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