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