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 struct sk_buff *receive_mergeable(struct net_device *dev, 1170 struct virtnet_info *vi, 1171 struct receive_queue *rq, 1172 void *buf, 1173 void *ctx, 1174 unsigned int len, 1175 unsigned int *xdp_xmit, 1176 struct virtnet_rq_stats *stats) 1177 { 1178 struct virtio_net_hdr_mrg_rxbuf *hdr = buf; 1179 int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers); 1180 struct page *page = virt_to_head_page(buf); 1181 int offset = buf - page_address(page); 1182 struct sk_buff *head_skb, *curr_skb; 1183 struct bpf_prog *xdp_prog; 1184 unsigned int truesize = mergeable_ctx_to_truesize(ctx); 1185 unsigned int headroom = mergeable_ctx_to_headroom(ctx); 1186 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 1187 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom); 1188 unsigned int frame_sz, xdp_room; 1189 int err; 1190 1191 head_skb = NULL; 1192 stats->bytes += len - vi->hdr_len; 1193 1194 if (unlikely(len > truesize - room)) { 1195 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 1196 dev->name, len, (unsigned long)(truesize - room)); 1197 dev->stats.rx_length_errors++; 1198 goto err_skb; 1199 } 1200 1201 if (likely(!vi->xdp_enabled)) { 1202 xdp_prog = NULL; 1203 goto skip_xdp; 1204 } 1205 1206 rcu_read_lock(); 1207 xdp_prog = rcu_dereference(rq->xdp_prog); 1208 if (xdp_prog) { 1209 unsigned int xdp_frags_truesz = 0; 1210 struct skb_shared_info *shinfo; 1211 struct xdp_frame *xdpf; 1212 struct page *xdp_page; 1213 struct xdp_buff xdp; 1214 void *data; 1215 u32 act; 1216 int i; 1217 1218 /* Transient failure which in theory could occur if 1219 * in-flight packets from before XDP was enabled reach 1220 * the receive path after XDP is loaded. 1221 */ 1222 if (unlikely(hdr->hdr.gso_type)) 1223 goto err_xdp; 1224 1225 /* Now XDP core assumes frag size is PAGE_SIZE, but buffers 1226 * with headroom may add hole in truesize, which 1227 * make their length exceed PAGE_SIZE. So we disabled the 1228 * hole mechanism for xdp. See add_recvbuf_mergeable(). 1229 */ 1230 frame_sz = truesize; 1231 1232 /* This happens when headroom is not enough because 1233 * of the buffer was prefilled before XDP is set. 1234 * This should only happen for the first several packets. 1235 * In fact, vq reset can be used here to help us clean up 1236 * the prefilled buffers, but many existing devices do not 1237 * support it, and we don't want to bother users who are 1238 * using xdp normally. 1239 */ 1240 if (!xdp_prog->aux->xdp_has_frags && 1241 (num_buf > 1 || headroom < virtnet_get_headroom(vi))) { 1242 /* linearize data for XDP */ 1243 xdp_page = xdp_linearize_page(rq, &num_buf, 1244 page, offset, 1245 VIRTIO_XDP_HEADROOM, 1246 &len); 1247 frame_sz = PAGE_SIZE; 1248 1249 if (!xdp_page) 1250 goto err_xdp; 1251 offset = VIRTIO_XDP_HEADROOM; 1252 1253 put_page(page); 1254 page = xdp_page; 1255 } else if (unlikely(headroom < virtnet_get_headroom(vi))) { 1256 xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM + 1257 sizeof(struct skb_shared_info)); 1258 if (len + xdp_room > PAGE_SIZE) 1259 goto err_xdp; 1260 1261 xdp_page = alloc_page(GFP_ATOMIC); 1262 if (!xdp_page) 1263 goto err_xdp; 1264 1265 memcpy(page_address(xdp_page) + VIRTIO_XDP_HEADROOM, 1266 page_address(page) + offset, len); 1267 frame_sz = PAGE_SIZE; 1268 offset = VIRTIO_XDP_HEADROOM; 1269 1270 put_page(page); 1271 page = xdp_page; 1272 } 1273 1274 data = page_address(page) + offset; 1275 err = virtnet_build_xdp_buff_mrg(dev, vi, rq, &xdp, data, len, frame_sz, 1276 &num_buf, &xdp_frags_truesz, stats); 1277 if (unlikely(err)) 1278 goto err_xdp_frags; 1279 1280 act = bpf_prog_run_xdp(xdp_prog, &xdp); 1281 stats->xdp_packets++; 1282 1283 switch (act) { 1284 case XDP_PASS: 1285 head_skb = build_skb_from_xdp_buff(dev, vi, &xdp, xdp_frags_truesz); 1286 if (unlikely(!head_skb)) 1287 goto err_xdp_frags; 1288 1289 rcu_read_unlock(); 1290 return head_skb; 1291 case XDP_TX: 1292 stats->xdp_tx++; 1293 xdpf = xdp_convert_buff_to_frame(&xdp); 1294 if (unlikely(!xdpf)) { 1295 netdev_dbg(dev, "convert buff to frame failed for xdp\n"); 1296 goto err_xdp_frags; 1297 } 1298 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0); 1299 if (unlikely(!err)) { 1300 xdp_return_frame_rx_napi(xdpf); 1301 } else if (unlikely(err < 0)) { 1302 trace_xdp_exception(vi->dev, xdp_prog, act); 1303 goto err_xdp_frags; 1304 } 1305 *xdp_xmit |= VIRTIO_XDP_TX; 1306 rcu_read_unlock(); 1307 goto xdp_xmit; 1308 case XDP_REDIRECT: 1309 stats->xdp_redirects++; 1310 err = xdp_do_redirect(dev, &xdp, xdp_prog); 1311 if (err) 1312 goto err_xdp_frags; 1313 *xdp_xmit |= VIRTIO_XDP_REDIR; 1314 rcu_read_unlock(); 1315 goto xdp_xmit; 1316 default: 1317 bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act); 1318 fallthrough; 1319 case XDP_ABORTED: 1320 trace_xdp_exception(vi->dev, xdp_prog, act); 1321 fallthrough; 1322 case XDP_DROP: 1323 goto err_xdp_frags; 1324 } 1325 err_xdp_frags: 1326 if (xdp_buff_has_frags(&xdp)) { 1327 shinfo = xdp_get_shared_info_from_buff(&xdp); 1328 for (i = 0; i < shinfo->nr_frags; i++) { 1329 xdp_page = skb_frag_page(&shinfo->frags[i]); 1330 put_page(xdp_page); 1331 } 1332 } 1333 1334 goto err_xdp; 1335 } 1336 rcu_read_unlock(); 1337 1338 skip_xdp: 1339 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, headroom); 1340 curr_skb = head_skb; 1341 1342 if (unlikely(!curr_skb)) 1343 goto err_skb; 1344 while (--num_buf) { 1345 int num_skb_frags; 1346 1347 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx); 1348 if (unlikely(!buf)) { 1349 pr_debug("%s: rx error: %d buffers out of %d missing\n", 1350 dev->name, num_buf, 1351 virtio16_to_cpu(vi->vdev, 1352 hdr->num_buffers)); 1353 dev->stats.rx_length_errors++; 1354 goto err_buf; 1355 } 1356 1357 stats->bytes += len; 1358 page = virt_to_head_page(buf); 1359 1360 truesize = mergeable_ctx_to_truesize(ctx); 1361 headroom = mergeable_ctx_to_headroom(ctx); 1362 tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 1363 room = SKB_DATA_ALIGN(headroom + tailroom); 1364 if (unlikely(len > truesize - room)) { 1365 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 1366 dev->name, len, (unsigned long)(truesize - room)); 1367 dev->stats.rx_length_errors++; 1368 goto err_skb; 1369 } 1370 1371 num_skb_frags = skb_shinfo(curr_skb)->nr_frags; 1372 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) { 1373 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC); 1374 1375 if (unlikely(!nskb)) 1376 goto err_skb; 1377 if (curr_skb == head_skb) 1378 skb_shinfo(curr_skb)->frag_list = nskb; 1379 else 1380 curr_skb->next = nskb; 1381 curr_skb = nskb; 1382 head_skb->truesize += nskb->truesize; 1383 num_skb_frags = 0; 1384 } 1385 if (curr_skb != head_skb) { 1386 head_skb->data_len += len; 1387 head_skb->len += len; 1388 head_skb->truesize += truesize; 1389 } 1390 offset = buf - page_address(page); 1391 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) { 1392 put_page(page); 1393 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1, 1394 len, truesize); 1395 } else { 1396 skb_add_rx_frag(curr_skb, num_skb_frags, page, 1397 offset, len, truesize); 1398 } 1399 } 1400 1401 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len); 1402 return head_skb; 1403 1404 err_xdp: 1405 rcu_read_unlock(); 1406 stats->xdp_drops++; 1407 err_skb: 1408 put_page(page); 1409 while (num_buf-- > 1) { 1410 buf = virtqueue_get_buf(rq->vq, &len); 1411 if (unlikely(!buf)) { 1412 pr_debug("%s: rx error: %d buffers missing\n", 1413 dev->name, num_buf); 1414 dev->stats.rx_length_errors++; 1415 break; 1416 } 1417 stats->bytes += len; 1418 page = virt_to_head_page(buf); 1419 put_page(page); 1420 } 1421 err_buf: 1422 stats->drops++; 1423 dev_kfree_skb(head_skb); 1424 xdp_xmit: 1425 return NULL; 1426 } 1427 1428 static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash, 1429 struct sk_buff *skb) 1430 { 1431 enum pkt_hash_types rss_hash_type; 1432 1433 if (!hdr_hash || !skb) 1434 return; 1435 1436 switch (__le16_to_cpu(hdr_hash->hash_report)) { 1437 case VIRTIO_NET_HASH_REPORT_TCPv4: 1438 case VIRTIO_NET_HASH_REPORT_UDPv4: 1439 case VIRTIO_NET_HASH_REPORT_TCPv6: 1440 case VIRTIO_NET_HASH_REPORT_UDPv6: 1441 case VIRTIO_NET_HASH_REPORT_TCPv6_EX: 1442 case VIRTIO_NET_HASH_REPORT_UDPv6_EX: 1443 rss_hash_type = PKT_HASH_TYPE_L4; 1444 break; 1445 case VIRTIO_NET_HASH_REPORT_IPv4: 1446 case VIRTIO_NET_HASH_REPORT_IPv6: 1447 case VIRTIO_NET_HASH_REPORT_IPv6_EX: 1448 rss_hash_type = PKT_HASH_TYPE_L3; 1449 break; 1450 case VIRTIO_NET_HASH_REPORT_NONE: 1451 default: 1452 rss_hash_type = PKT_HASH_TYPE_NONE; 1453 } 1454 skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type); 1455 } 1456 1457 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq, 1458 void *buf, unsigned int len, void **ctx, 1459 unsigned int *xdp_xmit, 1460 struct virtnet_rq_stats *stats) 1461 { 1462 struct net_device *dev = vi->dev; 1463 struct sk_buff *skb; 1464 struct virtio_net_hdr_mrg_rxbuf *hdr; 1465 1466 if (unlikely(len < vi->hdr_len + ETH_HLEN)) { 1467 pr_debug("%s: short packet %i\n", dev->name, len); 1468 dev->stats.rx_length_errors++; 1469 virtnet_rq_free_unused_buf(rq->vq, buf); 1470 return; 1471 } 1472 1473 if (vi->mergeable_rx_bufs) 1474 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit, 1475 stats); 1476 else if (vi->big_packets) 1477 skb = receive_big(dev, vi, rq, buf, len, stats); 1478 else 1479 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats); 1480 1481 if (unlikely(!skb)) 1482 return; 1483 1484 hdr = skb_vnet_hdr(skb); 1485 if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report) 1486 virtio_skb_set_hash((const struct virtio_net_hdr_v1_hash *)hdr, skb); 1487 1488 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) 1489 skb->ip_summed = CHECKSUM_UNNECESSARY; 1490 1491 if (virtio_net_hdr_to_skb(skb, &hdr->hdr, 1492 virtio_is_little_endian(vi->vdev))) { 1493 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n", 1494 dev->name, hdr->hdr.gso_type, 1495 hdr->hdr.gso_size); 1496 goto frame_err; 1497 } 1498 1499 skb_record_rx_queue(skb, vq2rxq(rq->vq)); 1500 skb->protocol = eth_type_trans(skb, dev); 1501 pr_debug("Receiving skb proto 0x%04x len %i type %i\n", 1502 ntohs(skb->protocol), skb->len, skb->pkt_type); 1503 1504 napi_gro_receive(&rq->napi, skb); 1505 return; 1506 1507 frame_err: 1508 dev->stats.rx_frame_errors++; 1509 dev_kfree_skb(skb); 1510 } 1511 1512 /* Unlike mergeable buffers, all buffers are allocated to the 1513 * same size, except for the headroom. For this reason we do 1514 * not need to use mergeable_len_to_ctx here - it is enough 1515 * to store the headroom as the context ignoring the truesize. 1516 */ 1517 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq, 1518 gfp_t gfp) 1519 { 1520 struct page_frag *alloc_frag = &rq->alloc_frag; 1521 char *buf; 1522 unsigned int xdp_headroom = virtnet_get_headroom(vi); 1523 void *ctx = (void *)(unsigned long)xdp_headroom; 1524 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom; 1525 int err; 1526 1527 len = SKB_DATA_ALIGN(len) + 1528 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 1529 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp))) 1530 return -ENOMEM; 1531 1532 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1533 get_page(alloc_frag->page); 1534 alloc_frag->offset += len; 1535 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom, 1536 vi->hdr_len + GOOD_PACKET_LEN); 1537 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1538 if (err < 0) 1539 put_page(virt_to_head_page(buf)); 1540 return err; 1541 } 1542 1543 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq, 1544 gfp_t gfp) 1545 { 1546 struct page *first, *list = NULL; 1547 char *p; 1548 int i, err, offset; 1549 1550 sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2); 1551 1552 /* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */ 1553 for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) { 1554 first = get_a_page(rq, gfp); 1555 if (!first) { 1556 if (list) 1557 give_pages(rq, list); 1558 return -ENOMEM; 1559 } 1560 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE); 1561 1562 /* chain new page in list head to match sg */ 1563 first->private = (unsigned long)list; 1564 list = first; 1565 } 1566 1567 first = get_a_page(rq, gfp); 1568 if (!first) { 1569 give_pages(rq, list); 1570 return -ENOMEM; 1571 } 1572 p = page_address(first); 1573 1574 /* rq->sg[0], rq->sg[1] share the same page */ 1575 /* a separated rq->sg[0] for header - required in case !any_header_sg */ 1576 sg_set_buf(&rq->sg[0], p, vi->hdr_len); 1577 1578 /* rq->sg[1] for data packet, from offset */ 1579 offset = sizeof(struct padded_vnet_hdr); 1580 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset); 1581 1582 /* chain first in list head */ 1583 first->private = (unsigned long)list; 1584 err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2, 1585 first, gfp); 1586 if (err < 0) 1587 give_pages(rq, first); 1588 1589 return err; 1590 } 1591 1592 static unsigned int get_mergeable_buf_len(struct receive_queue *rq, 1593 struct ewma_pkt_len *avg_pkt_len, 1594 unsigned int room) 1595 { 1596 struct virtnet_info *vi = rq->vq->vdev->priv; 1597 const size_t hdr_len = vi->hdr_len; 1598 unsigned int len; 1599 1600 if (room) 1601 return PAGE_SIZE - room; 1602 1603 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len), 1604 rq->min_buf_len, PAGE_SIZE - hdr_len); 1605 1606 return ALIGN(len, L1_CACHE_BYTES); 1607 } 1608 1609 static int add_recvbuf_mergeable(struct virtnet_info *vi, 1610 struct receive_queue *rq, gfp_t gfp) 1611 { 1612 struct page_frag *alloc_frag = &rq->alloc_frag; 1613 unsigned int headroom = virtnet_get_headroom(vi); 1614 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 1615 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom); 1616 char *buf; 1617 void *ctx; 1618 int err; 1619 unsigned int len, hole; 1620 1621 /* Extra tailroom is needed to satisfy XDP's assumption. This 1622 * means rx frags coalescing won't work, but consider we've 1623 * disabled GSO for XDP, it won't be a big issue. 1624 */ 1625 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room); 1626 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp))) 1627 return -ENOMEM; 1628 1629 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1630 buf += headroom; /* advance address leaving hole at front of pkt */ 1631 get_page(alloc_frag->page); 1632 alloc_frag->offset += len + room; 1633 hole = alloc_frag->size - alloc_frag->offset; 1634 if (hole < len + room) { 1635 /* To avoid internal fragmentation, if there is very likely not 1636 * enough space for another buffer, add the remaining space to 1637 * the current buffer. 1638 * XDP core assumes that frame_size of xdp_buff and the length 1639 * of the frag are PAGE_SIZE, so we disable the hole mechanism. 1640 */ 1641 if (!headroom) 1642 len += hole; 1643 alloc_frag->offset += hole; 1644 } 1645 1646 sg_init_one(rq->sg, buf, len); 1647 ctx = mergeable_len_to_ctx(len + room, headroom); 1648 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1649 if (err < 0) 1650 put_page(virt_to_head_page(buf)); 1651 1652 return err; 1653 } 1654 1655 /* 1656 * Returns false if we couldn't fill entirely (OOM). 1657 * 1658 * Normally run in the receive path, but can also be run from ndo_open 1659 * before we're receiving packets, or from refill_work which is 1660 * careful to disable receiving (using napi_disable). 1661 */ 1662 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq, 1663 gfp_t gfp) 1664 { 1665 int err; 1666 bool oom; 1667 1668 do { 1669 if (vi->mergeable_rx_bufs) 1670 err = add_recvbuf_mergeable(vi, rq, gfp); 1671 else if (vi->big_packets) 1672 err = add_recvbuf_big(vi, rq, gfp); 1673 else 1674 err = add_recvbuf_small(vi, rq, gfp); 1675 1676 oom = err == -ENOMEM; 1677 if (err) 1678 break; 1679 } while (rq->vq->num_free); 1680 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) { 1681 unsigned long flags; 1682 1683 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp); 1684 rq->stats.kicks++; 1685 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags); 1686 } 1687 1688 return !oom; 1689 } 1690 1691 static void skb_recv_done(struct virtqueue *rvq) 1692 { 1693 struct virtnet_info *vi = rvq->vdev->priv; 1694 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)]; 1695 1696 virtqueue_napi_schedule(&rq->napi, rvq); 1697 } 1698 1699 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi) 1700 { 1701 napi_enable(napi); 1702 1703 /* If all buffers were filled by other side before we napi_enabled, we 1704 * won't get another interrupt, so process any outstanding packets now. 1705 * Call local_bh_enable after to trigger softIRQ processing. 1706 */ 1707 local_bh_disable(); 1708 virtqueue_napi_schedule(napi, vq); 1709 local_bh_enable(); 1710 } 1711 1712 static void virtnet_napi_tx_enable(struct virtnet_info *vi, 1713 struct virtqueue *vq, 1714 struct napi_struct *napi) 1715 { 1716 if (!napi->weight) 1717 return; 1718 1719 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only 1720 * enable the feature if this is likely affine with the transmit path. 1721 */ 1722 if (!vi->affinity_hint_set) { 1723 napi->weight = 0; 1724 return; 1725 } 1726 1727 return virtnet_napi_enable(vq, napi); 1728 } 1729 1730 static void virtnet_napi_tx_disable(struct napi_struct *napi) 1731 { 1732 if (napi->weight) 1733 napi_disable(napi); 1734 } 1735 1736 static void refill_work(struct work_struct *work) 1737 { 1738 struct virtnet_info *vi = 1739 container_of(work, struct virtnet_info, refill.work); 1740 bool still_empty; 1741 int i; 1742 1743 for (i = 0; i < vi->curr_queue_pairs; i++) { 1744 struct receive_queue *rq = &vi->rq[i]; 1745 1746 napi_disable(&rq->napi); 1747 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL); 1748 virtnet_napi_enable(rq->vq, &rq->napi); 1749 1750 /* In theory, this can happen: if we don't get any buffers in 1751 * we will *never* try to fill again. 1752 */ 1753 if (still_empty) 1754 schedule_delayed_work(&vi->refill, HZ/2); 1755 } 1756 } 1757 1758 static int virtnet_receive(struct receive_queue *rq, int budget, 1759 unsigned int *xdp_xmit) 1760 { 1761 struct virtnet_info *vi = rq->vq->vdev->priv; 1762 struct virtnet_rq_stats stats = {}; 1763 unsigned int len; 1764 void *buf; 1765 int i; 1766 1767 if (!vi->big_packets || vi->mergeable_rx_bufs) { 1768 void *ctx; 1769 1770 while (stats.packets < budget && 1771 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) { 1772 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats); 1773 stats.packets++; 1774 } 1775 } else { 1776 while (stats.packets < budget && 1777 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) { 1778 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats); 1779 stats.packets++; 1780 } 1781 } 1782 1783 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) { 1784 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) { 1785 spin_lock(&vi->refill_lock); 1786 if (vi->refill_enabled) 1787 schedule_delayed_work(&vi->refill, 0); 1788 spin_unlock(&vi->refill_lock); 1789 } 1790 } 1791 1792 u64_stats_update_begin(&rq->stats.syncp); 1793 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) { 1794 size_t offset = virtnet_rq_stats_desc[i].offset; 1795 u64 *item; 1796 1797 item = (u64 *)((u8 *)&rq->stats + offset); 1798 *item += *(u64 *)((u8 *)&stats + offset); 1799 } 1800 u64_stats_update_end(&rq->stats.syncp); 1801 1802 return stats.packets; 1803 } 1804 1805 static void virtnet_poll_cleantx(struct receive_queue *rq) 1806 { 1807 struct virtnet_info *vi = rq->vq->vdev->priv; 1808 unsigned int index = vq2rxq(rq->vq); 1809 struct send_queue *sq = &vi->sq[index]; 1810 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index); 1811 1812 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index)) 1813 return; 1814 1815 if (__netif_tx_trylock(txq)) { 1816 if (sq->reset) { 1817 __netif_tx_unlock(txq); 1818 return; 1819 } 1820 1821 do { 1822 virtqueue_disable_cb(sq->vq); 1823 free_old_xmit_skbs(sq, true); 1824 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq))); 1825 1826 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1827 netif_tx_wake_queue(txq); 1828 1829 __netif_tx_unlock(txq); 1830 } 1831 } 1832 1833 static int virtnet_poll(struct napi_struct *napi, int budget) 1834 { 1835 struct receive_queue *rq = 1836 container_of(napi, struct receive_queue, napi); 1837 struct virtnet_info *vi = rq->vq->vdev->priv; 1838 struct send_queue *sq; 1839 unsigned int received; 1840 unsigned int xdp_xmit = 0; 1841 1842 virtnet_poll_cleantx(rq); 1843 1844 received = virtnet_receive(rq, budget, &xdp_xmit); 1845 1846 if (xdp_xmit & VIRTIO_XDP_REDIR) 1847 xdp_do_flush(); 1848 1849 /* Out of packets? */ 1850 if (received < budget) 1851 virtqueue_napi_complete(napi, rq->vq, received); 1852 1853 if (xdp_xmit & VIRTIO_XDP_TX) { 1854 sq = virtnet_xdp_get_sq(vi); 1855 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) { 1856 u64_stats_update_begin(&sq->stats.syncp); 1857 sq->stats.kicks++; 1858 u64_stats_update_end(&sq->stats.syncp); 1859 } 1860 virtnet_xdp_put_sq(vi, sq); 1861 } 1862 1863 return received; 1864 } 1865 1866 static int virtnet_open(struct net_device *dev) 1867 { 1868 struct virtnet_info *vi = netdev_priv(dev); 1869 int i, err; 1870 1871 enable_delayed_refill(vi); 1872 1873 for (i = 0; i < vi->max_queue_pairs; i++) { 1874 if (i < vi->curr_queue_pairs) 1875 /* Make sure we have some buffers: if oom use wq. */ 1876 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 1877 schedule_delayed_work(&vi->refill, 0); 1878 1879 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id); 1880 if (err < 0) 1881 return err; 1882 1883 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq, 1884 MEM_TYPE_PAGE_SHARED, NULL); 1885 if (err < 0) { 1886 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 1887 return err; 1888 } 1889 1890 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 1891 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi); 1892 } 1893 1894 return 0; 1895 } 1896 1897 static int virtnet_poll_tx(struct napi_struct *napi, int budget) 1898 { 1899 struct send_queue *sq = container_of(napi, struct send_queue, napi); 1900 struct virtnet_info *vi = sq->vq->vdev->priv; 1901 unsigned int index = vq2txq(sq->vq); 1902 struct netdev_queue *txq; 1903 int opaque; 1904 bool done; 1905 1906 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) { 1907 /* We don't need to enable cb for XDP */ 1908 napi_complete_done(napi, 0); 1909 return 0; 1910 } 1911 1912 txq = netdev_get_tx_queue(vi->dev, index); 1913 __netif_tx_lock(txq, raw_smp_processor_id()); 1914 virtqueue_disable_cb(sq->vq); 1915 free_old_xmit_skbs(sq, true); 1916 1917 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1918 netif_tx_wake_queue(txq); 1919 1920 opaque = virtqueue_enable_cb_prepare(sq->vq); 1921 1922 done = napi_complete_done(napi, 0); 1923 1924 if (!done) 1925 virtqueue_disable_cb(sq->vq); 1926 1927 __netif_tx_unlock(txq); 1928 1929 if (done) { 1930 if (unlikely(virtqueue_poll(sq->vq, opaque))) { 1931 if (napi_schedule_prep(napi)) { 1932 __netif_tx_lock(txq, raw_smp_processor_id()); 1933 virtqueue_disable_cb(sq->vq); 1934 __netif_tx_unlock(txq); 1935 __napi_schedule(napi); 1936 } 1937 } 1938 } 1939 1940 return 0; 1941 } 1942 1943 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb) 1944 { 1945 struct virtio_net_hdr_mrg_rxbuf *hdr; 1946 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest; 1947 struct virtnet_info *vi = sq->vq->vdev->priv; 1948 int num_sg; 1949 unsigned hdr_len = vi->hdr_len; 1950 bool can_push; 1951 1952 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest); 1953 1954 can_push = vi->any_header_sg && 1955 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) && 1956 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len; 1957 /* Even if we can, don't push here yet as this would skew 1958 * csum_start offset below. */ 1959 if (can_push) 1960 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len); 1961 else 1962 hdr = skb_vnet_hdr(skb); 1963 1964 if (virtio_net_hdr_from_skb(skb, &hdr->hdr, 1965 virtio_is_little_endian(vi->vdev), false, 1966 0)) 1967 return -EPROTO; 1968 1969 if (vi->mergeable_rx_bufs) 1970 hdr->num_buffers = 0; 1971 1972 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2)); 1973 if (can_push) { 1974 __skb_push(skb, hdr_len); 1975 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len); 1976 if (unlikely(num_sg < 0)) 1977 return num_sg; 1978 /* Pull header back to avoid skew in tx bytes calculations. */ 1979 __skb_pull(skb, hdr_len); 1980 } else { 1981 sg_set_buf(sq->sg, hdr, hdr_len); 1982 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len); 1983 if (unlikely(num_sg < 0)) 1984 return num_sg; 1985 num_sg++; 1986 } 1987 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC); 1988 } 1989 1990 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) 1991 { 1992 struct virtnet_info *vi = netdev_priv(dev); 1993 int qnum = skb_get_queue_mapping(skb); 1994 struct send_queue *sq = &vi->sq[qnum]; 1995 int err; 1996 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum); 1997 bool kick = !netdev_xmit_more(); 1998 bool use_napi = sq->napi.weight; 1999 2000 /* Free up any pending old buffers before queueing new ones. */ 2001 do { 2002 if (use_napi) 2003 virtqueue_disable_cb(sq->vq); 2004 2005 free_old_xmit_skbs(sq, false); 2006 2007 } while (use_napi && kick && 2008 unlikely(!virtqueue_enable_cb_delayed(sq->vq))); 2009 2010 /* timestamp packet in software */ 2011 skb_tx_timestamp(skb); 2012 2013 /* Try to transmit */ 2014 err = xmit_skb(sq, skb); 2015 2016 /* This should not happen! */ 2017 if (unlikely(err)) { 2018 dev->stats.tx_fifo_errors++; 2019 if (net_ratelimit()) 2020 dev_warn(&dev->dev, 2021 "Unexpected TXQ (%d) queue failure: %d\n", 2022 qnum, err); 2023 dev->stats.tx_dropped++; 2024 dev_kfree_skb_any(skb); 2025 return NETDEV_TX_OK; 2026 } 2027 2028 /* Don't wait up for transmitted skbs to be freed. */ 2029 if (!use_napi) { 2030 skb_orphan(skb); 2031 nf_reset_ct(skb); 2032 } 2033 2034 check_sq_full_and_disable(vi, dev, sq); 2035 2036 if (kick || netif_xmit_stopped(txq)) { 2037 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) { 2038 u64_stats_update_begin(&sq->stats.syncp); 2039 sq->stats.kicks++; 2040 u64_stats_update_end(&sq->stats.syncp); 2041 } 2042 } 2043 2044 return NETDEV_TX_OK; 2045 } 2046 2047 static int virtnet_rx_resize(struct virtnet_info *vi, 2048 struct receive_queue *rq, u32 ring_num) 2049 { 2050 bool running = netif_running(vi->dev); 2051 int err, qindex; 2052 2053 qindex = rq - vi->rq; 2054 2055 if (running) 2056 napi_disable(&rq->napi); 2057 2058 err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_free_unused_buf); 2059 if (err) 2060 netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err); 2061 2062 if (!try_fill_recv(vi, rq, GFP_KERNEL)) 2063 schedule_delayed_work(&vi->refill, 0); 2064 2065 if (running) 2066 virtnet_napi_enable(rq->vq, &rq->napi); 2067 return err; 2068 } 2069 2070 static int virtnet_tx_resize(struct virtnet_info *vi, 2071 struct send_queue *sq, u32 ring_num) 2072 { 2073 bool running = netif_running(vi->dev); 2074 struct netdev_queue *txq; 2075 int err, qindex; 2076 2077 qindex = sq - vi->sq; 2078 2079 if (running) 2080 virtnet_napi_tx_disable(&sq->napi); 2081 2082 txq = netdev_get_tx_queue(vi->dev, qindex); 2083 2084 /* 1. wait all ximt complete 2085 * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue() 2086 */ 2087 __netif_tx_lock_bh(txq); 2088 2089 /* Prevent rx poll from accessing sq. */ 2090 sq->reset = true; 2091 2092 /* Prevent the upper layer from trying to send packets. */ 2093 netif_stop_subqueue(vi->dev, qindex); 2094 2095 __netif_tx_unlock_bh(txq); 2096 2097 err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf); 2098 if (err) 2099 netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err); 2100 2101 __netif_tx_lock_bh(txq); 2102 sq->reset = false; 2103 netif_tx_wake_queue(txq); 2104 __netif_tx_unlock_bh(txq); 2105 2106 if (running) 2107 virtnet_napi_tx_enable(vi, sq->vq, &sq->napi); 2108 return err; 2109 } 2110 2111 /* 2112 * Send command via the control virtqueue and check status. Commands 2113 * supported by the hypervisor, as indicated by feature bits, should 2114 * never fail unless improperly formatted. 2115 */ 2116 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd, 2117 struct scatterlist *out) 2118 { 2119 struct scatterlist *sgs[4], hdr, stat; 2120 unsigned out_num = 0, tmp; 2121 int ret; 2122 2123 /* Caller should know better */ 2124 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)); 2125 2126 vi->ctrl->status = ~0; 2127 vi->ctrl->hdr.class = class; 2128 vi->ctrl->hdr.cmd = cmd; 2129 /* Add header */ 2130 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr)); 2131 sgs[out_num++] = &hdr; 2132 2133 if (out) 2134 sgs[out_num++] = out; 2135 2136 /* Add return status. */ 2137 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status)); 2138 sgs[out_num] = &stat; 2139 2140 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs)); 2141 ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC); 2142 if (ret < 0) { 2143 dev_warn(&vi->vdev->dev, 2144 "Failed to add sgs for command vq: %d\n.", ret); 2145 return false; 2146 } 2147 2148 if (unlikely(!virtqueue_kick(vi->cvq))) 2149 return vi->ctrl->status == VIRTIO_NET_OK; 2150 2151 /* Spin for a response, the kick causes an ioport write, trapping 2152 * into the hypervisor, so the request should be handled immediately. 2153 */ 2154 while (!virtqueue_get_buf(vi->cvq, &tmp) && 2155 !virtqueue_is_broken(vi->cvq)) 2156 cpu_relax(); 2157 2158 return vi->ctrl->status == VIRTIO_NET_OK; 2159 } 2160 2161 static int virtnet_set_mac_address(struct net_device *dev, void *p) 2162 { 2163 struct virtnet_info *vi = netdev_priv(dev); 2164 struct virtio_device *vdev = vi->vdev; 2165 int ret; 2166 struct sockaddr *addr; 2167 struct scatterlist sg; 2168 2169 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY)) 2170 return -EOPNOTSUPP; 2171 2172 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL); 2173 if (!addr) 2174 return -ENOMEM; 2175 2176 ret = eth_prepare_mac_addr_change(dev, addr); 2177 if (ret) 2178 goto out; 2179 2180 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 2181 sg_init_one(&sg, addr->sa_data, dev->addr_len); 2182 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 2183 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) { 2184 dev_warn(&vdev->dev, 2185 "Failed to set mac address by vq command.\n"); 2186 ret = -EINVAL; 2187 goto out; 2188 } 2189 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) && 2190 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) { 2191 unsigned int i; 2192 2193 /* Naturally, this has an atomicity problem. */ 2194 for (i = 0; i < dev->addr_len; i++) 2195 virtio_cwrite8(vdev, 2196 offsetof(struct virtio_net_config, mac) + 2197 i, addr->sa_data[i]); 2198 } 2199 2200 eth_commit_mac_addr_change(dev, p); 2201 ret = 0; 2202 2203 out: 2204 kfree(addr); 2205 return ret; 2206 } 2207 2208 static void virtnet_stats(struct net_device *dev, 2209 struct rtnl_link_stats64 *tot) 2210 { 2211 struct virtnet_info *vi = netdev_priv(dev); 2212 unsigned int start; 2213 int i; 2214 2215 for (i = 0; i < vi->max_queue_pairs; i++) { 2216 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops; 2217 struct receive_queue *rq = &vi->rq[i]; 2218 struct send_queue *sq = &vi->sq[i]; 2219 2220 do { 2221 start = u64_stats_fetch_begin(&sq->stats.syncp); 2222 tpackets = sq->stats.packets; 2223 tbytes = sq->stats.bytes; 2224 terrors = sq->stats.tx_timeouts; 2225 } while (u64_stats_fetch_retry(&sq->stats.syncp, start)); 2226 2227 do { 2228 start = u64_stats_fetch_begin(&rq->stats.syncp); 2229 rpackets = rq->stats.packets; 2230 rbytes = rq->stats.bytes; 2231 rdrops = rq->stats.drops; 2232 } while (u64_stats_fetch_retry(&rq->stats.syncp, start)); 2233 2234 tot->rx_packets += rpackets; 2235 tot->tx_packets += tpackets; 2236 tot->rx_bytes += rbytes; 2237 tot->tx_bytes += tbytes; 2238 tot->rx_dropped += rdrops; 2239 tot->tx_errors += terrors; 2240 } 2241 2242 tot->tx_dropped = dev->stats.tx_dropped; 2243 tot->tx_fifo_errors = dev->stats.tx_fifo_errors; 2244 tot->rx_length_errors = dev->stats.rx_length_errors; 2245 tot->rx_frame_errors = dev->stats.rx_frame_errors; 2246 } 2247 2248 static void virtnet_ack_link_announce(struct virtnet_info *vi) 2249 { 2250 rtnl_lock(); 2251 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE, 2252 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL)) 2253 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n"); 2254 rtnl_unlock(); 2255 } 2256 2257 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 2258 { 2259 struct scatterlist sg; 2260 struct net_device *dev = vi->dev; 2261 2262 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ)) 2263 return 0; 2264 2265 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs); 2266 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq)); 2267 2268 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ, 2269 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) { 2270 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n", 2271 queue_pairs); 2272 return -EINVAL; 2273 } else { 2274 vi->curr_queue_pairs = queue_pairs; 2275 /* virtnet_open() will refill when device is going to up. */ 2276 if (dev->flags & IFF_UP) 2277 schedule_delayed_work(&vi->refill, 0); 2278 } 2279 2280 return 0; 2281 } 2282 2283 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 2284 { 2285 int err; 2286 2287 rtnl_lock(); 2288 err = _virtnet_set_queues(vi, queue_pairs); 2289 rtnl_unlock(); 2290 return err; 2291 } 2292 2293 static int virtnet_close(struct net_device *dev) 2294 { 2295 struct virtnet_info *vi = netdev_priv(dev); 2296 int i; 2297 2298 /* Make sure NAPI doesn't schedule refill work */ 2299 disable_delayed_refill(vi); 2300 /* Make sure refill_work doesn't re-enable napi! */ 2301 cancel_delayed_work_sync(&vi->refill); 2302 2303 for (i = 0; i < vi->max_queue_pairs; i++) { 2304 virtnet_napi_tx_disable(&vi->sq[i].napi); 2305 napi_disable(&vi->rq[i].napi); 2306 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 2307 } 2308 2309 return 0; 2310 } 2311 2312 static void virtnet_set_rx_mode(struct net_device *dev) 2313 { 2314 struct virtnet_info *vi = netdev_priv(dev); 2315 struct scatterlist sg[2]; 2316 struct virtio_net_ctrl_mac *mac_data; 2317 struct netdev_hw_addr *ha; 2318 int uc_count; 2319 int mc_count; 2320 void *buf; 2321 int i; 2322 2323 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */ 2324 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX)) 2325 return; 2326 2327 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0); 2328 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0); 2329 2330 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc)); 2331 2332 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 2333 VIRTIO_NET_CTRL_RX_PROMISC, sg)) 2334 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n", 2335 vi->ctrl->promisc ? "en" : "dis"); 2336 2337 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti)); 2338 2339 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 2340 VIRTIO_NET_CTRL_RX_ALLMULTI, sg)) 2341 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n", 2342 vi->ctrl->allmulti ? "en" : "dis"); 2343 2344 uc_count = netdev_uc_count(dev); 2345 mc_count = netdev_mc_count(dev); 2346 /* MAC filter - use one buffer for both lists */ 2347 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) + 2348 (2 * sizeof(mac_data->entries)), GFP_ATOMIC); 2349 mac_data = buf; 2350 if (!buf) 2351 return; 2352 2353 sg_init_table(sg, 2); 2354 2355 /* Store the unicast list and count in the front of the buffer */ 2356 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count); 2357 i = 0; 2358 netdev_for_each_uc_addr(ha, dev) 2359 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 2360 2361 sg_set_buf(&sg[0], mac_data, 2362 sizeof(mac_data->entries) + (uc_count * ETH_ALEN)); 2363 2364 /* multicast list and count fill the end */ 2365 mac_data = (void *)&mac_data->macs[uc_count][0]; 2366 2367 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count); 2368 i = 0; 2369 netdev_for_each_mc_addr(ha, dev) 2370 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 2371 2372 sg_set_buf(&sg[1], mac_data, 2373 sizeof(mac_data->entries) + (mc_count * ETH_ALEN)); 2374 2375 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 2376 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg)) 2377 dev_warn(&dev->dev, "Failed to set MAC filter table.\n"); 2378 2379 kfree(buf); 2380 } 2381 2382 static int virtnet_vlan_rx_add_vid(struct net_device *dev, 2383 __be16 proto, u16 vid) 2384 { 2385 struct virtnet_info *vi = netdev_priv(dev); 2386 struct scatterlist sg; 2387 2388 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid); 2389 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid)); 2390 2391 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 2392 VIRTIO_NET_CTRL_VLAN_ADD, &sg)) 2393 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid); 2394 return 0; 2395 } 2396 2397 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, 2398 __be16 proto, u16 vid) 2399 { 2400 struct virtnet_info *vi = netdev_priv(dev); 2401 struct scatterlist sg; 2402 2403 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid); 2404 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid)); 2405 2406 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 2407 VIRTIO_NET_CTRL_VLAN_DEL, &sg)) 2408 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid); 2409 return 0; 2410 } 2411 2412 static void virtnet_clean_affinity(struct virtnet_info *vi) 2413 { 2414 int i; 2415 2416 if (vi->affinity_hint_set) { 2417 for (i = 0; i < vi->max_queue_pairs; i++) { 2418 virtqueue_set_affinity(vi->rq[i].vq, NULL); 2419 virtqueue_set_affinity(vi->sq[i].vq, NULL); 2420 } 2421 2422 vi->affinity_hint_set = false; 2423 } 2424 } 2425 2426 static void virtnet_set_affinity(struct virtnet_info *vi) 2427 { 2428 cpumask_var_t mask; 2429 int stragglers; 2430 int group_size; 2431 int i, j, cpu; 2432 int num_cpu; 2433 int stride; 2434 2435 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) { 2436 virtnet_clean_affinity(vi); 2437 return; 2438 } 2439 2440 num_cpu = num_online_cpus(); 2441 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1); 2442 stragglers = num_cpu >= vi->curr_queue_pairs ? 2443 num_cpu % vi->curr_queue_pairs : 2444 0; 2445 cpu = cpumask_first(cpu_online_mask); 2446 2447 for (i = 0; i < vi->curr_queue_pairs; i++) { 2448 group_size = stride + (i < stragglers ? 1 : 0); 2449 2450 for (j = 0; j < group_size; j++) { 2451 cpumask_set_cpu(cpu, mask); 2452 cpu = cpumask_next_wrap(cpu, cpu_online_mask, 2453 nr_cpu_ids, false); 2454 } 2455 virtqueue_set_affinity(vi->rq[i].vq, mask); 2456 virtqueue_set_affinity(vi->sq[i].vq, mask); 2457 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS); 2458 cpumask_clear(mask); 2459 } 2460 2461 vi->affinity_hint_set = true; 2462 free_cpumask_var(mask); 2463 } 2464 2465 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node) 2466 { 2467 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 2468 node); 2469 virtnet_set_affinity(vi); 2470 return 0; 2471 } 2472 2473 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node) 2474 { 2475 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 2476 node_dead); 2477 virtnet_set_affinity(vi); 2478 return 0; 2479 } 2480 2481 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node) 2482 { 2483 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 2484 node); 2485 2486 virtnet_clean_affinity(vi); 2487 return 0; 2488 } 2489 2490 static enum cpuhp_state virtionet_online; 2491 2492 static int virtnet_cpu_notif_add(struct virtnet_info *vi) 2493 { 2494 int ret; 2495 2496 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node); 2497 if (ret) 2498 return ret; 2499 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD, 2500 &vi->node_dead); 2501 if (!ret) 2502 return ret; 2503 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 2504 return ret; 2505 } 2506 2507 static void virtnet_cpu_notif_remove(struct virtnet_info *vi) 2508 { 2509 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 2510 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD, 2511 &vi->node_dead); 2512 } 2513 2514 static void virtnet_get_ringparam(struct net_device *dev, 2515 struct ethtool_ringparam *ring, 2516 struct kernel_ethtool_ringparam *kernel_ring, 2517 struct netlink_ext_ack *extack) 2518 { 2519 struct virtnet_info *vi = netdev_priv(dev); 2520 2521 ring->rx_max_pending = vi->rq[0].vq->num_max; 2522 ring->tx_max_pending = vi->sq[0].vq->num_max; 2523 ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq); 2524 ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq); 2525 } 2526 2527 static int virtnet_set_ringparam(struct net_device *dev, 2528 struct ethtool_ringparam *ring, 2529 struct kernel_ethtool_ringparam *kernel_ring, 2530 struct netlink_ext_ack *extack) 2531 { 2532 struct virtnet_info *vi = netdev_priv(dev); 2533 u32 rx_pending, tx_pending; 2534 struct receive_queue *rq; 2535 struct send_queue *sq; 2536 int i, err; 2537 2538 if (ring->rx_mini_pending || ring->rx_jumbo_pending) 2539 return -EINVAL; 2540 2541 rx_pending = virtqueue_get_vring_size(vi->rq[0].vq); 2542 tx_pending = virtqueue_get_vring_size(vi->sq[0].vq); 2543 2544 if (ring->rx_pending == rx_pending && 2545 ring->tx_pending == tx_pending) 2546 return 0; 2547 2548 if (ring->rx_pending > vi->rq[0].vq->num_max) 2549 return -EINVAL; 2550 2551 if (ring->tx_pending > vi->sq[0].vq->num_max) 2552 return -EINVAL; 2553 2554 for (i = 0; i < vi->max_queue_pairs; i++) { 2555 rq = vi->rq + i; 2556 sq = vi->sq + i; 2557 2558 if (ring->tx_pending != tx_pending) { 2559 err = virtnet_tx_resize(vi, sq, ring->tx_pending); 2560 if (err) 2561 return err; 2562 } 2563 2564 if (ring->rx_pending != rx_pending) { 2565 err = virtnet_rx_resize(vi, rq, ring->rx_pending); 2566 if (err) 2567 return err; 2568 } 2569 } 2570 2571 return 0; 2572 } 2573 2574 static bool virtnet_commit_rss_command(struct virtnet_info *vi) 2575 { 2576 struct net_device *dev = vi->dev; 2577 struct scatterlist sgs[4]; 2578 unsigned int sg_buf_size; 2579 2580 /* prepare sgs */ 2581 sg_init_table(sgs, 4); 2582 2583 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table); 2584 sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size); 2585 2586 sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1); 2587 sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size); 2588 2589 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key) 2590 - offsetof(struct virtio_net_ctrl_rss, max_tx_vq); 2591 sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size); 2592 2593 sg_buf_size = vi->rss_key_size; 2594 sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size); 2595 2596 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ, 2597 vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG 2598 : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) { 2599 dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n"); 2600 return false; 2601 } 2602 return true; 2603 } 2604 2605 static void virtnet_init_default_rss(struct virtnet_info *vi) 2606 { 2607 u32 indir_val = 0; 2608 int i = 0; 2609 2610 vi->ctrl->rss.hash_types = vi->rss_hash_types_supported; 2611 vi->rss_hash_types_saved = vi->rss_hash_types_supported; 2612 vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size 2613 ? vi->rss_indir_table_size - 1 : 0; 2614 vi->ctrl->rss.unclassified_queue = 0; 2615 2616 for (; i < vi->rss_indir_table_size; ++i) { 2617 indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs); 2618 vi->ctrl->rss.indirection_table[i] = indir_val; 2619 } 2620 2621 vi->ctrl->rss.max_tx_vq = vi->curr_queue_pairs; 2622 vi->ctrl->rss.hash_key_length = vi->rss_key_size; 2623 2624 netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size); 2625 } 2626 2627 static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info) 2628 { 2629 info->data = 0; 2630 switch (info->flow_type) { 2631 case TCP_V4_FLOW: 2632 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) { 2633 info->data = RXH_IP_SRC | RXH_IP_DST | 2634 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2635 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) { 2636 info->data = RXH_IP_SRC | RXH_IP_DST; 2637 } 2638 break; 2639 case TCP_V6_FLOW: 2640 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) { 2641 info->data = RXH_IP_SRC | RXH_IP_DST | 2642 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2643 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) { 2644 info->data = RXH_IP_SRC | RXH_IP_DST; 2645 } 2646 break; 2647 case UDP_V4_FLOW: 2648 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) { 2649 info->data = RXH_IP_SRC | RXH_IP_DST | 2650 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2651 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) { 2652 info->data = RXH_IP_SRC | RXH_IP_DST; 2653 } 2654 break; 2655 case UDP_V6_FLOW: 2656 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) { 2657 info->data = RXH_IP_SRC | RXH_IP_DST | 2658 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2659 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) { 2660 info->data = RXH_IP_SRC | RXH_IP_DST; 2661 } 2662 break; 2663 case IPV4_FLOW: 2664 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) 2665 info->data = RXH_IP_SRC | RXH_IP_DST; 2666 2667 break; 2668 case IPV6_FLOW: 2669 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) 2670 info->data = RXH_IP_SRC | RXH_IP_DST; 2671 2672 break; 2673 default: 2674 info->data = 0; 2675 break; 2676 } 2677 } 2678 2679 static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info) 2680 { 2681 u32 new_hashtypes = vi->rss_hash_types_saved; 2682 bool is_disable = info->data & RXH_DISCARD; 2683 bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3); 2684 2685 /* supports only 'sd', 'sdfn' and 'r' */ 2686 if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable)) 2687 return false; 2688 2689 switch (info->flow_type) { 2690 case TCP_V4_FLOW: 2691 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4); 2692 if (!is_disable) 2693 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4 2694 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0); 2695 break; 2696 case UDP_V4_FLOW: 2697 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4); 2698 if (!is_disable) 2699 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4 2700 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0); 2701 break; 2702 case IPV4_FLOW: 2703 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4; 2704 if (!is_disable) 2705 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4; 2706 break; 2707 case TCP_V6_FLOW: 2708 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6); 2709 if (!is_disable) 2710 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6 2711 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0); 2712 break; 2713 case UDP_V6_FLOW: 2714 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6); 2715 if (!is_disable) 2716 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6 2717 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0); 2718 break; 2719 case IPV6_FLOW: 2720 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6; 2721 if (!is_disable) 2722 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6; 2723 break; 2724 default: 2725 /* unsupported flow */ 2726 return false; 2727 } 2728 2729 /* if unsupported hashtype was set */ 2730 if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported)) 2731 return false; 2732 2733 if (new_hashtypes != vi->rss_hash_types_saved) { 2734 vi->rss_hash_types_saved = new_hashtypes; 2735 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved; 2736 if (vi->dev->features & NETIF_F_RXHASH) 2737 return virtnet_commit_rss_command(vi); 2738 } 2739 2740 return true; 2741 } 2742 2743 static void virtnet_get_drvinfo(struct net_device *dev, 2744 struct ethtool_drvinfo *info) 2745 { 2746 struct virtnet_info *vi = netdev_priv(dev); 2747 struct virtio_device *vdev = vi->vdev; 2748 2749 strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver)); 2750 strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version)); 2751 strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info)); 2752 2753 } 2754 2755 /* TODO: Eliminate OOO packets during switching */ 2756 static int virtnet_set_channels(struct net_device *dev, 2757 struct ethtool_channels *channels) 2758 { 2759 struct virtnet_info *vi = netdev_priv(dev); 2760 u16 queue_pairs = channels->combined_count; 2761 int err; 2762 2763 /* We don't support separate rx/tx channels. 2764 * We don't allow setting 'other' channels. 2765 */ 2766 if (channels->rx_count || channels->tx_count || channels->other_count) 2767 return -EINVAL; 2768 2769 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0) 2770 return -EINVAL; 2771 2772 /* For now we don't support modifying channels while XDP is loaded 2773 * also when XDP is loaded all RX queues have XDP programs so we only 2774 * need to check a single RX queue. 2775 */ 2776 if (vi->rq[0].xdp_prog) 2777 return -EINVAL; 2778 2779 cpus_read_lock(); 2780 err = _virtnet_set_queues(vi, queue_pairs); 2781 if (err) { 2782 cpus_read_unlock(); 2783 goto err; 2784 } 2785 virtnet_set_affinity(vi); 2786 cpus_read_unlock(); 2787 2788 netif_set_real_num_tx_queues(dev, queue_pairs); 2789 netif_set_real_num_rx_queues(dev, queue_pairs); 2790 err: 2791 return err; 2792 } 2793 2794 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data) 2795 { 2796 struct virtnet_info *vi = netdev_priv(dev); 2797 unsigned int i, j; 2798 u8 *p = data; 2799 2800 switch (stringset) { 2801 case ETH_SS_STATS: 2802 for (i = 0; i < vi->curr_queue_pairs; i++) { 2803 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) 2804 ethtool_sprintf(&p, "rx_queue_%u_%s", i, 2805 virtnet_rq_stats_desc[j].desc); 2806 } 2807 2808 for (i = 0; i < vi->curr_queue_pairs; i++) { 2809 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) 2810 ethtool_sprintf(&p, "tx_queue_%u_%s", i, 2811 virtnet_sq_stats_desc[j].desc); 2812 } 2813 break; 2814 } 2815 } 2816 2817 static int virtnet_get_sset_count(struct net_device *dev, int sset) 2818 { 2819 struct virtnet_info *vi = netdev_priv(dev); 2820 2821 switch (sset) { 2822 case ETH_SS_STATS: 2823 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN + 2824 VIRTNET_SQ_STATS_LEN); 2825 default: 2826 return -EOPNOTSUPP; 2827 } 2828 } 2829 2830 static void virtnet_get_ethtool_stats(struct net_device *dev, 2831 struct ethtool_stats *stats, u64 *data) 2832 { 2833 struct virtnet_info *vi = netdev_priv(dev); 2834 unsigned int idx = 0, start, i, j; 2835 const u8 *stats_base; 2836 size_t offset; 2837 2838 for (i = 0; i < vi->curr_queue_pairs; i++) { 2839 struct receive_queue *rq = &vi->rq[i]; 2840 2841 stats_base = (u8 *)&rq->stats; 2842 do { 2843 start = u64_stats_fetch_begin(&rq->stats.syncp); 2844 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 2845 offset = virtnet_rq_stats_desc[j].offset; 2846 data[idx + j] = *(u64 *)(stats_base + offset); 2847 } 2848 } while (u64_stats_fetch_retry(&rq->stats.syncp, start)); 2849 idx += VIRTNET_RQ_STATS_LEN; 2850 } 2851 2852 for (i = 0; i < vi->curr_queue_pairs; i++) { 2853 struct send_queue *sq = &vi->sq[i]; 2854 2855 stats_base = (u8 *)&sq->stats; 2856 do { 2857 start = u64_stats_fetch_begin(&sq->stats.syncp); 2858 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 2859 offset = virtnet_sq_stats_desc[j].offset; 2860 data[idx + j] = *(u64 *)(stats_base + offset); 2861 } 2862 } while (u64_stats_fetch_retry(&sq->stats.syncp, start)); 2863 idx += VIRTNET_SQ_STATS_LEN; 2864 } 2865 } 2866 2867 static void virtnet_get_channels(struct net_device *dev, 2868 struct ethtool_channels *channels) 2869 { 2870 struct virtnet_info *vi = netdev_priv(dev); 2871 2872 channels->combined_count = vi->curr_queue_pairs; 2873 channels->max_combined = vi->max_queue_pairs; 2874 channels->max_other = 0; 2875 channels->rx_count = 0; 2876 channels->tx_count = 0; 2877 channels->other_count = 0; 2878 } 2879 2880 static int virtnet_set_link_ksettings(struct net_device *dev, 2881 const struct ethtool_link_ksettings *cmd) 2882 { 2883 struct virtnet_info *vi = netdev_priv(dev); 2884 2885 return ethtool_virtdev_set_link_ksettings(dev, cmd, 2886 &vi->speed, &vi->duplex); 2887 } 2888 2889 static int virtnet_get_link_ksettings(struct net_device *dev, 2890 struct ethtool_link_ksettings *cmd) 2891 { 2892 struct virtnet_info *vi = netdev_priv(dev); 2893 2894 cmd->base.speed = vi->speed; 2895 cmd->base.duplex = vi->duplex; 2896 cmd->base.port = PORT_OTHER; 2897 2898 return 0; 2899 } 2900 2901 static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi, 2902 struct ethtool_coalesce *ec) 2903 { 2904 struct scatterlist sgs_tx, sgs_rx; 2905 struct virtio_net_ctrl_coal_tx coal_tx; 2906 struct virtio_net_ctrl_coal_rx coal_rx; 2907 2908 coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs); 2909 coal_tx.tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames); 2910 sg_init_one(&sgs_tx, &coal_tx, sizeof(coal_tx)); 2911 2912 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL, 2913 VIRTIO_NET_CTRL_NOTF_COAL_TX_SET, 2914 &sgs_tx)) 2915 return -EINVAL; 2916 2917 /* Save parameters */ 2918 vi->tx_usecs = ec->tx_coalesce_usecs; 2919 vi->tx_max_packets = ec->tx_max_coalesced_frames; 2920 2921 coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs); 2922 coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames); 2923 sg_init_one(&sgs_rx, &coal_rx, sizeof(coal_rx)); 2924 2925 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL, 2926 VIRTIO_NET_CTRL_NOTF_COAL_RX_SET, 2927 &sgs_rx)) 2928 return -EINVAL; 2929 2930 /* Save parameters */ 2931 vi->rx_usecs = ec->rx_coalesce_usecs; 2932 vi->rx_max_packets = ec->rx_max_coalesced_frames; 2933 2934 return 0; 2935 } 2936 2937 static int virtnet_coal_params_supported(struct ethtool_coalesce *ec) 2938 { 2939 /* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL 2940 * feature is negotiated. 2941 */ 2942 if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs) 2943 return -EOPNOTSUPP; 2944 2945 if (ec->tx_max_coalesced_frames > 1 || 2946 ec->rx_max_coalesced_frames != 1) 2947 return -EINVAL; 2948 2949 return 0; 2950 } 2951 2952 static int virtnet_set_coalesce(struct net_device *dev, 2953 struct ethtool_coalesce *ec, 2954 struct kernel_ethtool_coalesce *kernel_coal, 2955 struct netlink_ext_ack *extack) 2956 { 2957 struct virtnet_info *vi = netdev_priv(dev); 2958 int ret, i, napi_weight; 2959 bool update_napi = false; 2960 2961 /* Can't change NAPI weight if the link is up */ 2962 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0; 2963 if (napi_weight ^ vi->sq[0].napi.weight) { 2964 if (dev->flags & IFF_UP) 2965 return -EBUSY; 2966 else 2967 update_napi = true; 2968 } 2969 2970 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) 2971 ret = virtnet_send_notf_coal_cmds(vi, ec); 2972 else 2973 ret = virtnet_coal_params_supported(ec); 2974 2975 if (ret) 2976 return ret; 2977 2978 if (update_napi) { 2979 for (i = 0; i < vi->max_queue_pairs; i++) 2980 vi->sq[i].napi.weight = napi_weight; 2981 } 2982 2983 return ret; 2984 } 2985 2986 static int virtnet_get_coalesce(struct net_device *dev, 2987 struct ethtool_coalesce *ec, 2988 struct kernel_ethtool_coalesce *kernel_coal, 2989 struct netlink_ext_ack *extack) 2990 { 2991 struct virtnet_info *vi = netdev_priv(dev); 2992 2993 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) { 2994 ec->rx_coalesce_usecs = vi->rx_usecs; 2995 ec->tx_coalesce_usecs = vi->tx_usecs; 2996 ec->tx_max_coalesced_frames = vi->tx_max_packets; 2997 ec->rx_max_coalesced_frames = vi->rx_max_packets; 2998 } else { 2999 ec->rx_max_coalesced_frames = 1; 3000 3001 if (vi->sq[0].napi.weight) 3002 ec->tx_max_coalesced_frames = 1; 3003 } 3004 3005 return 0; 3006 } 3007 3008 static void virtnet_init_settings(struct net_device *dev) 3009 { 3010 struct virtnet_info *vi = netdev_priv(dev); 3011 3012 vi->speed = SPEED_UNKNOWN; 3013 vi->duplex = DUPLEX_UNKNOWN; 3014 } 3015 3016 static void virtnet_update_settings(struct virtnet_info *vi) 3017 { 3018 u32 speed; 3019 u8 duplex; 3020 3021 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX)) 3022 return; 3023 3024 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed); 3025 3026 if (ethtool_validate_speed(speed)) 3027 vi->speed = speed; 3028 3029 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex); 3030 3031 if (ethtool_validate_duplex(duplex)) 3032 vi->duplex = duplex; 3033 } 3034 3035 static u32 virtnet_get_rxfh_key_size(struct net_device *dev) 3036 { 3037 return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size; 3038 } 3039 3040 static u32 virtnet_get_rxfh_indir_size(struct net_device *dev) 3041 { 3042 return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size; 3043 } 3044 3045 static int virtnet_get_rxfh(struct net_device *dev, u32 *indir, u8 *key, u8 *hfunc) 3046 { 3047 struct virtnet_info *vi = netdev_priv(dev); 3048 int i; 3049 3050 if (indir) { 3051 for (i = 0; i < vi->rss_indir_table_size; ++i) 3052 indir[i] = vi->ctrl->rss.indirection_table[i]; 3053 } 3054 3055 if (key) 3056 memcpy(key, vi->ctrl->rss.key, vi->rss_key_size); 3057 3058 if (hfunc) 3059 *hfunc = ETH_RSS_HASH_TOP; 3060 3061 return 0; 3062 } 3063 3064 static int virtnet_set_rxfh(struct net_device *dev, const u32 *indir, const u8 *key, const u8 hfunc) 3065 { 3066 struct virtnet_info *vi = netdev_priv(dev); 3067 int i; 3068 3069 if (hfunc != ETH_RSS_HASH_NO_CHANGE && hfunc != ETH_RSS_HASH_TOP) 3070 return -EOPNOTSUPP; 3071 3072 if (indir) { 3073 for (i = 0; i < vi->rss_indir_table_size; ++i) 3074 vi->ctrl->rss.indirection_table[i] = indir[i]; 3075 } 3076 if (key) 3077 memcpy(vi->ctrl->rss.key, key, vi->rss_key_size); 3078 3079 virtnet_commit_rss_command(vi); 3080 3081 return 0; 3082 } 3083 3084 static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs) 3085 { 3086 struct virtnet_info *vi = netdev_priv(dev); 3087 int rc = 0; 3088 3089 switch (info->cmd) { 3090 case ETHTOOL_GRXRINGS: 3091 info->data = vi->curr_queue_pairs; 3092 break; 3093 case ETHTOOL_GRXFH: 3094 virtnet_get_hashflow(vi, info); 3095 break; 3096 default: 3097 rc = -EOPNOTSUPP; 3098 } 3099 3100 return rc; 3101 } 3102 3103 static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info) 3104 { 3105 struct virtnet_info *vi = netdev_priv(dev); 3106 int rc = 0; 3107 3108 switch (info->cmd) { 3109 case ETHTOOL_SRXFH: 3110 if (!virtnet_set_hashflow(vi, info)) 3111 rc = -EINVAL; 3112 3113 break; 3114 default: 3115 rc = -EOPNOTSUPP; 3116 } 3117 3118 return rc; 3119 } 3120 3121 static const struct ethtool_ops virtnet_ethtool_ops = { 3122 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES | 3123 ETHTOOL_COALESCE_USECS, 3124 .get_drvinfo = virtnet_get_drvinfo, 3125 .get_link = ethtool_op_get_link, 3126 .get_ringparam = virtnet_get_ringparam, 3127 .set_ringparam = virtnet_set_ringparam, 3128 .get_strings = virtnet_get_strings, 3129 .get_sset_count = virtnet_get_sset_count, 3130 .get_ethtool_stats = virtnet_get_ethtool_stats, 3131 .set_channels = virtnet_set_channels, 3132 .get_channels = virtnet_get_channels, 3133 .get_ts_info = ethtool_op_get_ts_info, 3134 .get_link_ksettings = virtnet_get_link_ksettings, 3135 .set_link_ksettings = virtnet_set_link_ksettings, 3136 .set_coalesce = virtnet_set_coalesce, 3137 .get_coalesce = virtnet_get_coalesce, 3138 .get_rxfh_key_size = virtnet_get_rxfh_key_size, 3139 .get_rxfh_indir_size = virtnet_get_rxfh_indir_size, 3140 .get_rxfh = virtnet_get_rxfh, 3141 .set_rxfh = virtnet_set_rxfh, 3142 .get_rxnfc = virtnet_get_rxnfc, 3143 .set_rxnfc = virtnet_set_rxnfc, 3144 }; 3145 3146 static void virtnet_freeze_down(struct virtio_device *vdev) 3147 { 3148 struct virtnet_info *vi = vdev->priv; 3149 3150 /* Make sure no work handler is accessing the device */ 3151 flush_work(&vi->config_work); 3152 3153 netif_tx_lock_bh(vi->dev); 3154 netif_device_detach(vi->dev); 3155 netif_tx_unlock_bh(vi->dev); 3156 if (netif_running(vi->dev)) 3157 virtnet_close(vi->dev); 3158 } 3159 3160 static int init_vqs(struct virtnet_info *vi); 3161 3162 static int virtnet_restore_up(struct virtio_device *vdev) 3163 { 3164 struct virtnet_info *vi = vdev->priv; 3165 int err; 3166 3167 err = init_vqs(vi); 3168 if (err) 3169 return err; 3170 3171 virtio_device_ready(vdev); 3172 3173 enable_delayed_refill(vi); 3174 3175 if (netif_running(vi->dev)) { 3176 err = virtnet_open(vi->dev); 3177 if (err) 3178 return err; 3179 } 3180 3181 netif_tx_lock_bh(vi->dev); 3182 netif_device_attach(vi->dev); 3183 netif_tx_unlock_bh(vi->dev); 3184 return err; 3185 } 3186 3187 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads) 3188 { 3189 struct scatterlist sg; 3190 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads); 3191 3192 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads)); 3193 3194 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS, 3195 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) { 3196 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n"); 3197 return -EINVAL; 3198 } 3199 3200 return 0; 3201 } 3202 3203 static int virtnet_clear_guest_offloads(struct virtnet_info *vi) 3204 { 3205 u64 offloads = 0; 3206 3207 if (!vi->guest_offloads) 3208 return 0; 3209 3210 return virtnet_set_guest_offloads(vi, offloads); 3211 } 3212 3213 static int virtnet_restore_guest_offloads(struct virtnet_info *vi) 3214 { 3215 u64 offloads = vi->guest_offloads; 3216 3217 if (!vi->guest_offloads) 3218 return 0; 3219 3220 return virtnet_set_guest_offloads(vi, offloads); 3221 } 3222 3223 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog, 3224 struct netlink_ext_ack *extack) 3225 { 3226 unsigned int room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM + 3227 sizeof(struct skb_shared_info)); 3228 unsigned int max_sz = PAGE_SIZE - room - ETH_HLEN; 3229 struct virtnet_info *vi = netdev_priv(dev); 3230 struct bpf_prog *old_prog; 3231 u16 xdp_qp = 0, curr_qp; 3232 int i, err; 3233 3234 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) 3235 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) || 3236 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) || 3237 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) || 3238 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) || 3239 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM) || 3240 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) || 3241 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6))) { 3242 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first"); 3243 return -EOPNOTSUPP; 3244 } 3245 3246 if (vi->mergeable_rx_bufs && !vi->any_header_sg) { 3247 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required"); 3248 return -EINVAL; 3249 } 3250 3251 if (prog && !prog->aux->xdp_has_frags && dev->mtu > max_sz) { 3252 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP without frags"); 3253 netdev_warn(dev, "single-buffer XDP requires MTU less than %u\n", max_sz); 3254 return -EINVAL; 3255 } 3256 3257 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs; 3258 if (prog) 3259 xdp_qp = nr_cpu_ids; 3260 3261 /* XDP requires extra queues for XDP_TX */ 3262 if (curr_qp + xdp_qp > vi->max_queue_pairs) { 3263 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", 3264 curr_qp + xdp_qp, vi->max_queue_pairs); 3265 xdp_qp = 0; 3266 } 3267 3268 old_prog = rtnl_dereference(vi->rq[0].xdp_prog); 3269 if (!prog && !old_prog) 3270 return 0; 3271 3272 if (prog) 3273 bpf_prog_add(prog, vi->max_queue_pairs - 1); 3274 3275 /* Make sure NAPI is not using any XDP TX queues for RX. */ 3276 if (netif_running(dev)) { 3277 for (i = 0; i < vi->max_queue_pairs; i++) { 3278 napi_disable(&vi->rq[i].napi); 3279 virtnet_napi_tx_disable(&vi->sq[i].napi); 3280 } 3281 } 3282 3283 if (!prog) { 3284 for (i = 0; i < vi->max_queue_pairs; i++) { 3285 rcu_assign_pointer(vi->rq[i].xdp_prog, prog); 3286 if (i == 0) 3287 virtnet_restore_guest_offloads(vi); 3288 } 3289 synchronize_net(); 3290 } 3291 3292 err = _virtnet_set_queues(vi, curr_qp + xdp_qp); 3293 if (err) 3294 goto err; 3295 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp); 3296 vi->xdp_queue_pairs = xdp_qp; 3297 3298 if (prog) { 3299 vi->xdp_enabled = true; 3300 for (i = 0; i < vi->max_queue_pairs; i++) { 3301 rcu_assign_pointer(vi->rq[i].xdp_prog, prog); 3302 if (i == 0 && !old_prog) 3303 virtnet_clear_guest_offloads(vi); 3304 } 3305 if (!old_prog) 3306 xdp_features_set_redirect_target(dev, true); 3307 } else { 3308 xdp_features_clear_redirect_target(dev); 3309 vi->xdp_enabled = false; 3310 } 3311 3312 for (i = 0; i < vi->max_queue_pairs; i++) { 3313 if (old_prog) 3314 bpf_prog_put(old_prog); 3315 if (netif_running(dev)) { 3316 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 3317 virtnet_napi_tx_enable(vi, vi->sq[i].vq, 3318 &vi->sq[i].napi); 3319 } 3320 } 3321 3322 return 0; 3323 3324 err: 3325 if (!prog) { 3326 virtnet_clear_guest_offloads(vi); 3327 for (i = 0; i < vi->max_queue_pairs; i++) 3328 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog); 3329 } 3330 3331 if (netif_running(dev)) { 3332 for (i = 0; i < vi->max_queue_pairs; i++) { 3333 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 3334 virtnet_napi_tx_enable(vi, vi->sq[i].vq, 3335 &vi->sq[i].napi); 3336 } 3337 } 3338 if (prog) 3339 bpf_prog_sub(prog, vi->max_queue_pairs - 1); 3340 return err; 3341 } 3342 3343 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp) 3344 { 3345 switch (xdp->command) { 3346 case XDP_SETUP_PROG: 3347 return virtnet_xdp_set(dev, xdp->prog, xdp->extack); 3348 default: 3349 return -EINVAL; 3350 } 3351 } 3352 3353 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf, 3354 size_t len) 3355 { 3356 struct virtnet_info *vi = netdev_priv(dev); 3357 int ret; 3358 3359 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY)) 3360 return -EOPNOTSUPP; 3361 3362 ret = snprintf(buf, len, "sby"); 3363 if (ret >= len) 3364 return -EOPNOTSUPP; 3365 3366 return 0; 3367 } 3368 3369 static int virtnet_set_features(struct net_device *dev, 3370 netdev_features_t features) 3371 { 3372 struct virtnet_info *vi = netdev_priv(dev); 3373 u64 offloads; 3374 int err; 3375 3376 if ((dev->features ^ features) & NETIF_F_GRO_HW) { 3377 if (vi->xdp_enabled) 3378 return -EBUSY; 3379 3380 if (features & NETIF_F_GRO_HW) 3381 offloads = vi->guest_offloads_capable; 3382 else 3383 offloads = vi->guest_offloads_capable & 3384 ~GUEST_OFFLOAD_GRO_HW_MASK; 3385 3386 err = virtnet_set_guest_offloads(vi, offloads); 3387 if (err) 3388 return err; 3389 vi->guest_offloads = offloads; 3390 } 3391 3392 if ((dev->features ^ features) & NETIF_F_RXHASH) { 3393 if (features & NETIF_F_RXHASH) 3394 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved; 3395 else 3396 vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE; 3397 3398 if (!virtnet_commit_rss_command(vi)) 3399 return -EINVAL; 3400 } 3401 3402 return 0; 3403 } 3404 3405 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue) 3406 { 3407 struct virtnet_info *priv = netdev_priv(dev); 3408 struct send_queue *sq = &priv->sq[txqueue]; 3409 struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue); 3410 3411 u64_stats_update_begin(&sq->stats.syncp); 3412 sq->stats.tx_timeouts++; 3413 u64_stats_update_end(&sq->stats.syncp); 3414 3415 netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n", 3416 txqueue, sq->name, sq->vq->index, sq->vq->name, 3417 jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start))); 3418 } 3419 3420 static const struct net_device_ops virtnet_netdev = { 3421 .ndo_open = virtnet_open, 3422 .ndo_stop = virtnet_close, 3423 .ndo_start_xmit = start_xmit, 3424 .ndo_validate_addr = eth_validate_addr, 3425 .ndo_set_mac_address = virtnet_set_mac_address, 3426 .ndo_set_rx_mode = virtnet_set_rx_mode, 3427 .ndo_get_stats64 = virtnet_stats, 3428 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid, 3429 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid, 3430 .ndo_bpf = virtnet_xdp, 3431 .ndo_xdp_xmit = virtnet_xdp_xmit, 3432 .ndo_features_check = passthru_features_check, 3433 .ndo_get_phys_port_name = virtnet_get_phys_port_name, 3434 .ndo_set_features = virtnet_set_features, 3435 .ndo_tx_timeout = virtnet_tx_timeout, 3436 }; 3437 3438 static void virtnet_config_changed_work(struct work_struct *work) 3439 { 3440 struct virtnet_info *vi = 3441 container_of(work, struct virtnet_info, config_work); 3442 u16 v; 3443 3444 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS, 3445 struct virtio_net_config, status, &v) < 0) 3446 return; 3447 3448 if (v & VIRTIO_NET_S_ANNOUNCE) { 3449 netdev_notify_peers(vi->dev); 3450 virtnet_ack_link_announce(vi); 3451 } 3452 3453 /* Ignore unknown (future) status bits */ 3454 v &= VIRTIO_NET_S_LINK_UP; 3455 3456 if (vi->status == v) 3457 return; 3458 3459 vi->status = v; 3460 3461 if (vi->status & VIRTIO_NET_S_LINK_UP) { 3462 virtnet_update_settings(vi); 3463 netif_carrier_on(vi->dev); 3464 netif_tx_wake_all_queues(vi->dev); 3465 } else { 3466 netif_carrier_off(vi->dev); 3467 netif_tx_stop_all_queues(vi->dev); 3468 } 3469 } 3470 3471 static void virtnet_config_changed(struct virtio_device *vdev) 3472 { 3473 struct virtnet_info *vi = vdev->priv; 3474 3475 schedule_work(&vi->config_work); 3476 } 3477 3478 static void virtnet_free_queues(struct virtnet_info *vi) 3479 { 3480 int i; 3481 3482 for (i = 0; i < vi->max_queue_pairs; i++) { 3483 __netif_napi_del(&vi->rq[i].napi); 3484 __netif_napi_del(&vi->sq[i].napi); 3485 } 3486 3487 /* We called __netif_napi_del(), 3488 * we need to respect an RCU grace period before freeing vi->rq 3489 */ 3490 synchronize_net(); 3491 3492 kfree(vi->rq); 3493 kfree(vi->sq); 3494 kfree(vi->ctrl); 3495 } 3496 3497 static void _free_receive_bufs(struct virtnet_info *vi) 3498 { 3499 struct bpf_prog *old_prog; 3500 int i; 3501 3502 for (i = 0; i < vi->max_queue_pairs; i++) { 3503 while (vi->rq[i].pages) 3504 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0); 3505 3506 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 3507 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL); 3508 if (old_prog) 3509 bpf_prog_put(old_prog); 3510 } 3511 } 3512 3513 static void free_receive_bufs(struct virtnet_info *vi) 3514 { 3515 rtnl_lock(); 3516 _free_receive_bufs(vi); 3517 rtnl_unlock(); 3518 } 3519 3520 static void free_receive_page_frags(struct virtnet_info *vi) 3521 { 3522 int i; 3523 for (i = 0; i < vi->max_queue_pairs; i++) 3524 if (vi->rq[i].alloc_frag.page) 3525 put_page(vi->rq[i].alloc_frag.page); 3526 } 3527 3528 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf) 3529 { 3530 if (!is_xdp_frame(buf)) 3531 dev_kfree_skb(buf); 3532 else 3533 xdp_return_frame(ptr_to_xdp(buf)); 3534 } 3535 3536 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf) 3537 { 3538 struct virtnet_info *vi = vq->vdev->priv; 3539 int i = vq2rxq(vq); 3540 3541 if (vi->mergeable_rx_bufs) 3542 put_page(virt_to_head_page(buf)); 3543 else if (vi->big_packets) 3544 give_pages(&vi->rq[i], buf); 3545 else 3546 put_page(virt_to_head_page(buf)); 3547 } 3548 3549 static void free_unused_bufs(struct virtnet_info *vi) 3550 { 3551 void *buf; 3552 int i; 3553 3554 for (i = 0; i < vi->max_queue_pairs; i++) { 3555 struct virtqueue *vq = vi->sq[i].vq; 3556 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) 3557 virtnet_sq_free_unused_buf(vq, buf); 3558 cond_resched(); 3559 } 3560 3561 for (i = 0; i < vi->max_queue_pairs; i++) { 3562 struct virtqueue *vq = vi->rq[i].vq; 3563 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) 3564 virtnet_rq_free_unused_buf(vq, buf); 3565 cond_resched(); 3566 } 3567 } 3568 3569 static void virtnet_del_vqs(struct virtnet_info *vi) 3570 { 3571 struct virtio_device *vdev = vi->vdev; 3572 3573 virtnet_clean_affinity(vi); 3574 3575 vdev->config->del_vqs(vdev); 3576 3577 virtnet_free_queues(vi); 3578 } 3579 3580 /* How large should a single buffer be so a queue full of these can fit at 3581 * least one full packet? 3582 * Logic below assumes the mergeable buffer header is used. 3583 */ 3584 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq) 3585 { 3586 const unsigned int hdr_len = vi->hdr_len; 3587 unsigned int rq_size = virtqueue_get_vring_size(vq); 3588 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu; 3589 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len; 3590 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size); 3591 3592 return max(max(min_buf_len, hdr_len) - hdr_len, 3593 (unsigned int)GOOD_PACKET_LEN); 3594 } 3595 3596 static int virtnet_find_vqs(struct virtnet_info *vi) 3597 { 3598 vq_callback_t **callbacks; 3599 struct virtqueue **vqs; 3600 int ret = -ENOMEM; 3601 int i, total_vqs; 3602 const char **names; 3603 bool *ctx; 3604 3605 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by 3606 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by 3607 * possible control vq. 3608 */ 3609 total_vqs = vi->max_queue_pairs * 2 + 3610 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ); 3611 3612 /* Allocate space for find_vqs parameters */ 3613 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL); 3614 if (!vqs) 3615 goto err_vq; 3616 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL); 3617 if (!callbacks) 3618 goto err_callback; 3619 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL); 3620 if (!names) 3621 goto err_names; 3622 if (!vi->big_packets || vi->mergeable_rx_bufs) { 3623 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL); 3624 if (!ctx) 3625 goto err_ctx; 3626 } else { 3627 ctx = NULL; 3628 } 3629 3630 /* Parameters for control virtqueue, if any */ 3631 if (vi->has_cvq) { 3632 callbacks[total_vqs - 1] = NULL; 3633 names[total_vqs - 1] = "control"; 3634 } 3635 3636 /* Allocate/initialize parameters for send/receive virtqueues */ 3637 for (i = 0; i < vi->max_queue_pairs; i++) { 3638 callbacks[rxq2vq(i)] = skb_recv_done; 3639 callbacks[txq2vq(i)] = skb_xmit_done; 3640 sprintf(vi->rq[i].name, "input.%d", i); 3641 sprintf(vi->sq[i].name, "output.%d", i); 3642 names[rxq2vq(i)] = vi->rq[i].name; 3643 names[txq2vq(i)] = vi->sq[i].name; 3644 if (ctx) 3645 ctx[rxq2vq(i)] = true; 3646 } 3647 3648 ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks, 3649 names, ctx, NULL); 3650 if (ret) 3651 goto err_find; 3652 3653 if (vi->has_cvq) { 3654 vi->cvq = vqs[total_vqs - 1]; 3655 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN)) 3656 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER; 3657 } 3658 3659 for (i = 0; i < vi->max_queue_pairs; i++) { 3660 vi->rq[i].vq = vqs[rxq2vq(i)]; 3661 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq); 3662 vi->sq[i].vq = vqs[txq2vq(i)]; 3663 } 3664 3665 /* run here: ret == 0. */ 3666 3667 3668 err_find: 3669 kfree(ctx); 3670 err_ctx: 3671 kfree(names); 3672 err_names: 3673 kfree(callbacks); 3674 err_callback: 3675 kfree(vqs); 3676 err_vq: 3677 return ret; 3678 } 3679 3680 static int virtnet_alloc_queues(struct virtnet_info *vi) 3681 { 3682 int i; 3683 3684 if (vi->has_cvq) { 3685 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL); 3686 if (!vi->ctrl) 3687 goto err_ctrl; 3688 } else { 3689 vi->ctrl = NULL; 3690 } 3691 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL); 3692 if (!vi->sq) 3693 goto err_sq; 3694 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL); 3695 if (!vi->rq) 3696 goto err_rq; 3697 3698 INIT_DELAYED_WORK(&vi->refill, refill_work); 3699 for (i = 0; i < vi->max_queue_pairs; i++) { 3700 vi->rq[i].pages = NULL; 3701 netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll, 3702 napi_weight); 3703 netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi, 3704 virtnet_poll_tx, 3705 napi_tx ? napi_weight : 0); 3706 3707 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg)); 3708 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len); 3709 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg)); 3710 3711 u64_stats_init(&vi->rq[i].stats.syncp); 3712 u64_stats_init(&vi->sq[i].stats.syncp); 3713 } 3714 3715 return 0; 3716 3717 err_rq: 3718 kfree(vi->sq); 3719 err_sq: 3720 kfree(vi->ctrl); 3721 err_ctrl: 3722 return -ENOMEM; 3723 } 3724 3725 static int init_vqs(struct virtnet_info *vi) 3726 { 3727 int ret; 3728 3729 /* Allocate send & receive queues */ 3730 ret = virtnet_alloc_queues(vi); 3731 if (ret) 3732 goto err; 3733 3734 ret = virtnet_find_vqs(vi); 3735 if (ret) 3736 goto err_free; 3737 3738 cpus_read_lock(); 3739 virtnet_set_affinity(vi); 3740 cpus_read_unlock(); 3741 3742 return 0; 3743 3744 err_free: 3745 virtnet_free_queues(vi); 3746 err: 3747 return ret; 3748 } 3749 3750 #ifdef CONFIG_SYSFS 3751 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue, 3752 char *buf) 3753 { 3754 struct virtnet_info *vi = netdev_priv(queue->dev); 3755 unsigned int queue_index = get_netdev_rx_queue_index(queue); 3756 unsigned int headroom = virtnet_get_headroom(vi); 3757 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 3758 struct ewma_pkt_len *avg; 3759 3760 BUG_ON(queue_index >= vi->max_queue_pairs); 3761 avg = &vi->rq[queue_index].mrg_avg_pkt_len; 3762 return sprintf(buf, "%u\n", 3763 get_mergeable_buf_len(&vi->rq[queue_index], avg, 3764 SKB_DATA_ALIGN(headroom + tailroom))); 3765 } 3766 3767 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute = 3768 __ATTR_RO(mergeable_rx_buffer_size); 3769 3770 static struct attribute *virtio_net_mrg_rx_attrs[] = { 3771 &mergeable_rx_buffer_size_attribute.attr, 3772 NULL 3773 }; 3774 3775 static const struct attribute_group virtio_net_mrg_rx_group = { 3776 .name = "virtio_net", 3777 .attrs = virtio_net_mrg_rx_attrs 3778 }; 3779 #endif 3780 3781 static bool virtnet_fail_on_feature(struct virtio_device *vdev, 3782 unsigned int fbit, 3783 const char *fname, const char *dname) 3784 { 3785 if (!virtio_has_feature(vdev, fbit)) 3786 return false; 3787 3788 dev_err(&vdev->dev, "device advertises feature %s but not %s", 3789 fname, dname); 3790 3791 return true; 3792 } 3793 3794 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \ 3795 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit) 3796 3797 static bool virtnet_validate_features(struct virtio_device *vdev) 3798 { 3799 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) && 3800 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX, 3801 "VIRTIO_NET_F_CTRL_VQ") || 3802 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN, 3803 "VIRTIO_NET_F_CTRL_VQ") || 3804 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE, 3805 "VIRTIO_NET_F_CTRL_VQ") || 3806 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") || 3807 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR, 3808 "VIRTIO_NET_F_CTRL_VQ") || 3809 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS, 3810 "VIRTIO_NET_F_CTRL_VQ") || 3811 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT, 3812 "VIRTIO_NET_F_CTRL_VQ") || 3813 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL, 3814 "VIRTIO_NET_F_CTRL_VQ"))) { 3815 return false; 3816 } 3817 3818 return true; 3819 } 3820 3821 #define MIN_MTU ETH_MIN_MTU 3822 #define MAX_MTU ETH_MAX_MTU 3823 3824 static int virtnet_validate(struct virtio_device *vdev) 3825 { 3826 if (!vdev->config->get) { 3827 dev_err(&vdev->dev, "%s failure: config access disabled\n", 3828 __func__); 3829 return -EINVAL; 3830 } 3831 3832 if (!virtnet_validate_features(vdev)) 3833 return -EINVAL; 3834 3835 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 3836 int mtu = virtio_cread16(vdev, 3837 offsetof(struct virtio_net_config, 3838 mtu)); 3839 if (mtu < MIN_MTU) 3840 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU); 3841 } 3842 3843 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY) && 3844 !virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) { 3845 dev_warn(&vdev->dev, "device advertises feature VIRTIO_NET_F_STANDBY but not VIRTIO_NET_F_MAC, disabling standby"); 3846 __virtio_clear_bit(vdev, VIRTIO_NET_F_STANDBY); 3847 } 3848 3849 return 0; 3850 } 3851 3852 static bool virtnet_check_guest_gso(const struct virtnet_info *vi) 3853 { 3854 return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) || 3855 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) || 3856 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) || 3857 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) || 3858 (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) && 3859 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6)); 3860 } 3861 3862 static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu) 3863 { 3864 bool guest_gso = virtnet_check_guest_gso(vi); 3865 3866 /* If device can receive ANY guest GSO packets, regardless of mtu, 3867 * allocate packets of maximum size, otherwise limit it to only 3868 * mtu size worth only. 3869 */ 3870 if (mtu > ETH_DATA_LEN || guest_gso) { 3871 vi->big_packets = true; 3872 vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE); 3873 } 3874 } 3875 3876 static int virtnet_probe(struct virtio_device *vdev) 3877 { 3878 int i, err = -ENOMEM; 3879 struct net_device *dev; 3880 struct virtnet_info *vi; 3881 u16 max_queue_pairs; 3882 int mtu = 0; 3883 3884 /* Find if host supports multiqueue/rss virtio_net device */ 3885 max_queue_pairs = 1; 3886 if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS)) 3887 max_queue_pairs = 3888 virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs)); 3889 3890 /* We need at least 2 queue's */ 3891 if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN || 3892 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX || 3893 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 3894 max_queue_pairs = 1; 3895 3896 /* Allocate ourselves a network device with room for our info */ 3897 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs); 3898 if (!dev) 3899 return -ENOMEM; 3900 3901 /* Set up network device as normal. */ 3902 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE | 3903 IFF_TX_SKB_NO_LINEAR; 3904 dev->netdev_ops = &virtnet_netdev; 3905 dev->features = NETIF_F_HIGHDMA; 3906 3907 dev->ethtool_ops = &virtnet_ethtool_ops; 3908 SET_NETDEV_DEV(dev, &vdev->dev); 3909 3910 /* Do we support "hardware" checksums? */ 3911 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) { 3912 /* This opens up the world of extra features. */ 3913 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG; 3914 if (csum) 3915 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG; 3916 3917 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) { 3918 dev->hw_features |= NETIF_F_TSO 3919 | NETIF_F_TSO_ECN | NETIF_F_TSO6; 3920 } 3921 /* Individual feature bits: what can host handle? */ 3922 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4)) 3923 dev->hw_features |= NETIF_F_TSO; 3924 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6)) 3925 dev->hw_features |= NETIF_F_TSO6; 3926 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN)) 3927 dev->hw_features |= NETIF_F_TSO_ECN; 3928 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_USO)) 3929 dev->hw_features |= NETIF_F_GSO_UDP_L4; 3930 3931 dev->features |= NETIF_F_GSO_ROBUST; 3932 3933 if (gso) 3934 dev->features |= dev->hw_features & NETIF_F_ALL_TSO; 3935 /* (!csum && gso) case will be fixed by register_netdev() */ 3936 } 3937 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM)) 3938 dev->features |= NETIF_F_RXCSUM; 3939 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) || 3940 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6)) 3941 dev->features |= NETIF_F_GRO_HW; 3942 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) 3943 dev->hw_features |= NETIF_F_GRO_HW; 3944 3945 dev->vlan_features = dev->features; 3946 dev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT; 3947 3948 /* MTU range: 68 - 65535 */ 3949 dev->min_mtu = MIN_MTU; 3950 dev->max_mtu = MAX_MTU; 3951 3952 /* Configuration may specify what MAC to use. Otherwise random. */ 3953 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) { 3954 u8 addr[ETH_ALEN]; 3955 3956 virtio_cread_bytes(vdev, 3957 offsetof(struct virtio_net_config, mac), 3958 addr, ETH_ALEN); 3959 eth_hw_addr_set(dev, addr); 3960 } else { 3961 eth_hw_addr_random(dev); 3962 dev_info(&vdev->dev, "Assigned random MAC address %pM\n", 3963 dev->dev_addr); 3964 } 3965 3966 /* Set up our device-specific information */ 3967 vi = netdev_priv(dev); 3968 vi->dev = dev; 3969 vi->vdev = vdev; 3970 vdev->priv = vi; 3971 3972 INIT_WORK(&vi->config_work, virtnet_config_changed_work); 3973 spin_lock_init(&vi->refill_lock); 3974 3975 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) { 3976 vi->mergeable_rx_bufs = true; 3977 dev->xdp_features |= NETDEV_XDP_ACT_RX_SG; 3978 } 3979 3980 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) { 3981 vi->rx_usecs = 0; 3982 vi->tx_usecs = 0; 3983 vi->tx_max_packets = 0; 3984 vi->rx_max_packets = 0; 3985 } 3986 3987 if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT)) 3988 vi->has_rss_hash_report = true; 3989 3990 if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS)) 3991 vi->has_rss = true; 3992 3993 if (vi->has_rss || vi->has_rss_hash_report) { 3994 vi->rss_indir_table_size = 3995 virtio_cread16(vdev, offsetof(struct virtio_net_config, 3996 rss_max_indirection_table_length)); 3997 vi->rss_key_size = 3998 virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size)); 3999 4000 vi->rss_hash_types_supported = 4001 virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types)); 4002 vi->rss_hash_types_supported &= 4003 ~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX | 4004 VIRTIO_NET_RSS_HASH_TYPE_TCP_EX | 4005 VIRTIO_NET_RSS_HASH_TYPE_UDP_EX); 4006 4007 dev->hw_features |= NETIF_F_RXHASH; 4008 } 4009 4010 if (vi->has_rss_hash_report) 4011 vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash); 4012 else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) || 4013 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 4014 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 4015 else 4016 vi->hdr_len = sizeof(struct virtio_net_hdr); 4017 4018 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) || 4019 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 4020 vi->any_header_sg = true; 4021 4022 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 4023 vi->has_cvq = true; 4024 4025 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 4026 mtu = virtio_cread16(vdev, 4027 offsetof(struct virtio_net_config, 4028 mtu)); 4029 if (mtu < dev->min_mtu) { 4030 /* Should never trigger: MTU was previously validated 4031 * in virtnet_validate. 4032 */ 4033 dev_err(&vdev->dev, 4034 "device MTU appears to have changed it is now %d < %d", 4035 mtu, dev->min_mtu); 4036 err = -EINVAL; 4037 goto free; 4038 } 4039 4040 dev->mtu = mtu; 4041 dev->max_mtu = mtu; 4042 } 4043 4044 virtnet_set_big_packets(vi, mtu); 4045 4046 if (vi->any_header_sg) 4047 dev->needed_headroom = vi->hdr_len; 4048 4049 /* Enable multiqueue by default */ 4050 if (num_online_cpus() >= max_queue_pairs) 4051 vi->curr_queue_pairs = max_queue_pairs; 4052 else 4053 vi->curr_queue_pairs = num_online_cpus(); 4054 vi->max_queue_pairs = max_queue_pairs; 4055 4056 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */ 4057 err = init_vqs(vi); 4058 if (err) 4059 goto free; 4060 4061 #ifdef CONFIG_SYSFS 4062 if (vi->mergeable_rx_bufs) 4063 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group; 4064 #endif 4065 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs); 4066 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs); 4067 4068 virtnet_init_settings(dev); 4069 4070 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) { 4071 vi->failover = net_failover_create(vi->dev); 4072 if (IS_ERR(vi->failover)) { 4073 err = PTR_ERR(vi->failover); 4074 goto free_vqs; 4075 } 4076 } 4077 4078 if (vi->has_rss || vi->has_rss_hash_report) 4079 virtnet_init_default_rss(vi); 4080 4081 /* serialize netdev register + virtio_device_ready() with ndo_open() */ 4082 rtnl_lock(); 4083 4084 err = register_netdevice(dev); 4085 if (err) { 4086 pr_debug("virtio_net: registering device failed\n"); 4087 rtnl_unlock(); 4088 goto free_failover; 4089 } 4090 4091 virtio_device_ready(vdev); 4092 4093 /* a random MAC address has been assigned, notify the device. 4094 * We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not there 4095 * because many devices work fine without getting MAC explicitly 4096 */ 4097 if (!virtio_has_feature(vdev, VIRTIO_NET_F_MAC) && 4098 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 4099 struct scatterlist sg; 4100 4101 sg_init_one(&sg, dev->dev_addr, dev->addr_len); 4102 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 4103 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) { 4104 pr_debug("virtio_net: setting MAC address failed\n"); 4105 rtnl_unlock(); 4106 err = -EINVAL; 4107 goto free_unregister_netdev; 4108 } 4109 } 4110 4111 rtnl_unlock(); 4112 4113 err = virtnet_cpu_notif_add(vi); 4114 if (err) { 4115 pr_debug("virtio_net: registering cpu notifier failed\n"); 4116 goto free_unregister_netdev; 4117 } 4118 4119 virtnet_set_queues(vi, vi->curr_queue_pairs); 4120 4121 /* Assume link up if device can't report link status, 4122 otherwise get link status from config. */ 4123 netif_carrier_off(dev); 4124 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) { 4125 schedule_work(&vi->config_work); 4126 } else { 4127 vi->status = VIRTIO_NET_S_LINK_UP; 4128 virtnet_update_settings(vi); 4129 netif_carrier_on(dev); 4130 } 4131 4132 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++) 4133 if (virtio_has_feature(vi->vdev, guest_offloads[i])) 4134 set_bit(guest_offloads[i], &vi->guest_offloads); 4135 vi->guest_offloads_capable = vi->guest_offloads; 4136 4137 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n", 4138 dev->name, max_queue_pairs); 4139 4140 return 0; 4141 4142 free_unregister_netdev: 4143 unregister_netdev(dev); 4144 free_failover: 4145 net_failover_destroy(vi->failover); 4146 free_vqs: 4147 virtio_reset_device(vdev); 4148 cancel_delayed_work_sync(&vi->refill); 4149 free_receive_page_frags(vi); 4150 virtnet_del_vqs(vi); 4151 free: 4152 free_netdev(dev); 4153 return err; 4154 } 4155 4156 static void remove_vq_common(struct virtnet_info *vi) 4157 { 4158 virtio_reset_device(vi->vdev); 4159 4160 /* Free unused buffers in both send and recv, if any. */ 4161 free_unused_bufs(vi); 4162 4163 free_receive_bufs(vi); 4164 4165 free_receive_page_frags(vi); 4166 4167 virtnet_del_vqs(vi); 4168 } 4169 4170 static void virtnet_remove(struct virtio_device *vdev) 4171 { 4172 struct virtnet_info *vi = vdev->priv; 4173 4174 virtnet_cpu_notif_remove(vi); 4175 4176 /* Make sure no work handler is accessing the device. */ 4177 flush_work(&vi->config_work); 4178 4179 unregister_netdev(vi->dev); 4180 4181 net_failover_destroy(vi->failover); 4182 4183 remove_vq_common(vi); 4184 4185 free_netdev(vi->dev); 4186 } 4187 4188 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev) 4189 { 4190 struct virtnet_info *vi = vdev->priv; 4191 4192 virtnet_cpu_notif_remove(vi); 4193 virtnet_freeze_down(vdev); 4194 remove_vq_common(vi); 4195 4196 return 0; 4197 } 4198 4199 static __maybe_unused int virtnet_restore(struct virtio_device *vdev) 4200 { 4201 struct virtnet_info *vi = vdev->priv; 4202 int err; 4203 4204 err = virtnet_restore_up(vdev); 4205 if (err) 4206 return err; 4207 virtnet_set_queues(vi, vi->curr_queue_pairs); 4208 4209 err = virtnet_cpu_notif_add(vi); 4210 if (err) { 4211 virtnet_freeze_down(vdev); 4212 remove_vq_common(vi); 4213 return err; 4214 } 4215 4216 return 0; 4217 } 4218 4219 static struct virtio_device_id id_table[] = { 4220 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID }, 4221 { 0 }, 4222 }; 4223 4224 #define VIRTNET_FEATURES \ 4225 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \ 4226 VIRTIO_NET_F_MAC, \ 4227 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \ 4228 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \ 4229 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \ 4230 VIRTIO_NET_F_HOST_USO, VIRTIO_NET_F_GUEST_USO4, VIRTIO_NET_F_GUEST_USO6, \ 4231 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \ 4232 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \ 4233 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \ 4234 VIRTIO_NET_F_CTRL_MAC_ADDR, \ 4235 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \ 4236 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \ 4237 VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL, \ 4238 VIRTIO_NET_F_GUEST_HDRLEN 4239 4240 static unsigned int features[] = { 4241 VIRTNET_FEATURES, 4242 }; 4243 4244 static unsigned int features_legacy[] = { 4245 VIRTNET_FEATURES, 4246 VIRTIO_NET_F_GSO, 4247 VIRTIO_F_ANY_LAYOUT, 4248 }; 4249 4250 static struct virtio_driver virtio_net_driver = { 4251 .feature_table = features, 4252 .feature_table_size = ARRAY_SIZE(features), 4253 .feature_table_legacy = features_legacy, 4254 .feature_table_size_legacy = ARRAY_SIZE(features_legacy), 4255 .driver.name = KBUILD_MODNAME, 4256 .driver.owner = THIS_MODULE, 4257 .id_table = id_table, 4258 .validate = virtnet_validate, 4259 .probe = virtnet_probe, 4260 .remove = virtnet_remove, 4261 .config_changed = virtnet_config_changed, 4262 #ifdef CONFIG_PM_SLEEP 4263 .freeze = virtnet_freeze, 4264 .restore = virtnet_restore, 4265 #endif 4266 }; 4267 4268 static __init int virtio_net_driver_init(void) 4269 { 4270 int ret; 4271 4272 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online", 4273 virtnet_cpu_online, 4274 virtnet_cpu_down_prep); 4275 if (ret < 0) 4276 goto out; 4277 virtionet_online = ret; 4278 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead", 4279 NULL, virtnet_cpu_dead); 4280 if (ret) 4281 goto err_dead; 4282 ret = register_virtio_driver(&virtio_net_driver); 4283 if (ret) 4284 goto err_virtio; 4285 return 0; 4286 err_virtio: 4287 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 4288 err_dead: 4289 cpuhp_remove_multi_state(virtionet_online); 4290 out: 4291 return ret; 4292 } 4293 module_init(virtio_net_driver_init); 4294 4295 static __exit void virtio_net_driver_exit(void) 4296 { 4297 unregister_virtio_driver(&virtio_net_driver); 4298 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 4299 cpuhp_remove_multi_state(virtionet_online); 4300 } 4301 module_exit(virtio_net_driver_exit); 4302 4303 MODULE_DEVICE_TABLE(virtio, id_table); 4304 MODULE_DESCRIPTION("Virtio network driver"); 4305 MODULE_LICENSE("GPL"); 4306