_t gfp) 1317 { 1318 struct page *first, *list = NULL; 1319 char *p; 1320 int i, err, offset; 1321 1322 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2); 1323 1324 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */ 1325 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) { 1326 first = get_a_page(rq, gfp); 1327 if (!first) { 1328 if (list) 1329 give_pages(rq, list); 1330 return -ENOMEM; 1331 } 1332 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE); 1333 1334 /* chain new page in list head to match sg */ 1335 first->private = (unsigned long)list; 1336 list = first; 1337 } 1338 1339 first = get_a_page(rq, gfp); 1340 if (!first) { 1341 give_pages(rq, list); 1342 return -ENOMEM; 1343 } 1344 p = page_address(first); 1345 1346 /* rq->sg[0], rq->sg[1] share the same page */ 1347 /* a separated rq->sg[0] for header - required in case !any_header_sg */ 1348 sg_set_buf(&rq->sg[0], p, vi->hdr_len); 1349 1350 /* rq->sg[1] for data packet, from offset */ 1351 offset = sizeof(struct padded_vnet_hdr); 1352 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset); 1353 1354 /* chain first in list head */ 1355 first->private = (unsigned long)list; 1356 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2, 1357 first, gfp); 1358 if (err < 0) 1359 give_pages(rq, first); 1360 1361 return err; 1362 } 1363 1364 static unsigned int get_mergeable_buf_len(struct receive_queue *rq, 1365 struct ewma_pkt_len *avg_pkt_len, 1366 unsigned int room) 1367 { 1368 struct virtnet_info *vi = rq->vq->vdev->priv; 1369 const size_t hdr_len = vi->hdr_len; 1370 unsigned int len; 1371 1372 if (room) 1373 return PAGE_SIZE - room; 1374 1375 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len), 1376 rq->min_buf_len, PAGE_SIZE - hdr_len); 1377 1378 return ALIGN(len, L1_CACHE_BYTES); 1379 } 1380 1381 static int add_recvbuf_mergeable(struct virtnet_info *vi, 1382 struct receive_queue *rq, gfp_t gfp) 1383 { 1384 struct page_frag *alloc_frag = &rq->alloc_frag; 1385 unsigned int headroom = virtnet_get_headroom(vi); 1386 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 1387 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom); 1388 char *buf; 1389 void *ctx; 1390 int err; 1391 unsigned int len, hole; 1392 1393 /* Extra tailroom is needed to satisfy XDP's assumption. This 1394 * means rx frags coalescing won't work, but consider we've 1395 * disabled GSO for XDP, it won't be a big issue. 1396 */ 1397 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room); 1398 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp))) 1399 return -ENOMEM; 1400 1401 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1402 buf += headroom; /* advance address leaving hole at front of pkt */ 1403 get_page(alloc_frag->page); 1404 alloc_frag->offset += len + room; 1405 hole = alloc_frag->size - alloc_frag->offset; 1406 if (hole < len + room) { 1407 /* To avoid internal fragmentation, if there is very likely not 1408 * enough space for another buffer, add the remaining space to 1409 * the current buffer. 1410 */ 1411 len += hole; 1412 alloc_frag->offset += hole; 1413 } 1414 1415 sg_init_one(rq->sg, buf, len); 1416 ctx = mergeable_len_to_ctx(len, headroom); 1417 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1418 if (err < 0) 1419 put_page(virt_to_head_page(buf)); 1420 1421 return err; 1422 } 1423 1424 /* 1425 * Returns false if we couldn't fill entirely (OOM). 1426 * 1427 * Normally run in the receive path, but can also be run from ndo_open 1428 * before we're receiving packets, or from refill_work which is 1429 * careful to disable receiving (using napi_disable). 1430 */ 1431 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq, 1432 gfp_t gfp) 1433 { 1434 int err; 1435 bool oom; 1436 1437 do { 1438 if (vi->mergeable_rx_bufs) 1439 err = add_recvbuf_mergeable(vi, rq, gfp); 1440 else if (vi->big_packets) 1441 err = add_recvbuf_big(vi, rq, gfp); 1442 else 1443 err = add_recvbuf_small(vi, rq, gfp); 1444 1445 oom = err == -ENOMEM; 1446 if (err) 1447 break; 1448 } while (rq->vq->num_free); 1449 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) { 1450 unsigned long flags; 1451 1452 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp); 1453 rq->stats.kicks++; 1454 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags); 1455 } 1456 1457 return !oom; 1458 } 1459 1460 static void skb_recv_done(struct virtqueue *rvq) 1461 { 1462 struct virtnet_info *vi = rvq->vdev->priv; 1463 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)]; 1464 1465 virtqueue_napi_schedule(&rq->napi, rvq); 1466 } 1467 1468 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi) 1469 { 1470 napi_enable(napi); 1471 1472 /* If all buffers were filled by other side before we napi_enabled, we 1473 * won't get another interrupt, so process any outstanding packets now. 1474 * Call local_bh_enable after to trigger softIRQ processing. 1475 */ 1476 local_bh_disable(); 1477 virtqueue_napi_schedule(napi, vq); 1478 local_bh_enable(); 1479 } 1480 1481 static void virtnet_napi_tx_enable(struct virtnet_info *vi, 1482 struct virtqueue *vq, 1483 struct napi_struct *napi) 1484 { 1485 if (!napi->weight) 1486 return; 1487 1488 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only 1489 * enable the feature if this is likely affine with the transmit path. 1490 */ 1491 if (!vi->affinity_hint_set) { 1492 napi->weight = 0; 1493 return; 1494 } 1495 1496 return virtnet_napi_enable(vq, napi); 1497 } 1498 1499 static void virtnet_napi_tx_disable(struct napi_struct *napi) 1500 { 1501 if (napi->weight) 1502 napi_disable(napi); 1503 } 1504 1505 static void refill_work(struct work_struct *work) 1506 { 1507 struct virtnet_info *vi = 1508 container_of(work, struct virtnet_info, refill.work); 1509 bool still_empty; 1510 int i; 1511 1512 for (i = 0; i < vi->curr_queue_pairs; i++) { 1513 struct receive_queue *rq = &vi->rq[i]; 1514 1515 napi_disable(&rq->napi); 1516 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL); 1517 virtnet_napi_enable(rq->vq, &rq->napi); 1518 1519 /* In theory, this can happen: if we don't get any buffers in 1520 * we will *never* try to fill again. 1521 */ 1522 if (still_empty) 1523 schedule_delayed_work(&vi->refill, HZ/2); 1524 } 1525 } 1526 1527 static int virtnet_receive(struct receive_queue *rq, int budget, 1528 unsigned int *xdp_xmit) 1529 { 1530 struct virtnet_info *vi = rq->vq->vdev->priv; 1531 struct virtnet_rq_stats stats = {}; 1532 unsigned int len; 1533 void *buf; 1534 int i; 1535 1536 if (!vi->big_packets || vi->mergeable_rx_bufs) { 1537 void *ctx; 1538 1539 while (stats.packets < budget && 1540 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) { 1541 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats); 1542 stats.packets++; 1543 } 1544 } else { 1545 while (stats.packets < budget && 1546 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) { 1547 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats); 1548 stats.packets++; 1549 } 1550 } 1551 1552 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) { 1553 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) { 1554 spin_lock(&vi->refill_lock); 1555 if (vi->refill_enabled) 1556 schedule_delayed_work(&vi->refill, 0); 1557 spin_unlock(&vi->refill_lock); 1558 } 1559 } 1560 1561 u64_stats_update_begin(&rq->stats.syncp); 1562 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) { 1563 size_t offset = virtnet_rq_stats_desc[i].offset; 1564 u64 *item; 1565 1566 item = (u64 *)((u8 *)&rq->stats + offset); 1567 *item += *(u64 *)((u8 *)&stats + offset); 1568 } 1569 u64_stats_update_end(&rq->stats.syncp); 1570 1571 return stats.packets; 1572 } 1573 1574 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi) 1575 { 1576 unsigned int len; 1577 unsigned int packets = 0; 1578 unsigned int bytes = 0; 1579 void *ptr; 1580 1581 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) { 1582 if (likely(!is_xdp_frame(ptr))) { 1583 struct sk_buff *skb = ptr; 1584 1585 pr_debug("Sent skb %p\n", skb); 1586 1587 bytes += skb->len; 1588 napi_consume_skb(skb, in_napi); 1589 } else { 1590 struct xdp_frame *frame = ptr_to_xdp(ptr); 1591 1592 bytes += frame->len; 1593 xdp_return_frame(frame); 1594 } 1595 packets++; 1596 } 1597 1598 /* Avoid overhead when no packets have been processed 1599 * happens when called speculatively from start_xmit. 1600 */ 1601 if (!packets) 1602 return; 1603 1604 u64_stats_update_begin(&sq->stats.syncp); 1605 sq->stats.bytes += bytes; 1606 sq->stats.packets += packets; 1607 u64_stats_update_end(&sq->stats.syncp); 1608 } 1609 1610 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q) 1611 { 1612 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs)) 1613 return false; 1614 else if (q < vi->curr_queue_pairs) 1615 return true; 1616 else 1617 return false; 1618 } 1619 1620 static void virtnet_poll_cleantx(struct receive_queue *rq) 1621 { 1622 struct virtnet_info *vi = rq->vq->vdev->priv; 1623 unsigned int index = vq2rxq(rq->vq); 1624 struct send_queue *sq = &vi->sq[index]; 1625 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index); 1626 1627 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index)) 1628 return; 1629 1630 if (__netif_tx_trylock(txq)) { 1631 do { 1632 virtqueue_disable_cb(sq->vq); 1633 free_old_xmit_skbs(sq, true); 1634 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq))); 1635 1636 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1637 netif_tx_wake_queue(txq); 1638 1639 __netif_tx_unlock(txq); 1640 } 1641 } 1642 1643 static int virtnet_poll(struct napi_struct *napi, int budget) 1644 { 1645 struct receive_queue *rq = 1646 container_of(napi, struct receive_queue, napi); 1647 struct virtnet_info *vi = rq->vq->vdev->priv; 1648 struct send_queue *sq; 1649 unsigned int received; 1650 unsigned int xdp_xmit = 0; 1651 1652 virtnet_poll_cleantx(rq); 1653 1654 received = virtnet_receive(rq, budget, &xdp_xmit); 1655 1656 /* Out of packets? */ 1657 if (received < budget) 1658 virtqueue_napi_complete(napi, rq->vq, received); 1659 1660 if (xdp_xmit & VIRTIO_XDP_REDIR) 1661 xdp_do_flush(); 1662 1663 if (xdp_xmit & VIRTIO_XDP_TX) { 1664 sq = virtnet_xdp_get_sq(vi); 1665 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) { 1666 u64_stats_update_begin(&sq->stats.syncp); 1667 sq->stats.kicks++; 1668 u64_stats_update_end(&sq->stats.syncp); 1669 } 1670 virtnet_xdp_put_sq(vi, sq); 1671 } 1672 1673 return received; 1674 } 1675 1676 static int virtnet_open(struct net_device *dev) 1677 { 1678 struct virtnet_info *vi = netdev_priv(dev); 1679 int i, err; 1680 1681 enable_delayed_refill(vi); 1682 1683 for (i = 0; i < vi->max_queue_pairs; i++) { 1684 if (i < vi->curr_queue_pairs) 1685 /* Make sure we have some buffers: if oom use wq. */ 1686 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 1687 schedule_delayed_work(&vi->refill, 0); 1688 1689 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id); 1690 if (err < 0) 1691 return err; 1692 1693 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq, 1694 MEM_TYPE_PAGE_SHARED, NULL); 1695 if (err < 0) { 1696 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 1697 return err; 1698 } 1699 1700 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 1701 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi); 1702 } 1703 1704 return 0; 1705 } 1706 1707 static int virtnet_poll_tx(struct napi_struct *napi, int budget) 1708 { 1709 struct send_queue *sq = container_of(napi, struct send_queue, napi); 1710 struct virtnet_info *vi = sq->vq->vdev->priv; 1711 unsigned int index = vq2txq(sq->vq); 1712 struct netdev_queue *txq; 1713 int opaque; 1714 bool done; 1715 1716 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) { 1717 /* We don't need to enable cb for XDP */ 1718 napi_complete_done(napi, 0); 1719 return 0; 1720 } 1721 1722 txq = netdev_get_tx_queue(vi->dev, index); 1723 __netif_tx_lock(txq, raw_smp_processor_id()); 1724 virtqueue_disable_cb(sq->vq); 1725 free_old_xmit_skbs(sq, true); 1726 1727 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1728 netif_tx_wake_queue(txq); 1729 1730 opaque = virtqueue_enable_cb_prepare(sq->vq); 1731 1732 done = napi_complete_done(napi, 0); 1733 1734 if (!done) 1735 virtqueue_disable_cb(sq->vq); 1736 1737 __netif_tx_unlock(txq); 1738 1739 if (done) { 1740 if (unlikely(virtqueue_poll(sq->vq, opaque))) { 1741 if (napi_schedule_prep(napi)) { 1742 __netif_tx_lock(txq, raw_smp_processor_id()); 1743 virtqueue_disable_cb(sq->vq); 1744 __netif_tx_unlock(txq); 1745 __napi_schedule(napi); 1746 } 1747 } 1748 } 1749 1750 return 0; 1751 } 1752 1753 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb) 1754 { 1755 struct virtio_net_hdr_mrg_rxbuf *hdr; 1756 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest; 1757 struct virtnet_info *vi = sq->vq->vdev->priv; 1758 int num_sg; 1759 unsigned hdr_len = vi->hdr_len; 1760 bool can_push; 1761 1762 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest); 1763 1764 can_push = vi->any_header_sg && 1765 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) && 1766 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len; 1767 /* Even if we can, don't push here yet as this would skew 1768 * csum_start offset below. */ 1769 if (can_push) 1770 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len); 1771 else 1772 hdr = skb_vnet_hdr(skb); 1773 1774 if (virtio_net_hdr_from_skb(skb, &hdr->hdr, 1775 virtio_is_little_endian(vi->vdev), false, 1776 0)) 1777 return -EPROTO; 1778 1779 if (vi->mergeable_rx_bufs) 1780 hdr->num_buffers = 0; 1781 1782 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2)); 1783 if (can_push) { 1784 __skb_push(skb, hdr_len); 1785 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len); 1786 if (unlikely(num_sg < 0)) 1787 return num_sg; 1788 /* Pull header back to avoid skew in tx bytes calculations. */ 1789 __skb_pull(skb, hdr_len); 1790 } else { 1791 sg_set_buf(sq->sg, hdr, hdr_len); 1792 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len); 1793 if (unlikely(num_sg < 0)) 1794 return num_sg; 1795 num_sg++; 1796 } 1797 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC); 1798 } 1799 1800 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) 1801 { 1802 struct virtnet_info *vi = netdev_priv(dev); 1803 int qnum = skb_get_queue_mapping(skb); 1804 struct send_queue *sq = &vi->sq[qnum]; 1805 int err; 1806 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum); 1807 bool kick = !netdev_xmit_more(); 1808 bool use_napi = sq->napi.weight; 1809 1810 /* Free up any pending old buffers before queueing new ones. */ 1811 do { 1812 if (use_napi) 1813 virtqueue_disable_cb(sq->vq); 1814 1815 free_old_xmit_skbs(sq, false); 1816 1817 } while (use_napi && kick && 1818 unlikely(!virtqueue_enable_cb_delayed(sq->vq))); 1819 1820 /* timestamp packet in software */ 1821 skb_tx_timestamp(skb); 1822 1823 /* Try to transmit */ 1824 err = xmit_skb(sq, skb); 1825 1826 /* This should not happen! */ 1827 if (unlikely(err)) { 1828 dev->stats.tx_fifo_errors++; 1829 if (net_ratelimit()) 1830 dev_warn(&dev->dev, 1831 "Unexpected TXQ (%d) queue failure: %d\n", 1832 qnum, err); 1833 dev->stats.tx_dropped++; 1834 dev_kfree_skb_any(skb); 1835 return NETDEV_TX_OK; 1836 } 1837 1838 /* Don't wait up for transmitted skbs to be freed. */ 1839 if (!use_napi) { 1840 skb_orphan(skb); 1841 nf_reset_ct(skb); 1842 } 1843 1844 /* If running out of space, stop queue to avoid getting packets that we 1845 * are then unable to transmit. 1846 * An alternative would be to force queuing layer to requeue the skb by 1847 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be 1848 * returned in a normal path of operation: it means that driver is not 1849 * maintaining the TX queue stop/start state properly, and causes 1850 * the stack to do a non-trivial amount of useless work. 1851 * Since most packets only take 1 or 2 ring slots, stopping the queue 1852 * early means 16 slots are typically wasted. 1853 */ 1854 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) { 1855 netif_stop_subqueue(dev, qnum); 1856 if (!use_napi && 1857 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) { 1858 /* More just got used, free them then recheck. */ 1859 free_old_xmit_skbs(sq, false); 1860 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) { 1861 netif_start_subqueue(dev, qnum); 1862 virtqueue_disable_cb(sq->vq); 1863 } 1864 } 1865 } 1866 1867 if (kick || netif_xmit_stopped(txq)) { 1868 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) { 1869 u64_stats_update_begin(&sq->stats.syncp); 1870 sq->stats.kicks++; 1871 u64_stats_update_end(&sq->stats.syncp); 1872 } 1873 } 1874 1875 return NETDEV_TX_OK; 1876 } 1877 1878 /* 1879 * Send command via the control virtqueue and check status. Commands 1880 * supported by the hypervisor, as indicated by feature bits, should 1881 * never fail unless improperly formatted. 1882 */ 1883 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd, 1884 struct scatterlist *out) 1885 { 1886 struct scatterlist *sgs[4], hdr, stat; 1887 unsigned out_num = 0, tmp; 1888 int ret; 1889 1890 /* Caller should know better */ 1891 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)); 1892 1893 vi->ctrl->status = ~0; 1894 vi->ctrl->hdr.class = class; 1895 vi->ctrl->hdr.cmd = cmd; 1896 /* Add header */ 1897 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr)); 1898 sgs[out_num++] = &hdr; 1899 1900 if (out) 1901 sgs[out_num++] = out; 1902 1903 /* Add return status. */ 1904 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status)); 1905 sgs[out_num] = &stat; 1906 1907 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs)); 1908 ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC); 1909 if (ret < 0) { 1910 dev_warn(&vi->vdev->dev, 1911 "Failed to add sgs for command vq: %d\n.", ret); 1912 return false; 1913 } 1914 1915 if (unlikely(!virtqueue_kick(vi->cvq))) 1916 return vi->ctrl->status == VIRTIO_NET_OK; 1917 1918 /* Spin for a response, the kick causes an ioport write, trapping 1919 * into the hypervisor, so the request should be handled immediately. 1920 */ 1921 while (!virtqueue_get_buf(vi->cvq, &tmp) && 1922 !virtqueue_is_broken(vi->cvq)) 1923 cpu_relax(); 1924 1925 return vi->ctrl->status == VIRTIO_NET_OK; 1926 } 1927 1928 static int virtnet_set_mac_address(struct net_device *dev, void *p) 1929 { 1930 struct virtnet_info *vi = netdev_priv(dev); 1931 struct virtio_device *vdev = vi->vdev; 1932 int ret; 1933 struct sockaddr *addr; 1934 struct scatterlist sg; 1935 1936 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY)) 1937 return -EOPNOTSUPP; 1938 1939 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL); 1940 if (!addr) 1941 return -ENOMEM; 1942 1943 ret = eth_prepare_mac_addr_change(dev, addr); 1944 if (ret) 1945 goto out; 1946 1947 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 1948 sg_init_one(&sg, addr->sa_data, dev->addr_len); 1949 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 1950 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) { 1951 dev_warn(&vdev->dev, 1952 "Failed to set mac address by vq command.\n"); 1953 ret = -EINVAL; 1954 goto out; 1955 } 1956 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) && 1957 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) { 1958 unsigned int i; 1959 1960 /* Naturally, this has an atomicity problem. */ 1961 for (i = 0; i < dev->addr_len; i++) 1962 virtio_cwrite8(vdev, 1963 offsetof(struct virtio_net_config, mac) + 1964 i, addr->sa_data[i]); 1965 } 1966 1967 eth_commit_mac_addr_change(dev, p); 1968 ret = 0; 1969 1970 out: 1971 kfree(addr); 1972 return ret; 1973 } 1974 1975 static void virtnet_stats(struct net_device *dev, 1976 struct rtnl_link_stats64 *tot) 1977 { 1978 struct virtnet_info *vi = netdev_priv(dev); 1979 unsigned int start; 1980 int i; 1981 1982 for (i = 0; i < vi->max_queue_pairs; i++) { 1983 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops; 1984 struct receive_queue *rq = &vi->rq[i]; 1985 struct send_queue *sq = &vi->sq[i]; 1986 1987 do { 1988 start = u64_stats_fetch_begin_irq(&sq->stats.syncp); 1989 tpackets = sq->stats.packets; 1990 tbytes = sq->stats.bytes; 1991 terrors = sq->stats.tx_timeouts; 1992 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start)); 1993 1994 do { 1995 start = u64_stats_fetch_begin_irq(&rq->stats.syncp); 1996 rpackets = rq->stats.packets; 1997 rbytes = rq->stats.bytes; 1998 rdrops = rq->stats.drops; 1999 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start)); 2000 2001 tot->rx_packets += rpackets; 2002 tot->tx_packets += tpackets; 2003 tot->rx_bytes += rbytes; 2004 tot->tx_bytes += tbytes; 2005 tot->rx_dropped += rdrops; 2006 tot->tx_errors += terrors; 2007 } 2008 2009 tot->tx_dropped = dev->stats.tx_dropped; 2010 tot->tx_fifo_errors = dev->stats.tx_fifo_errors; 2011 tot->rx_length_errors = dev->stats.rx_length_errors; 2012 tot->rx_frame_errors = dev->stats.rx_frame_errors; 2013 } 2014 2015 static void virtnet_ack_link_announce(struct virtnet_info *vi) 2016 { 2017 rtnl_lock(); 2018 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE, 2019 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL)) 2020 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n"); 2021 rtnl_unlock(); 2022 } 2023 2024 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 2025 { 2026 struct scatterlist sg; 2027 struct net_device *dev = vi->dev; 2028 2029 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ)) 2030 return 0; 2031 2032 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs); 2033 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq)); 2034 2035 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ, 2036 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) { 2037 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n", 2038 queue_pairs); 2039 return -EINVAL; 2040 } else { 2041 vi->curr_queue_pairs = queue_pairs; 2042 /* virtnet_open() will refill when device is going to up. */ 2043 if (dev->flags & IFF_UP) 2044 schedule_delayed_work(&vi->refill, 0); 2045 } 2046 2047 return 0; 2048 } 2049 2050 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 2051 { 2052 int err; 2053 2054 rtnl_lock(); 2055 err = _virtnet_set_queues(vi, queue_pairs); 2056 rtnl_unlock(); 2057 return err; 2058 } 2059 2060 static int virtnet_close(struct net_device *dev) 2061 { 2062 struct virtnet_info *vi = netdev_priv(dev); 2063 int i; 2064 2065 /* Make sure NAPI doesn't schedule refill work */ 2066 disable_delayed_refill(vi); 2067 /* Make sure refill_work doesn't re-enable napi! */ 2068 cancel_delayed_work_sync(&vi->refill); 2069 2070 for (i = 0; i < vi->max_queue_pairs; i++) { 2071 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 2072 napi_disable(&vi->rq[i].napi); 2073 virtnet_napi_tx_disable(&vi->sq[i].napi); 2074 } 2075 2076 return 0; 2077 } 2078 2079 static void virtnet_set_rx_mode(struct net_device *dev) 2080 { 2081 struct virtnet_info *vi = netdev_priv(dev); 2082 struct scatterlist sg[2]; 2083 struct virtio_net_ctrl_mac *mac_data; 2084 struct netdev_hw_addr *ha; 2085 int uc_count; 2086 int mc_count; 2087 void *buf; 2088 int i; 2089 2090 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */ 2091 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX)) 2092 return; 2093 2094 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0); 2095 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0); 2096 2097 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc)); 2098 2099 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 2100 VIRTIO_NET_CTRL_RX_PROMISC, sg)) 2101 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n", 2102 vi->ctrl->promisc ? "en" : "dis"); 2103 2104 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti)); 2105 2106 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 2107 VIRTIO_NET_CTRL_RX_ALLMULTI, sg)) 2108 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n", 2109 vi->ctrl->allmulti ? "en" : "dis"); 2110 2111 uc_count = netdev_uc_count(dev); 2112 mc_count = netdev_mc_count(dev); 2113 /* MAC filter - use one buffer for both lists */ 2114 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) + 2115 (2 * sizeof(mac_data->entries)), GFP_ATOMIC); 2116 mac_data = buf; 2117 if (!buf) 2118 return; 2119 2120 sg_init_table(sg, 2); 2121 2122 /* Store the unicast list and count in the front of the buffer */ 2123 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count); 2124 i = 0; 2125 netdev_for_each_uc_addr(ha, dev) 2126 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 2127 2128 sg_set_buf(&sg[0], mac_data, 2129 sizeof(mac_data->entries) + (uc_count * ETH_ALEN)); 2130 2131 /* multicast list and count fill the end */ 2132 mac_data = (void *)&mac_data->macs[uc_count][0]; 2133 2134 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count); 2135 i = 0; 2136 netdev_for_each_mc_addr(ha, dev) 2137 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 2138 2139 sg_set_buf(&sg[1], mac_data, 2140 sizeof(mac_data->entries) + (mc_count * ETH_ALEN)); 2141 2142 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 2143 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg)) 2144 dev_warn(&dev->dev, "Failed to set MAC filter table.\n"); 2145 2146 kfree(buf); 2147 } 2148 2149 static int virtnet_vlan_rx_add_vid(struct net_device *dev, 2150 __be16 proto, u16 vid) 2151 { 2152 struct virtnet_info *vi = netdev_priv(dev); 2153 struct scatterlist sg; 2154 2155 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid); 2156 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid)); 2157 2158 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 2159 VIRTIO_NET_CTRL_VLAN_ADD, &sg)) 2160 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid); 2161 return 0; 2162 } 2163 2164 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, 2165 __be16 proto, u16 vid) 2166 { 2167 struct virtnet_info *vi = netdev_priv(dev); 2168 struct scatterlist sg; 2169 2170 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid); 2171 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid)); 2172 2173 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 2174 VIRTIO_NET_CTRL_VLAN_DEL, &sg)) 2175 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid); 2176 return 0; 2177 } 2178 2179 static void virtnet_clean_affinity(struct virtnet_info *vi) 2180 { 2181 int i; 2182 2183 if (vi->affinity_hint_set) { 2184 for (i = 0; i < vi->max_queue_pairs; i++) { 2185 virtqueue_set_affinity(vi->rq[i].vq, NULL); 2186 virtqueue_set_affinity(vi->sq[i].vq, NULL); 2187 } 2188 2189 vi->affinity_hint_set = false; 2190 } 2191 } 2192 2193 static void virtnet_set_affinity(struct virtnet_info *vi) 2194 { 2195 cpumask_var_t mask; 2196 int stragglers; 2197 int group_size; 2198 int i, j, cpu; 2199 int num_cpu; 2200 int stride; 2201 2202 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) { 2203 virtnet_clean_affinity(vi); 2204 return; 2205 } 2206 2207 num_cpu = num_online_cpus(); 2208 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1); 2209 stragglers = num_cpu >= vi->curr_queue_pairs ? 2210 num_cpu % vi->curr_queue_pairs : 2211 0; 2212 cpu = cpumask_first(cpu_online_mask); 2213 2214 for (i = 0; i < vi->curr_queue_pairs; i++) { 2215 group_size = stride + (i < stragglers ? 1 : 0); 2216 2217 for (j = 0; j < group_size; j++) { 2218 cpumask_set_cpu(cpu, mask); 2219 cpu = cpumask_next_wrap(cpu, cpu_online_mask, 2220 nr_cpu_ids, false); 2221 } 2222 virtqueue_set_affinity(vi->rq[i].vq, mask); 2223 virtqueue_set_affinity(vi->sq[i].vq, mask); 2224 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS); 2225 cpumask_clear(mask); 2226 } 2227 2228 vi->affinity_hint_set = true; 2229 free_cpumask_var(mask); 2230 } 2231 2232 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node) 2233 { 2234 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 2235 node); 2236 virtnet_set_affinity(vi); 2237 return 0; 2238 } 2239 2240 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node) 2241 { 2242 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 2243 node_dead); 2244 virtnet_set_affinity(vi); 2245 return 0; 2246 } 2247 2248 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node) 2249 { 2250 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 2251 node); 2252 2253 virtnet_clean_affinity(vi); 2254 return 0; 2255 } 2256 2257 static enum cpuhp_state virtionet_online; 2258 2259 static int virtnet_cpu_notif_add(struct virtnet_info *vi) 2260 { 2261 int ret; 2262 2263 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node); 2264 if (ret) 2265 return ret; 2266 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD, 2267 &vi->node_dead); 2268 if (!ret) 2269 return ret; 2270 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 2271 return ret; 2272 } 2273 2274 static void virtnet_cpu_notif_remove(struct virtnet_info *vi) 2275 { 2276 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 2277 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD, 2278 &vi->node_dead); 2279 } 2280 2281 static void virtnet_get_ringparam(struct net_device *dev, 2282 struct ethtool_ringparam *ring, 2283 struct kernel_ethtool_ringparam *kernel_ring, 2284 struct netlink_ext_ack *extack) 2285 { 2286 struct virtnet_info *vi = netdev_priv(dev); 2287 2288 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq); 2289 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq); 2290 ring->rx_pending = ring->rx_max_pending; 2291 ring->tx_pending = ring->tx_max_pending; 2292 } 2293 2294 static bool virtnet_commit_rss_command(struct virtnet_info *vi) 2295 { 2296 struct net_device *dev = vi->dev; 2297 struct scatterlist sgs[4]; 2298 unsigned int sg_buf_size; 2299 2300 /* prepare sgs */ 2301 sg_init_table(sgs, 4); 2302 2303 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table); 2304 sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size); 2305 2306 sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1); 2307 sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size); 2308 2309 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key) 2310 - offsetof(struct virtio_net_ctrl_rss, max_tx_vq); 2311 sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size); 2312 2313 sg_buf_size = vi->rss_key_size; 2314 sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size); 2315 2316 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ, 2317 vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG 2318 : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) { 2319 dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n"); 2320 return false; 2321 } 2322 return true; 2323 } 2324 2325 static void virtnet_init_default_rss(struct virtnet_info *vi) 2326 { 2327 u32 indir_val = 0; 2328 int i = 0; 2329 2330 vi->ctrl->rss.hash_types = vi->rss_hash_types_supported; 2331 vi->rss_hash_types_saved = vi->rss_hash_types_supported; 2332 vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size 2333 ? vi->rss_indir_table_size - 1 : 0; 2334 vi->ctrl->rss.unclassified_queue = 0; 2335 2336 for (; i < vi->rss_indir_table_size; ++i) { 2337 indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs); 2338 vi->ctrl->rss.indirection_table[i] = indir_val; 2339 } 2340 2341 vi->ctrl->rss.max_tx_vq = vi->curr_queue_pairs; 2342 vi->ctrl->rss.hash_key_length = vi->rss_key_size; 2343 2344 netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size); 2345 } 2346 2347 static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info) 2348 { 2349 info->data = 0; 2350 switch (info->flow_type) { 2351 case TCP_V4_FLOW: 2352 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) { 2353 info->data = RXH_IP_SRC | RXH_IP_DST | 2354 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2355 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) { 2356 info->data = RXH_IP_SRC | RXH_IP_DST; 2357 } 2358 break; 2359 case TCP_V6_FLOW: 2360 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) { 2361 info->data = RXH_IP_SRC | RXH_IP_DST | 2362 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2363 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) { 2364 info->data = RXH_IP_SRC | RXH_IP_DST; 2365 } 2366 break; 2367 case UDP_V4_FLOW: 2368 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) { 2369 info->data = RXH_IP_SRC | RXH_IP_DST | 2370 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2371 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) { 2372 info->data = RXH_IP_SRC | RXH_IP_DST; 2373 } 2374 break; 2375 case UDP_V6_FLOW: 2376 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) { 2377 info->data = RXH_IP_SRC | RXH_IP_DST | 2378 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2379 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) { 2380 info->data = RXH_IP_SRC | RXH_IP_DST; 2381 } 2382 break; 2383 case IPV4_FLOW: 2384 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) 2385 info->data = RXH_IP_SRC | RXH_IP_DST; 2386 2387 break; 2388 case IPV6_FLOW: 2389 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) 2390 info->data = RXH_IP_SRC | RXH_IP_DST; 2391 2392 break; 2393 default: 2394 info->data = 0; 2395 break; 2396 } 2397 } 2398 2399 static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info) 2400 { 2401 u32 new_hashtypes = vi->rss_hash_types_saved; 2402 bool is_disable = info->data & RXH_DISCARD; 2403 bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3); 2404 2405 /* supports only 'sd', 'sdfn' and 'r' */ 2406 if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable)) 2407 return false; 2408 2409 switch (info->flow_type) { 2410 case TCP_V4_FLOW: 2411 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4); 2412 if (!is_disable) 2413 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4 2414 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0); 2415 break; 2416 case UDP_V4_FLOW: 2417 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4); 2418 if (!is_disable) 2419 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4 2420 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0); 2421 break; 2422 case IPV4_FLOW: 2423 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4; 2424 if (!is_disable) 2425 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4; 2426 break; 2427 case TCP_V6_FLOW: 2428 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6); 2429 if (!is_disable) 2430 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6 2431 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0); 2432 break; 2433 case UDP_V6_FLOW: 2434 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6); 2435 if (!is_disable) 2436 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6 2437 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0); 2438 break; 2439 case IPV6_FLOW: 2440 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6; 2441 if (!is_disable) 2442 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6; 2443 break; 2444 default: 2445 /* unsupported flow */ 2446 return false; 2447 } 2448 2449 /* if unsupported hashtype was set */ 2450 if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported)) 2451 return false; 2452 2453 if (new_hashtypes != vi->rss_hash_types_saved) { 2454 vi->rss_hash_types_saved = new_hashtypes; 2455 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved; 2456 if (vi->dev->features & NETIF_F_RXHASH) 2457 return virtnet_commit_rss_command(vi); 2458 } 2459 2460 return true; 2461 } 2462 2463 static void virtnet_get_drvinfo(struct net_device *dev, 2464 struct ethtool_drvinfo *info) 2465 { 2466 struct virtnet_info *vi = netdev_priv(dev); 2467 struct virtio_device *vdev = vi->vdev; 2468 2469 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver)); 2470 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version)); 2471 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info)); 2472 2473 } 2474 2475 /* TODO: Eliminate OOO packets during switching */ 2476 static int virtnet_set_channels(struct net_device *dev, 2477 struct ethtool_channels *channels) 2478 { 2479 struct virtnet_info *vi = netdev_priv(dev); 2480 u16 queue_pairs = channels->combined_count; 2481 int err; 2482 2483 /* We don't support separate rx/tx channels. 2484 * We don't allow setting 'other' channels. 2485 */ 2486 if (channels->rx_count || channels->tx_count || channels->other_count) 2487 return -EINVAL; 2488 2489 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0) 2490 return -EINVAL; 2491 2492 /* For now we don't support modifying channels while XDP is loaded 2493 * also when XDP is loaded all RX queues have XDP programs so we only 2494 * need to check a single RX queue. 2495 */ 2496 if (vi->rq[0].xdp_prog) 2497 return -EINVAL; 2498 2499 cpus_read_lock(); 2500 err = _virtnet_set_queues(vi, queue_pairs); 2501 if (err) { 2502 cpus_read_unlock(); 2503 goto err; 2504 } 2505 virtnet_set_affinity(vi); 2506 cpus_read_unlock(); 2507 2508 netif_set_real_num_tx_queues(dev, queue_pairs); 2509 netif_set_real_num_rx_queues(dev, queue_pairs); 2510 err: 2511 return err; 2512 } 2513 2514 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data) 2515 { 2516 struct virtnet_info *vi = netdev_priv(dev); 2517 unsigned int i, j; 2518 u8 *p = data; 2519 2520 switch (stringset) { 2521 case ETH_SS_STATS: 2522 for (i = 0; i < vi->curr_queue_pairs; i++) { 2523 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) 2524 ethtool_sprintf(&p, "rx_queue_%u_%s", i, 2525 virtnet_rq_stats_desc[j].desc); 2526 } 2527 2528 for (i = 0; i < vi->curr_queue_pairs; i++) { 2529 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) 2530 ethtool_sprintf(&p, "tx_queue_%u_%s", i, 2531 virtnet_sq_stats_desc[j].desc); 2532 } 2533 break; 2534 } 2535 } 2536 2537 static int virtnet_get_sset_count(struct net_device *dev, int sset) 2538 { 2539 struct virtnet_info *vi = netdev_priv(dev); 2540 2541 switch (sset) { 2542 case ETH_SS_STATS: 2543 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN + 2544 VIRTNET_SQ_STATS_LEN); 2545 default: 2546 return -EOPNOTSUPP; 2547 } 2548 } 2549 2550 static void virtnet_get_ethtool_stats(struct net_device *dev, 2551 struct ethtool_stats *stats, u64 *data) 2552 { 2553 struct virtnet_info *vi = netdev_priv(dev); 2554 unsigned int idx = 0, start, i, j; 2555 const u8 *stats_base; 2556 size_t offset; 2557 2558 for (i = 0; i < vi->curr_queue_pairs; i++) { 2559 struct receive_queue *rq = &vi->rq[i]; 2560 2561 stats_base = (u8 *)&rq->stats; 2562 do { 2563 start = u64_stats_fetch_begin_irq(&rq->stats.syncp); 2564 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 2565 offset = virtnet_rq_stats_desc[j].offset; 2566 data[idx + j] = *(u64 *)(stats_base + offset); 2567 } 2568 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start)); 2569 idx += VIRTNET_RQ_STATS_LEN; 2570 } 2571 2572 for (i = 0; i < vi->curr_queue_pairs; i++) { 2573 struct send_queue *sq = &vi->sq[i]; 2574 2575 stats_base = (u8 *)&sq->stats; 2576 do { 2577 start = u64_stats_fetch_begin_irq(&sq->stats.syncp); 2578 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 2579 offset = virtnet_sq_stats_desc[j].offset; 2580 data[idx + j] = *(u64 *)(stats_base + offset); 2581 } 2582 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start)); 2583 idx += VIRTNET_SQ_STATS_LEN; 2584 } 2585 } 2586 2587 static void virtnet_get_channels(struct net_device *dev, 2588 struct ethtool_channels *channels) 2589 { 2590 struct virtnet_info *vi = netdev_priv(dev); 2591 2592 channels->combined_count = vi->curr_queue_pairs; 2593 channels->max_combined = vi->max_queue_pairs; 2594 channels->max_other = 0; 2595 channels->rx_count = 0; 2596 channels->tx_count = 0; 2597 channels->other_count = 0; 2598 } 2599 2600 static int virtnet_set_link_ksettings(struct net_device *dev, 2601 const struct ethtool_link_ksettings *cmd) 2602 { 2603 struct virtnet_info *vi = netdev_priv(dev); 2604 2605 return ethtool_virtdev_set_link_ksettings(dev, cmd, 2606 &vi->speed, &vi->duplex); 2607 } 2608 2609 static int virtnet_get_link_ksettings(struct net_device *dev, 2610 struct ethtool_link_ksettings *cmd) 2611 { 2612 struct virtnet_info *vi = netdev_priv(dev); 2613 2614 cmd->base.speed = vi->speed; 2615 cmd->base.duplex = vi->duplex; 2616 cmd->base.port = PORT_OTHER; 2617 2618 return 0; 2619 } 2620 2621 static int virtnet_set_coalesce(struct net_device *dev, 2622 struct ethtool_coalesce *ec, 2623 struct kernel_ethtool_coalesce *kernel_coal, 2624 struct netlink_ext_ack *extack) 2625 { 2626 struct virtnet_info *vi = netdev_priv(dev); 2627 int i, napi_weight; 2628 2629 if (ec->tx_max_coalesced_frames > 1 || 2630 ec->rx_max_coalesced_frames != 1) 2631 return -EINVAL; 2632 2633 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0; 2634 if (napi_weight ^ vi->sq[0].napi.weight) { 2635 if (dev->flags & IFF_UP) 2636 return -EBUSY; 2637 for (i = 0; i < vi->max_queue_pairs; i++) 2638 vi->sq[i].napi.weight = napi_weight; 2639 } 2640 2641 return 0; 2642 } 2643 2644 static int virtnet_get_coalesce(struct net_device *dev, 2645 struct ethtool_coalesce *ec, 2646 struct kernel_ethtool_coalesce *kernel_coal, 2647 struct netlink_ext_ack *extack) 2648 { 2649 struct ethtool_coalesce ec_default = { 2650 .cmd = ETHTOOL_GCOALESCE, 2651 .rx_max_coalesced_frames = 1, 2652 }; 2653 struct virtnet_info *vi = netdev_priv(dev); 2654 2655 memcpy(ec, &ec_default, sizeof(ec_default)); 2656 2657 if (vi->sq[0].napi.weight) 2658 ec->tx_max_coalesced_frames = 1; 2659 2660 return 0; 2661 } 2662 2663 static void virtnet_init_settings(struct net_device *dev) 2664 { 2665 struct virtnet_info *vi = netdev_priv(dev); 2666 2667 vi->speed = SPEED_UNKNOWN; 2668 vi->duplex = DUPLEX_UNKNOWN; 2669 } 2670 2671 static void virtnet_update_settings(struct virtnet_info *vi) 2672 { 2673 u32 speed; 2674 u8 duplex; 2675 2676 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX)) 2677 return; 2678 2679 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed); 2680 2681 if (ethtool_validate_speed(speed)) 2682 vi->speed = speed; 2683 2684 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex); 2685 2686 if (ethtool_validate_duplex(duplex)) 2687 vi->duplex = duplex; 2688 } 2689 2690 static u32 virtnet_get_rxfh_key_size(struct net_device *dev) 2691 { 2692 return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size; 2693 } 2694 2695 static u32 virtnet_get_rxfh_indir_size(struct net_device *dev) 2696 { 2697 return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size; 2698 } 2699 2700 static int virtnet_get_rxfh(struct net_device *dev, u32 *indir, u8 *key, u8 *hfunc) 2701 { 2702 struct virtnet_info *vi = netdev_priv(dev); 2703 int i; 2704 2705 if (indir) { 2706 for (i = 0; i < vi->rss_indir_table_size; ++i) 2707 indir[i] = vi->ctrl->rss.indirection_table[i]; 2708 } 2709 2710 if (key) 2711 memcpy(key, vi->ctrl->rss.key, vi->rss_key_size); 2712 2713 if (hfunc) 2714 *hfunc = ETH_RSS_HASH_TOP; 2715 2716 return 0; 2717 } 2718 2719 static int virtnet_set_rxfh(struct net_device *dev, const u32 *indir, const u8 *key, const u8 hfunc) 2720 { 2721 struct virtnet_info *vi = netdev_priv(dev); 2722 int i; 2723 2724 if (hfunc != ETH_RSS_HASH_NO_CHANGE && hfunc != ETH_RSS_HASH_TOP) 2725 return -EOPNOTSUPP; 2726 2727 if (indir) { 2728 for (i = 0; i < vi->rss_indir_table_size; ++i) 2729 vi->ctrl->rss.indirection_table[i] = indir[i]; 2730 } 2731 if (key) 2732 memcpy(vi->ctrl->rss.key, key, vi->rss_key_size); 2733 2734 virtnet_commit_rss_command(vi); 2735 2736 return 0; 2737 } 2738 2739 static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs) 2740 { 2741 struct virtnet_info *vi = netdev_priv(dev); 2742 int rc = 0; 2743 2744 switch (info->cmd) { 2745 case ETHTOOL_GRXRINGS: 2746 info->data = vi->curr_queue_pairs; 2747 break; 2748 case ETHTOOL_GRXFH: 2749 virtnet_get_hashflow(vi, info); 2750 break; 2751 default: 2752 rc = -EOPNOTSUPP; 2753 } 2754 2755 return rc; 2756 } 2757 2758 static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info) 2759 { 2760 struct virtnet_info *vi = netdev_priv(dev); 2761 int rc = 0; 2762 2763 switch (info->cmd) { 2764 case ETHTOOL_SRXFH: 2765 if (!virtnet_set_hashflow(vi, info)) 2766 rc = -EINVAL; 2767 2768 break; 2769 default: 2770 rc = -EOPNOTSUPP; 2771 } 2772 2773 return rc; 2774 } 2775 2776 static const struct ethtool_ops virtnet_ethtool_ops = { 2777 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES, 2778 .get_drvinfo = virtnet_get_drvinfo, 2779 .get_link = ethtool_op_get_link, 2780 .get_ringparam = virtnet_get_ringparam, 2781 .get_strings = virtnet_get_strings, 2782 .get_sset_count = virtnet_get_sset_count, 2783 .get_ethtool_stats = virtnet_get_ethtool_stats, 2784 .set_channels = virtnet_set_channels, 2785 .get_channels = virtnet_get_channels, 2786 .get_ts_info = ethtool_op_get_ts_info, 2787 .get_link_ksettings = virtnet_get_link_ksettings, 2788 .set_link_ksettings = virtnet_set_link_ksettings, 2789 .set_coalesce = virtnet_set_coalesce, 2790 .get_coalesce = virtnet_get_coalesce, 2791 .get_rxfh_key_size = virtnet_get_rxfh_key_size, 2792 .get_rxfh_indir_size = virtnet_get_rxfh_indir_size, 2793 .get_rxfh = virtnet_get_rxfh, 2794 .set_rxfh = virtnet_set_rxfh, 2795 .get_rxnfc = virtnet_get_rxnfc, 2796 .set_rxnfc = virtnet_set_rxnfc, 2797 }; 2798 2799 static void virtnet_freeze_down(struct virtio_device *vdev) 2800 { 2801 struct virtnet_info *vi = vdev->priv; 2802 2803 /* Make sure no work handler is accessing the device */ 2804 flush_work(&vi->config_work); 2805 2806 netif_tx_lock_bh(vi->dev); 2807 netif_device_detach(vi->dev); 2808 netif_tx_unlock_bh(vi->dev); 2809 if (netif_running(vi->dev)) 2810 virtnet_close(vi->dev); 2811 } 2812 2813 static int init_vqs(struct virtnet_info *vi); 2814 2815 static int virtnet_restore_up(struct virtio_device *vdev) 2816 { 2817 struct virtnet_info *vi = vdev->priv; 2818 int err; 2819 2820 err = init_vqs(vi); 2821 if (err) 2822 return err; 2823 2824 virtio_device_ready(vdev); 2825 2826 enable_delayed_refill(vi); 2827 2828 if (netif_running(vi->dev)) { 2829 err = virtnet_open(vi->dev); 2830 if (err) 2831 return err; 2832 } 2833 2834 netif_tx_lock_bh(vi->dev); 2835 netif_device_attach(vi->dev); 2836 netif_tx_unlock_bh(vi->dev); 2837 return err; 2838 } 2839 2840 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads) 2841 { 2842 struct scatterlist sg; 2843 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads); 2844 2845 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads)); 2846 2847 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS, 2848 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) { 2849 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n"); 2850 return -EINVAL; 2851 } 2852 2853 return 0; 2854 } 2855 2856 static int virtnet_clear_guest_offloads(struct virtnet_info *vi) 2857 { 2858 u64 offloads = 0; 2859 2860 if (!vi->guest_offloads) 2861 return 0; 2862 2863 return virtnet_set_guest_offloads(vi, offloads); 2864 } 2865 2866 static int virtnet_restore_guest_offloads(struct virtnet_info *vi) 2867 { 2868 u64 offloads = vi->guest_offloads; 2869 2870 if (!vi->guest_offloads) 2871 return 0; 2872 2873 return virtnet_set_guest_offloads(vi, offloads); 2874 } 2875 2876 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog, 2877 struct netlink_ext_ack *extack) 2878 { 2879 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr); 2880 struct virtnet_info *vi = netdev_priv(dev); 2881 struct bpf_prog *old_prog; 2882 u16 xdp_qp = 0, curr_qp; 2883 int i, err; 2884 2885 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) 2886 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) || 2887 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) || 2888 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) || 2889 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) || 2890 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) { 2891 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first"); 2892 return -EOPNOTSUPP; 2893 } 2894 2895 if (vi->mergeable_rx_bufs && !vi->any_header_sg) { 2896 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required"); 2897 return -EINVAL; 2898 } 2899 2900 if (dev->mtu > max_sz) { 2901 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP"); 2902 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz); 2903 return -EINVAL; 2904 } 2905 2906 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs; 2907 if (prog) 2908 xdp_qp = nr_cpu_ids; 2909 2910 /* XDP requires extra queues for XDP_TX */ 2911 if (curr_qp + xdp_qp > vi->max_queue_pairs) { 2912 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", 2913 curr_qp + xdp_qp, vi->max_queue_pairs); 2914 xdp_qp = 0; 2915 } 2916 2917 old_prog = rtnl_dereference(vi->rq[0].xdp_prog); 2918 if (!prog && !old_prog) 2919 return 0; 2920 2921 if (prog) 2922 bpf_prog_add(prog, vi->max_queue_pairs - 1); 2923 2924 /* Make sure NAPI is not using any XDP TX queues for RX. */ 2925 if (netif_running(dev)) { 2926 for (i = 0; i < vi->max_queue_pairs; i++) { 2927 napi_disable(&vi->rq[i].napi); 2928 virtnet_napi_tx_disable(&vi->sq[i].napi); 2929 } 2930 } 2931 2932 if (!prog) { 2933 for (i = 0; i < vi->max_queue_pairs; i++) { 2934 rcu_assign_pointer(vi->rq[i].xdp_prog, prog); 2935 if (i == 0) 2936 virtnet_restore_guest_offloads(vi); 2937 } 2938 synchronize_net(); 2939 } 2940 2941 err = _virtnet_set_queues(vi, curr_qp + xdp_qp); 2942 if (err) 2943 goto err; 2944 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp); 2945 vi->xdp_queue_pairs = xdp_qp; 2946 2947 if (prog) { 2948 vi->xdp_enabled = true; 2949 for (i = 0; i < vi->max_queue_pairs; i++) { 2950 rcu_assign_pointer(vi->rq[i].xdp_prog, prog); 2951 if (i == 0 && !old_prog) 2952 virtnet_clear_guest_offloads(vi); 2953 } 2954 } else { 2955 vi->xdp_enabled = false; 2956 } 2957 2958 for (i = 0; i < vi->max_queue_pairs; i++) { 2959 if (old_prog) 2960 bpf_prog_put(old_prog); 2961 if (netif_running(dev)) { 2962 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2963 virtnet_napi_tx_enable(vi, vi->sq[i].vq, 2964 &vi->sq[i].napi); 2965 } 2966 } 2967 2968 return 0; 2969 2970 err: 2971 if (!prog) { 2972 virtnet_clear_guest_offloads(vi); 2973 for (i = 0; i < vi->max_queue_pairs; i++) 2974 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog); 2975 } 2976 2977 if (netif_running(dev)) { 2978 for (i = 0; i < vi->max_queue_pairs; i++) { 2979 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2980 virtnet_napi_tx_enable(vi, vi->sq[i].vq, 2981 &vi->sq[i].napi); 2982 } 2983 } 2984 if (prog) 2985 bpf_prog_sub(prog, vi->max_queue_pairs - 1); 2986 return err; 2987 } 2988 2989 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp) 2990 { 2991 switch (xdp->command) { 2992 case XDP_SETUP_PROG: 2993 return virtnet_xdp_set(dev, xdp->prog, xdp->extack); 2994 default: 2995 return -EINVAL; 2996 } 2997 } 2998 2999 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf, 3000 size_t len) 3001 { 3002 struct virtnet_info *vi = netdev_priv(dev); 3003 int ret; 3004 3005 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY)) 3006 return -EOPNOTSUPP; 3007 3008 ret = snprintf(buf, len, "sby"); 3009 if (ret >= len) 3010 return -EOPNOTSUPP; 3011 3012 return 0; 3013 } 3014 3015 static int virtnet_set_features(struct net_device *dev, 3016 netdev_features_t features) 3017 { 3018 struct virtnet_info *vi = netdev_priv(dev); 3019 u64 offloads; 3020 int err; 3021 3022 if ((dev->features ^ features) & NETIF_F_GRO_HW) { 3023 if (vi->xdp_enabled) 3024 return -EBUSY; 3025 3026 if (features & NETIF_F_GRO_HW) 3027 offloads = vi->guest_offloads_capable; 3028 else 3029 offloads = vi->guest_offloads_capable & 3030 ~GUEST_OFFLOAD_GRO_HW_MASK; 3031 3032 err = virtnet_set_guest_offloads(vi, offloads); 3033 if (err) 3034 return err; 3035 vi->guest_offloads = offloads; 3036 } 3037 3038 if ((dev->features ^ features) & NETIF_F_RXHASH) { 3039 if (features & NETIF_F_RXHASH) 3040 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved; 3041 else 3042 vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE; 3043 3044 if (!virtnet_commit_rss_command(vi)) 3045 return -EINVAL; 3046 } 3047 3048 return 0; 3049 } 3050 3051 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue) 3052 { 3053 struct virtnet_info *priv = netdev_priv(dev); 3054 struct send_queue *sq = &priv->sq[txqueue]; 3055 struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue); 3056 3057 u64_stats_update_begin(&sq->stats.syncp); 3058 sq->stats.tx_timeouts++; 3059 u64_stats_update_end(&sq->stats.syncp); 3060 3061 netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n", 3062 txqueue, sq->name, sq->vq->index, sq->vq->name, 3063 jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start))); 3064 } 3065 3066 static const struct net_device_ops virtnet_netdev = { 3067 .ndo_open = virtnet_open, 3068 .ndo_stop = virtnet_close, 3069 .ndo_start_xmit = start_xmit, 3070 .ndo_validate_addr = eth_validate_addr, 3071 .ndo_set_mac_address = virtnet_set_mac_address, 3072 .ndo_set_rx_mode = virtnet_set_rx_mode, 3073 .ndo_get_stats64 = virtnet_stats, 3074 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid, 3075 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid, 3076 .ndo_bpf = virtnet_xdp, 3077 .ndo_xdp_xmit = virtnet_xdp_xmit, 3078 .ndo_features_check = passthru_features_check, 3079 .ndo_get_phys_port_name = virtnet_get_phys_port_name, 3080 .ndo_set_features = virtnet_set_features, 3081 .ndo_tx_timeout = virtnet_tx_timeout, 3082 }; 3083 3084 static void virtnet_config_changed_work(struct work_struct *work) 3085 { 3086 struct virtnet_info *vi = 3087 container_of(work, struct virtnet_info, config_work); 3088 u16 v; 3089 3090 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS, 3091 struct virtio_net_config, status, &v) < 0) 3092 return; 3093 3094 if (v & VIRTIO_NET_S_ANNOUNCE) { 3095 netdev_notify_peers(vi->dev); 3096 virtnet_ack_link_announce(vi); 3097 } 3098 3099 /* Ignore unknown (future) status bits */ 3100 v &= VIRTIO_NET_S_LINK_UP; 3101 3102 if (vi->status == v) 3103 return; 3104 3105 vi->status = v; 3106 3107 if (vi->status & VIRTIO_NET_S_LINK_UP) { 3108 virtnet_update_settings(vi); 3109 netif_carrier_on(vi->dev); 3110 netif_tx_wake_all_queues(vi->dev); 3111 } else { 3112 netif_carrier_off(vi->dev); 3113 netif_tx_stop_all_queues(vi->dev); 3114 } 3115 } 3116 3117 static void virtnet_config_changed(struct virtio_device *vdev) 3118 { 3119 struct virtnet_info *vi = vdev->priv; 3120 3121 schedule_work(&vi->config_work); 3122 } 3123 3124 static void virtnet_free_queues(struct virtnet_info *vi) 3125 { 3126 int i; 3127 3128 for (i = 0; i < vi->max_queue_pairs; i++) { 3129 __netif_napi_del(&vi->rq[i].napi); 3130 __netif_napi_del(&vi->sq[i].napi); 3131 } 3132 3133 /* We called __netif_napi_del(), 3134 * we need to respect an RCU grace period before freeing vi->rq 3135 */ 3136 synchronize_net(); 3137 3138 kfree(vi->rq); 3139 kfree(vi->sq); 3140 kfree(vi->ctrl); 3141 } 3142 3143 static void _free_receive_bufs(struct virtnet_info *vi) 3144 { 3145 struct bpf_prog *old_prog; 3146 int i; 3147 3148 for (i = 0; i < vi->max_queue_pairs; i++) { 3149 while (vi->rq[i].pages) 3150 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0); 3151 3152 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 3153 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL); 3154 if (old_prog) 3155 bpf_prog_put(old_prog); 3156 } 3157 } 3158 3159 static void free_receive_bufs(struct virtnet_info *vi) 3160 { 3161 rtnl_lock(); 3162 _free_receive_bufs(vi); 3163 rtnl_unlock(); 3164 } 3165 3166 static void free_receive_page_frags(struct virtnet_info *vi) 3167 { 3168 int i; 3169 for (i = 0; i < vi->max_queue_pairs; i++) 3170 if (vi->rq[i].alloc_frag.page) 3171 put_page(vi->rq[i].alloc_frag.page); 3172 } 3173 3174 static void free_unused_bufs(struct virtnet_info *vi) 3175 { 3176 void *buf; 3177 int i; 3178 3179 for (i = 0; i < vi->max_queue_pairs; i++) { 3180 struct virtqueue *vq = vi->sq[i].vq; 3181 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) { 3182 if (!is_xdp_frame(buf)) 3183 dev_kfree_skb(buf); 3184 else 3185 xdp_return_frame(ptr_to_xdp(buf)); 3186 } 3187 } 3188 3189 for (i = 0; i < vi->max_queue_pairs; i++) { 3190 struct virtqueue *vq = vi->rq[i].vq; 3191 3192 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) { 3193 if (vi->mergeable_rx_bufs) { 3194 put_page(virt_to_head_page(buf)); 3195 } else if (vi->big_packets) { 3196 give_pages(&vi->rq[i], buf); 3197 } else { 3198 put_page(virt_to_head_page(buf)); 3199 } 3200 } 3201 } 3202 } 3203 3204 static void virtnet_del_vqs(struct virtnet_info *vi) 3205 { 3206 struct virtio_device *vdev = vi->vdev; 3207 3208 virtnet_clean_affinity(vi); 3209 3210 vdev->config->del_vqs(vdev); 3211 3212 virtnet_free_queues(vi); 3213 } 3214 3215 /* How large should a single buffer be so a queue full of these can fit at 3216 * least one full packet? 3217 * Logic below assumes the mergeable buffer header is used. 3218 */ 3219 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq) 3220 { 3221 const unsigned int hdr_len = vi->hdr_len; 3222 unsigned int rq_size = virtqueue_get_vring_size(vq); 3223 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu; 3224 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len; 3225 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size); 3226 3227 return max(max(min_buf_len, hdr_len) - hdr_len, 3228 (unsigned int)GOOD_PACKET_LEN); 3229 } 3230 3231 static int virtnet_find_vqs(struct virtnet_info *vi) 3232 { 3233 vq_callback_t **callbacks; 3234 struct virtqueue **vqs; 3235 int ret = -ENOMEM; 3236 int i, total_vqs; 3237 const char **names; 3238 bool *ctx; 3239 3240 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by 3241 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by 3242 * possible control vq. 3243 */ 3244 total_vqs = vi->max_queue_pairs * 2 + 3245 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ); 3246 3247 /* Allocate space for find_vqs parameters */ 3248 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL); 3249 if (!vqs) 3250 goto err_vq; 3251 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL); 3252 if (!callbacks) 3253 goto err_callback; 3254 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL); 3255 if (!names) 3256 goto err_names; 3257 if (!vi->big_packets || vi->mergeable_rx_bufs) { 3258 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL); 3259 if (!ctx) 3260 goto err_ctx; 3261 } else { 3262 ctx = NULL; 3263 } 3264 3265 /* Parameters for control virtqueue, if any */ 3266 if (vi->has_cvq) { 3267 callbacks[total_vqs - 1] = NULL; 3268 names[total_vqs - 1] = "control"; 3269 } 3270 3271 /* Allocate/initialize parameters for send/receive virtqueues */ 3272 for (i = 0; i < vi->max_queue_pairs; i++) { 3273 callbacks[rxq2vq(i)] = skb_recv_done; 3274 callbacks[txq2vq(i)] = skb_xmit_done; 3275 sprintf(vi->rq[i].name, "input.%d", i); 3276 sprintf(vi->sq[i].name, "output.%d", i); 3277 names[rxq2vq(i)] = vi->rq[i].name; 3278 names[txq2vq(i)] = vi->sq[i].name; 3279 if (ctx) 3280 ctx[rxq2vq(i)] = true; 3281 } 3282 3283 ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks, 3284 names, ctx, NULL); 3285 if (ret) 3286 goto err_find; 3287 3288 if (vi->has_cvq) { 3289 vi->cvq = vqs[total_vqs - 1]; 3290 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN)) 3291 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER; 3292 } 3293 3294 for (i = 0; i < vi->max_queue_pairs; i++) { 3295 vi->rq[i].vq = vqs[rxq2vq(i)]; 3296 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq); 3297 vi->sq[i].vq = vqs[txq2vq(i)]; 3298 } 3299 3300 /* run here: ret == 0. */ 3301 3302 3303 err_find: 3304 kfree(ctx); 3305 err_ctx: 3306 kfree(names); 3307 err_names: 3308 kfree(callbacks); 3309 err_callback: 3310 kfree(vqs); 3311 err_vq: 3312 return ret; 3313 } 3314 3315 static int virtnet_alloc_queues(struct virtnet_info *vi) 3316 { 3317 int i; 3318 3319 if (vi->has_cvq) { 3320 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL); 3321 if (!vi->ctrl) 3322 goto err_ctrl; 3323 } else { 3324 vi->ctrl = NULL; 3325 } 3326 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL); 3327 if (!vi->sq) 3328 goto err_sq; 3329 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL); 3330 if (!vi->rq) 3331 goto err_rq; 3332 3333 INIT_DELAYED_WORK(&vi->refill, refill_work); 3334 for (i = 0; i < vi->max_queue_pairs; i++) { 3335 vi->rq[i].pages = NULL; 3336 netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll, 3337 napi_weight); 3338 netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi, 3339 virtnet_poll_tx, 3340 napi_tx ? napi_weight : 0); 3341 3342 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg)); 3343 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len); 3344 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg)); 3345 3346 u64_stats_init(&vi->rq[i].stats.syncp); 3347 u64_stats_init(&vi->sq[i].stats.syncp); 3348 } 3349 3350 return 0; 3351 3352 err_rq: 3353 kfree(vi->sq); 3354 err_sq: 3355 kfree(vi->ctrl); 3356 err_ctrl: 3357 return -ENOMEM; 3358 } 3359 3360 static int init_vqs(struct virtnet_info *vi) 3361 { 3362 int ret; 3363 3364 /* Allocate send & receive queues */ 3365 ret = virtnet_alloc_queues(vi); 3366 if (ret) 3367 goto err; 3368 3369 ret = virtnet_find_vqs(vi); 3370 if (ret) 3371 goto err_free; 3372 3373 cpus_read_lock(); 3374 virtnet_set_affinity(vi); 3375 cpus_read_unlock(); 3376 3377 return 0; 3378 3379 err_free: 3380 virtnet_free_queues(vi); 3381 err: 3382 return ret; 3383 } 3384 3385 #ifdef CONFIG_SYSFS 3386 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue, 3387 char *buf) 3388 { 3389 struct virtnet_info *vi = netdev_priv(queue->dev); 3390 unsigned int queue_index = get_netdev_rx_queue_index(queue); 3391 unsigned int headroom = virtnet_get_headroom(vi); 3392 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 3393 struct ewma_pkt_len *avg; 3394 3395 BUG_ON(queue_index >= vi->max_queue_pairs); 3396 avg = &vi->rq[queue_index].mrg_avg_pkt_len; 3397 return sprintf(buf, "%u\n", 3398 get_mergeable_buf_len(&vi->rq[queue_index], avg, 3399 SKB_DATA_ALIGN(headroom + tailroom))); 3400 } 3401 3402 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute = 3403 __ATTR_RO(mergeable_rx_buffer_size); 3404 3405 static struct attribute *virtio_net_mrg_rx_attrs[] = { 3406 &mergeable_rx_buffer_size_attribute.attr, 3407 NULL 3408 }; 3409 3410 static const struct attribute_group virtio_net_mrg_rx_group = { 3411 .name = "virtio_net", 3412 .attrs = virtio_net_mrg_rx_attrs 3413 }; 3414 #endif 3415 3416 static bool virtnet_fail_on_feature(struct virtio_device *vdev, 3417 unsigned int fbit, 3418 const char *fname, const char *dname) 3419 { 3420 if (!virtio_has_feature(vdev, fbit)) 3421 return false; 3422 3423 dev_err(&vdev->dev, "device advertises feature %s but not %s", 3424 fname, dname); 3425 3426 return true; 3427 } 3428 3429 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \ 3430 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit) 3431 3432 static bool virtnet_validate_features(struct virtio_device *vdev) 3433 { 3434 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) && 3435 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX, 3436 "VIRTIO_NET_F_CTRL_VQ") || 3437 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN, 3438 "VIRTIO_NET_F_CTRL_VQ") || 3439 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE, 3440 "VIRTIO_NET_F_CTRL_VQ") || 3441 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") || 3442 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR, 3443 "VIRTIO_NET_F_CTRL_VQ") || 3444 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS, 3445 "VIRTIO_NET_F_CTRL_VQ") || 3446 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT, 3447 "VIRTIO_NET_F_CTRL_VQ"))) { 3448 return false; 3449 } 3450 3451 return true; 3452 } 3453 3454 #define MIN_MTU ETH_MIN_MTU 3455 #define MAX_MTU ETH_MAX_MTU 3456 3457 static int virtnet_validate(struct virtio_device *vdev) 3458 { 3459 if (!vdev->config->get) { 3460 dev_err(&vdev->dev, "%s failure: config access disabled\n", 3461 __func__); 3462 return -EINVAL; 3463 } 3464 3465 if (!virtnet_validate_features(vdev)) 3466 return -EINVAL; 3467 3468 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 3469 int mtu = virtio_cread16(vdev, 3470 offsetof(struct virtio_net_config, 3471 mtu)); 3472 if (mtu < MIN_MTU) 3473 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU); 3474 } 3475 3476 return 0; 3477 } 3478 3479 static int virtnet_probe(struct virtio_device *vdev) 3480 { 3481 int i, err = -ENOMEM; 3482 struct net_device *dev; 3483 struct virtnet_info *vi; 3484 u16 max_queue_pairs; 3485 int mtu; 3486 3487 /* Find if host supports multiqueue/rss virtio_net device */ 3488 max_queue_pairs = 1; 3489 if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS)) 3490 max_queue_pairs = 3491 virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs)); 3492 3493 /* We need at least 2 queue's */ 3494 if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN || 3495 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX || 3496 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 3497 max_queue_pairs = 1; 3498 3499 /* Allocate ourselves a network device with room for our info */ 3500 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs); 3501 if (!dev) 3502 return -ENOMEM; 3503 3504 /* Set up network device as normal. */ 3505 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE | 3506 IFF_TX_SKB_NO_LINEAR; 3507 dev->netdev_ops = &virtnet_netdev; 3508 dev->features = NETIF_F_HIGHDMA; 3509 3510 dev->ethtool_ops = &virtnet_ethtool_ops; 3511 SET_NETDEV_DEV(dev, &vdev->dev); 3512 3513 /* Do we support "hardware" checksums? */ 3514 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) { 3515 /* This opens up the world of extra features. */ 3516 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG; 3517 if (csum) 3518 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG; 3519 3520 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) { 3521 dev->hw_features |= NETIF_F_TSO 3522 | NETIF_F_TSO_ECN | NETIF_F_TSO6; 3523 } 3524 /* Individual feature bits: what can host handle? */ 3525 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4)) 3526 dev->hw_features |= NETIF_F_TSO; 3527 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6)) 3528 dev->hw_features |= NETIF_F_TSO6; 3529 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN)) 3530 dev->hw_features |= NETIF_F_TSO_ECN; 3531 3532 dev->features |= NETIF_F_GSO_ROBUST; 3533 3534 if (gso) 3535 dev->features |= dev->hw_features & NETIF_F_ALL_TSO; 3536 /* (!csum && gso) case will be fixed by register_netdev() */ 3537 } 3538 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM)) 3539 dev->features |= NETIF_F_RXCSUM; 3540 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) || 3541 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6)) 3542 dev->features |= NETIF_F_GRO_HW; 3543 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) 3544 dev->hw_features |= NETIF_F_GRO_HW; 3545 3546 dev->vlan_features = dev->features; 3547 3548 /* MTU range: 68 - 65535 */ 3549 dev->min_mtu = MIN_MTU; 3550 dev->max_mtu = MAX_MTU; 3551 3552 /* Configuration may specify what MAC to use. Otherwise random. */ 3553 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) { 3554 u8 addr[ETH_ALEN]; 3555 3556 virtio_cread_bytes(vdev, 3557 offsetof(struct virtio_net_config, mac), 3558 addr, ETH_ALEN); 3559 eth_hw_addr_set(dev, addr); 3560 } else { 3561 eth_hw_addr_random(dev); 3562 } 3563 3564 /* Set up our device-specific information */ 3565 vi = netdev_priv(dev); 3566 vi->dev = dev; 3567 vi->vdev = vdev; 3568 vdev->priv = vi; 3569 3570 INIT_WORK(&vi->config_work, virtnet_config_changed_work); 3571 spin_lock_init(&vi->refill_lock); 3572 3573 /* If we can receive ANY GSO packets, we must allocate large ones. */ 3574 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) || 3575 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) || 3576 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) || 3577 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO)) 3578 vi->big_packets = true; 3579 3580 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) 3581 vi->mergeable_rx_bufs = true; 3582 3583 if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT)) 3584 vi->has_rss_hash_report = true; 3585 3586 if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS)) 3587 vi->has_rss = true; 3588 3589 if (vi->has_rss || vi->has_rss_hash_report) { 3590 vi->rss_indir_table_size = 3591 virtio_cread16(vdev, offsetof(struct virtio_net_config, 3592 rss_max_indirection_table_length)); 3593 vi->rss_key_size = 3594 virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size)); 3595 3596 vi->rss_hash_types_supported = 3597 virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types)); 3598 vi->rss_hash_types_supported &= 3599 ~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX | 3600 VIRTIO_NET_RSS_HASH_TYPE_TCP_EX | 3601 VIRTIO_NET_RSS_HASH_TYPE_UDP_EX); 3602 3603 dev->hw_features |= NETIF_F_RXHASH; 3604 } 3605 3606 if (vi->has_rss_hash_report) 3607 vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash); 3608 else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) || 3609 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 3610 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 3611 else 3612 vi->hdr_len = sizeof(struct virtio_net_hdr); 3613 3614 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) || 3615 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 3616 vi->any_header_sg = true; 3617 3618 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 3619 vi->has_cvq = true; 3620 3621 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 3622 mtu = virtio_cread16(vdev, 3623 offsetof(struct virtio_net_config, 3624 mtu)); 3625 if (mtu < dev->min_mtu) { 3626 /* Should never trigger: MTU was previously validated 3627 * in virtnet_validate. 3628 */ 3629 dev_err(&vdev->dev, 3630 "device MTU appears to have changed it is now %d < %d", 3631 mtu, dev->min_mtu); 3632 err = -EINVAL; 3633 goto free; 3634 } 3635 3636 dev->mtu = mtu; 3637 dev->max_mtu = mtu; 3638 3639 /* TODO: size buffers correctly in this case. */ 3640 if (dev->mtu > ETH_DATA_LEN) 3641 vi->big_packets = true; 3642 } 3643 3644 if (vi->any_header_sg) 3645 dev->needed_headroom = vi->hdr_len; 3646 3647 /* Enable multiqueue by default */ 3648 if (num_online_cpus() >= max_queue_pairs) 3649 vi->curr_queue_pairs = max_queue_pairs; 3650 else 3651 vi->curr_queue_pairs = num_online_cpus(); 3652 vi->max_queue_pairs = max_queue_pairs; 3653 3654 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */ 3655 err = init_vqs(vi); 3656 if (err) 3657 goto free; 3658 3659 #ifdef CONFIG_SYSFS 3660 if (vi->mergeable_rx_bufs) 3661 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group; 3662 #endif 3663 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs); 3664 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs); 3665 3666 virtnet_init_settings(dev); 3667 3668 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) { 3669 vi->failover = net_failover_create(vi->dev); 3670 if (IS_ERR(vi->failover)) { 3671 err = PTR_ERR(vi->failover); 3672 goto free_vqs; 3673 } 3674 } 3675 3676 if (vi->has_rss || vi->has_rss_hash_report) 3677 virtnet_init_default_rss(vi); 3678 3679 /* serialize netdev register + virtio_device_ready() with ndo_open() */ 3680 rtnl_lock(); 3681 3682 err = register_netdevice(dev); 3683 if (err) { 3684 pr_debug("virtio_net: registering device failed\n"); 3685 rtnl_unlock(); 3686 goto free_failover; 3687 } 3688 3689 virtio_device_ready(vdev); 3690 3691 rtnl_unlock(); 3692 3693 err = virtnet_cpu_notif_add(vi); 3694 if (err) { 3695 pr_debug("virtio_net: registering cpu notifier failed\n"); 3696 goto free_unregister_netdev; 3697 } 3698 3699 virtnet_set_queues(vi, vi->curr_queue_pairs); 3700 3701 /* Assume link up if device can't report link status, 3702 otherwise get link status from config. */ 3703 netif_carrier_off(dev); 3704 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) { 3705 schedule_work(&vi->config_work); 3706 } else { 3707 vi->status = VIRTIO_NET_S_LINK_UP; 3708 virtnet_update_settings(vi); 3709 netif_carrier_on(dev); 3710 } 3711 3712 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++) 3713 if (virtio_has_feature(vi->vdev, guest_offloads[i])) 3714 set_bit(guest_offloads[i], &vi->guest_offloads); 3715 vi->guest_offloads_capable = vi->guest_offloads; 3716 3717 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n", 3718 dev->name, max_queue_pairs); 3719 3720 return 0; 3721 3722 free_unregister_netdev: 3723 virtio_reset_device(vdev); 3724 3725 unregister_netdev(dev); 3726 free_failover: 3727 net_failover_destroy(vi->failover); 3728 free_vqs: 3729 cancel_delayed_work_sync(&vi->refill); 3730 free_receive_page_frags(vi); 3731 virtnet_del_vqs(vi); 3732 free: 3733 free_netdev(dev); 3734 return err; 3735 } 3736 3737 static void remove_vq_common(struct virtnet_info *vi) 3738 { 3739 virtio_reset_device(vi->vdev); 3740 3741 /* Free unused buffers in both send and recv, if any. */ 3742 free_unused_bufs(vi); 3743 3744 free_receive_bufs(vi); 3745 3746 free_receive_page_frags(vi); 3747 3748 virtnet_del_vqs(vi); 3749 } 3750 3751 static void virtnet_remove(struct virtio_device *vdev) 3752 { 3753 struct virtnet_info *vi = vdev->priv; 3754 3755 virtnet_cpu_notif_remove(vi); 3756 3757 /* Make sure no work handler is accessing the device. */ 3758 flush_work(&vi->config_work); 3759 3760 unregister_netdev(vi->dev); 3761 3762 net_failover_destroy(vi->failover); 3763 3764 remove_vq_common(vi); 3765 3766 free_netdev(vi->dev); 3767 } 3768 3769 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev) 3770 { 3771 struct virtnet_info *vi = vdev->priv; 3772 3773 virtnet_cpu_notif_remove(vi); 3774 virtnet_freeze_down(vdev); 3775 remove_vq_common(vi); 3776 3777 return 0; 3778 } 3779 3780 static __maybe_unused int virtnet_restore(struct virtio_device *vdev) 3781 { 3782 struct virtnet_info *vi = vdev->priv; 3783 int err; 3784 3785 err = virtnet_restore_up(vdev); 3786 if (err) 3787 return err; 3788 virtnet_set_queues(vi, vi->curr_queue_pairs); 3789 3790 err = virtnet_cpu_notif_add(vi); 3791 if (err) { 3792 virtnet_freeze_down(vdev); 3793 remove_vq_common(vi); 3794 return err; 3795 } 3796 3797 return 0; 3798 } 3799 3800 static struct virtio_device_id id_table[] = { 3801 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID }, 3802 { 0 }, 3803 }; 3804 3805 #define VIRTNET_FEATURES \ 3806 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \ 3807 VIRTIO_NET_F_MAC, \ 3808 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \ 3809 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \ 3810 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \ 3811 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \ 3812 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \ 3813 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \ 3814 VIRTIO_NET_F_CTRL_MAC_ADDR, \ 3815 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \ 3816 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \ 3817 VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT 3818 3819 static unsigned int features[] = { 3820 VIRTNET_FEATURES, 3821 }; 3822 3823 static unsigned int features_legacy[] = { 3824 VIRTNET_FEATURES, 3825 VIRTIO_NET_F_GSO, 3826 VIRTIO_F_ANY_LAYOUT, 3827 }; 3828 3829 static struct virtio_driver virtio_net_driver = { 3830 .feature_table = features, 3831 .feature_table_size = ARRAY_SIZE(features), 3832 .feature_table_legacy = features_legacy, 3833 .feature_table_size_legacy = ARRAY_SIZE(features_legacy), 3834 .driver.name = KBUILD_MODNAME, 3835 .driver.owner = THIS_MODULE, 3836 .id_table = id_table, 3837 .validate = virtnet_validate, 3838 .probe = virtnet_probe, 3839 .remove = virtnet_remove, 3840 .config_changed = virtnet_config_changed, 3841 #ifdef CONFIG_PM_SLEEP 3842 .freeze = virtnet_freeze, 3843 .restore = virtnet_restore, 3844 #endif 3845 }; 3846 3847 static __init int virtio_net_driver_init(void) 3848 { 3849 int ret; 3850 3851 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online", 3852 virtnet_cpu_online, 3853 virtnet_cpu_down_prep); 3854 if (ret < 0) 3855 goto out; 3856 virtionet_online = ret; 3857 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead", 3858 NULL, virtnet_cpu_dead); 3859 if (ret) 3860 goto err_dead; 3861 ret = register_virtio_driver(&virtio_net_driver); 3862 if (ret) 3863 goto err_virtio; 3864 return 0; 3865 err_virtio: 3866 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 3867 err_dead: 3868 cpuhp_remove_multi_state(virtionet_online); 3869 out: 3870 return ret; 3871 } 3872 module_init(virtio_net_driver_init); 3873 3874 static __exit void virtio_net_driver_exit(void) 3875 { 3876 unregister_virtio_driver(&virtio_net_driver); 3877 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 3878 cpuhp_remove_multi_state(virtionet_online); 3879 } 3880 module_exit(virtio_net_driver_exit); 3881 3882 MODULE_DEVICE_TABLE(virtio, id_table); 3883 MODULE_DESCRIPTION("Virtio network driver"); 3884 MODULE_LICENSE("GPL"); 3885