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