1 /* A network driver using virtio. 2 * 3 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation 4 * 5 * This program is free software; you can redistribute it and/or modify 6 * it under the terms of the GNU General Public License as published by 7 * the Free Software Foundation; either version 2 of the License, or 8 * (at your option) any later version. 9 * 10 * This program is distributed in the hope that it will be useful, 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 13 * GNU General Public License for more details. 14 * 15 * You should have received a copy of the GNU General Public License 16 * along with this program; if not, see <http://www.gnu.org/licenses/>. 17 */ 18 //#define DEBUG 19 #include <linux/netdevice.h> 20 #include <linux/etherdevice.h> 21 #include <linux/ethtool.h> 22 #include <linux/module.h> 23 #include <linux/virtio.h> 24 #include <linux/virtio_net.h> 25 #include <linux/bpf.h> 26 #include <linux/bpf_trace.h> 27 #include <linux/scatterlist.h> 28 #include <linux/if_vlan.h> 29 #include <linux/slab.h> 30 #include <linux/cpu.h> 31 #include <linux/average.h> 32 #include <linux/filter.h> 33 #include <net/route.h> 34 #include <net/xdp.h> 35 36 static int napi_weight = NAPI_POLL_WEIGHT; 37 module_param(napi_weight, int, 0444); 38 39 static bool csum = true, gso = true, napi_tx; 40 module_param(csum, bool, 0444); 41 module_param(gso, bool, 0444); 42 module_param(napi_tx, bool, 0644); 43 44 /* FIXME: MTU in config. */ 45 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN) 46 #define GOOD_COPY_LEN 128 47 48 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD) 49 50 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */ 51 #define VIRTIO_XDP_HEADROOM 256 52 53 /* RX packet size EWMA. The average packet size is used to determine the packet 54 * buffer size when refilling RX rings. As the entire RX ring may be refilled 55 * at once, the weight is chosen so that the EWMA will be insensitive to short- 56 * term, transient changes in packet size. 57 */ 58 DECLARE_EWMA(pkt_len, 0, 64) 59 60 #define VIRTNET_DRIVER_VERSION "1.0.0" 61 62 static const unsigned long guest_offloads[] = { 63 VIRTIO_NET_F_GUEST_TSO4, 64 VIRTIO_NET_F_GUEST_TSO6, 65 VIRTIO_NET_F_GUEST_ECN, 66 VIRTIO_NET_F_GUEST_UFO 67 }; 68 69 struct virtnet_stat_desc { 70 char desc[ETH_GSTRING_LEN]; 71 size_t offset; 72 }; 73 74 struct virtnet_sq_stats { 75 struct u64_stats_sync syncp; 76 u64 packets; 77 u64 bytes; 78 }; 79 80 struct virtnet_rq_stats { 81 struct u64_stats_sync syncp; 82 u64 packets; 83 u64 bytes; 84 }; 85 86 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m) 87 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m) 88 89 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = { 90 { "packets", VIRTNET_SQ_STAT(packets) }, 91 { "bytes", VIRTNET_SQ_STAT(bytes) }, 92 }; 93 94 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = { 95 { "packets", VIRTNET_RQ_STAT(packets) }, 96 { "bytes", VIRTNET_RQ_STAT(bytes) }, 97 }; 98 99 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc) 100 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc) 101 102 /* Internal representation of a send virtqueue */ 103 struct send_queue { 104 /* Virtqueue associated with this send _queue */ 105 struct virtqueue *vq; 106 107 /* TX: fragments + linear part + virtio header */ 108 struct scatterlist sg[MAX_SKB_FRAGS + 2]; 109 110 /* Name of the send queue: output.$index */ 111 char name[40]; 112 113 struct virtnet_sq_stats stats; 114 115 struct napi_struct napi; 116 }; 117 118 /* Internal representation of a receive virtqueue */ 119 struct receive_queue { 120 /* Virtqueue associated with this receive_queue */ 121 struct virtqueue *vq; 122 123 struct napi_struct napi; 124 125 struct bpf_prog __rcu *xdp_prog; 126 127 struct virtnet_rq_stats stats; 128 129 /* Chain pages by the private ptr. */ 130 struct page *pages; 131 132 /* Average packet length for mergeable receive buffers. */ 133 struct ewma_pkt_len mrg_avg_pkt_len; 134 135 /* Page frag for packet buffer allocation. */ 136 struct page_frag alloc_frag; 137 138 /* RX: fragments + linear part + virtio header */ 139 struct scatterlist sg[MAX_SKB_FRAGS + 2]; 140 141 /* Min single buffer size for mergeable buffers case. */ 142 unsigned int min_buf_len; 143 144 /* Name of this receive queue: input.$index */ 145 char name[40]; 146 147 struct xdp_rxq_info xdp_rxq; 148 }; 149 150 struct virtnet_info { 151 struct virtio_device *vdev; 152 struct virtqueue *cvq; 153 struct net_device *dev; 154 struct send_queue *sq; 155 struct receive_queue *rq; 156 unsigned int status; 157 158 /* Max # of queue pairs supported by the device */ 159 u16 max_queue_pairs; 160 161 /* # of queue pairs currently used by the driver */ 162 u16 curr_queue_pairs; 163 164 /* # of XDP queue pairs currently used by the driver */ 165 u16 xdp_queue_pairs; 166 167 /* I like... big packets and I cannot lie! */ 168 bool big_packets; 169 170 /* Host will merge rx buffers for big packets (shake it! shake it!) */ 171 bool mergeable_rx_bufs; 172 173 /* Has control virtqueue */ 174 bool has_cvq; 175 176 /* Host can handle any s/g split between our header and packet data */ 177 bool any_header_sg; 178 179 /* Packet virtio header size */ 180 u8 hdr_len; 181 182 /* Work struct for refilling if we run low on memory. */ 183 struct delayed_work refill; 184 185 /* Work struct for config space updates */ 186 struct work_struct config_work; 187 188 /* Does the affinity hint is set for virtqueues? */ 189 bool affinity_hint_set; 190 191 /* CPU hotplug instances for online & dead */ 192 struct hlist_node node; 193 struct hlist_node node_dead; 194 195 /* Control VQ buffers: protected by the rtnl lock */ 196 struct virtio_net_ctrl_hdr ctrl_hdr; 197 virtio_net_ctrl_ack ctrl_status; 198 struct virtio_net_ctrl_mq ctrl_mq; 199 u8 ctrl_promisc; 200 u8 ctrl_allmulti; 201 u16 ctrl_vid; 202 u64 ctrl_offloads; 203 204 /* Ethtool settings */ 205 u8 duplex; 206 u32 speed; 207 208 unsigned long guest_offloads; 209 }; 210 211 struct padded_vnet_hdr { 212 struct virtio_net_hdr_mrg_rxbuf hdr; 213 /* 214 * hdr is in a separate sg buffer, and data sg buffer shares same page 215 * with this header sg. This padding makes next sg 16 byte aligned 216 * after the header. 217 */ 218 char padding[4]; 219 }; 220 221 /* Converting between virtqueue no. and kernel tx/rx queue no. 222 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq 223 */ 224 static int vq2txq(struct virtqueue *vq) 225 { 226 return (vq->index - 1) / 2; 227 } 228 229 static int txq2vq(int txq) 230 { 231 return txq * 2 + 1; 232 } 233 234 static int vq2rxq(struct virtqueue *vq) 235 { 236 return vq->index / 2; 237 } 238 239 static int rxq2vq(int rxq) 240 { 241 return rxq * 2; 242 } 243 244 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb) 245 { 246 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb; 247 } 248 249 /* 250 * private is used to chain pages for big packets, put the whole 251 * most recent used list in the beginning for reuse 252 */ 253 static void give_pages(struct receive_queue *rq, struct page *page) 254 { 255 struct page *end; 256 257 /* Find end of list, sew whole thing into vi->rq.pages. */ 258 for (end = page; end->private; end = (struct page *)end->private); 259 end->private = (unsigned long)rq->pages; 260 rq->pages = page; 261 } 262 263 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask) 264 { 265 struct page *p = rq->pages; 266 267 if (p) { 268 rq->pages = (struct page *)p->private; 269 /* clear private here, it is used to chain pages */ 270 p->private = 0; 271 } else 272 p = alloc_page(gfp_mask); 273 return p; 274 } 275 276 static void virtqueue_napi_schedule(struct napi_struct *napi, 277 struct virtqueue *vq) 278 { 279 if (napi_schedule_prep(napi)) { 280 virtqueue_disable_cb(vq); 281 __napi_schedule(napi); 282 } 283 } 284 285 static void virtqueue_napi_complete(struct napi_struct *napi, 286 struct virtqueue *vq, int processed) 287 { 288 int opaque; 289 290 opaque = virtqueue_enable_cb_prepare(vq); 291 if (napi_complete_done(napi, processed)) { 292 if (unlikely(virtqueue_poll(vq, opaque))) 293 virtqueue_napi_schedule(napi, vq); 294 } else { 295 virtqueue_disable_cb(vq); 296 } 297 } 298 299 static void skb_xmit_done(struct virtqueue *vq) 300 { 301 struct virtnet_info *vi = vq->vdev->priv; 302 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi; 303 304 /* Suppress further interrupts. */ 305 virtqueue_disable_cb(vq); 306 307 if (napi->weight) 308 virtqueue_napi_schedule(napi, vq); 309 else 310 /* We were probably waiting for more output buffers. */ 311 netif_wake_subqueue(vi->dev, vq2txq(vq)); 312 } 313 314 #define MRG_CTX_HEADER_SHIFT 22 315 static void *mergeable_len_to_ctx(unsigned int truesize, 316 unsigned int headroom) 317 { 318 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize); 319 } 320 321 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx) 322 { 323 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT; 324 } 325 326 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx) 327 { 328 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1); 329 } 330 331 /* Called from bottom half context */ 332 static struct sk_buff *page_to_skb(struct virtnet_info *vi, 333 struct receive_queue *rq, 334 struct page *page, unsigned int offset, 335 unsigned int len, unsigned int truesize) 336 { 337 struct sk_buff *skb; 338 struct virtio_net_hdr_mrg_rxbuf *hdr; 339 unsigned int copy, hdr_len, hdr_padded_len; 340 char *p; 341 342 p = page_address(page) + offset; 343 344 /* copy small packet so we can reuse these pages for small data */ 345 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN); 346 if (unlikely(!skb)) 347 return NULL; 348 349 hdr = skb_vnet_hdr(skb); 350 351 hdr_len = vi->hdr_len; 352 if (vi->mergeable_rx_bufs) 353 hdr_padded_len = sizeof(*hdr); 354 else 355 hdr_padded_len = sizeof(struct padded_vnet_hdr); 356 357 memcpy(hdr, p, hdr_len); 358 359 len -= hdr_len; 360 offset += hdr_padded_len; 361 p += hdr_padded_len; 362 363 copy = len; 364 if (copy > skb_tailroom(skb)) 365 copy = skb_tailroom(skb); 366 skb_put_data(skb, p, copy); 367 368 len -= copy; 369 offset += copy; 370 371 if (vi->mergeable_rx_bufs) { 372 if (len) 373 skb_add_rx_frag(skb, 0, page, offset, len, truesize); 374 else 375 put_page(page); 376 return skb; 377 } 378 379 /* 380 * Verify that we can indeed put this data into a skb. 381 * This is here to handle cases when the device erroneously 382 * tries to receive more than is possible. This is usually 383 * the case of a broken device. 384 */ 385 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) { 386 net_dbg_ratelimited("%s: too much data\n", skb->dev->name); 387 dev_kfree_skb(skb); 388 return NULL; 389 } 390 BUG_ON(offset >= PAGE_SIZE); 391 while (len) { 392 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len); 393 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset, 394 frag_size, truesize); 395 len -= frag_size; 396 page = (struct page *)page->private; 397 offset = 0; 398 } 399 400 if (page) 401 give_pages(rq, page); 402 403 return skb; 404 } 405 406 static void virtnet_xdp_flush(struct net_device *dev) 407 { 408 struct virtnet_info *vi = netdev_priv(dev); 409 struct send_queue *sq; 410 unsigned int qp; 411 412 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id(); 413 sq = &vi->sq[qp]; 414 415 virtqueue_kick(sq->vq); 416 } 417 418 static bool __virtnet_xdp_xmit(struct virtnet_info *vi, 419 struct xdp_buff *xdp) 420 { 421 struct virtio_net_hdr_mrg_rxbuf *hdr; 422 unsigned int len; 423 struct send_queue *sq; 424 unsigned int qp; 425 void *xdp_sent; 426 int err; 427 428 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id(); 429 sq = &vi->sq[qp]; 430 431 /* Free up any pending old buffers before queueing new ones. */ 432 while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) { 433 struct page *sent_page = virt_to_head_page(xdp_sent); 434 435 put_page(sent_page); 436 } 437 438 xdp->data -= vi->hdr_len; 439 /* Zero header and leave csum up to XDP layers */ 440 hdr = xdp->data; 441 memset(hdr, 0, vi->hdr_len); 442 443 sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data); 444 445 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp->data, GFP_ATOMIC); 446 if (unlikely(err)) { 447 struct page *page = virt_to_head_page(xdp->data); 448 449 put_page(page); 450 return false; 451 } 452 453 return true; 454 } 455 456 static int virtnet_xdp_xmit(struct net_device *dev, struct xdp_buff *xdp) 457 { 458 struct virtnet_info *vi = netdev_priv(dev); 459 bool sent = __virtnet_xdp_xmit(vi, xdp); 460 461 if (!sent) 462 return -ENOSPC; 463 return 0; 464 } 465 466 static unsigned int virtnet_get_headroom(struct virtnet_info *vi) 467 { 468 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0; 469 } 470 471 /* We copy the packet for XDP in the following cases: 472 * 473 * 1) Packet is scattered across multiple rx buffers. 474 * 2) Headroom space is insufficient. 475 * 476 * This is inefficient but it's a temporary condition that 477 * we hit right after XDP is enabled and until queue is refilled 478 * with large buffers with sufficient headroom - so it should affect 479 * at most queue size packets. 480 * Afterwards, the conditions to enable 481 * XDP should preclude the underlying device from sending packets 482 * across multiple buffers (num_buf > 1), and we make sure buffers 483 * have enough headroom. 484 */ 485 static struct page *xdp_linearize_page(struct receive_queue *rq, 486 u16 *num_buf, 487 struct page *p, 488 int offset, 489 int page_off, 490 unsigned int *len) 491 { 492 struct page *page = alloc_page(GFP_ATOMIC); 493 494 if (!page) 495 return NULL; 496 497 memcpy(page_address(page) + page_off, page_address(p) + offset, *len); 498 page_off += *len; 499 500 while (--*num_buf) { 501 unsigned int buflen; 502 void *buf; 503 int off; 504 505 buf = virtqueue_get_buf(rq->vq, &buflen); 506 if (unlikely(!buf)) 507 goto err_buf; 508 509 p = virt_to_head_page(buf); 510 off = buf - page_address(p); 511 512 /* guard against a misconfigured or uncooperative backend that 513 * is sending packet larger than the MTU. 514 */ 515 if ((page_off + buflen) > PAGE_SIZE) { 516 put_page(p); 517 goto err_buf; 518 } 519 520 memcpy(page_address(page) + page_off, 521 page_address(p) + off, buflen); 522 page_off += buflen; 523 put_page(p); 524 } 525 526 /* Headroom does not contribute to packet length */ 527 *len = page_off - VIRTIO_XDP_HEADROOM; 528 return page; 529 err_buf: 530 __free_pages(page, 0); 531 return NULL; 532 } 533 534 static struct sk_buff *receive_small(struct net_device *dev, 535 struct virtnet_info *vi, 536 struct receive_queue *rq, 537 void *buf, void *ctx, 538 unsigned int len, 539 bool *xdp_xmit) 540 { 541 struct sk_buff *skb; 542 struct bpf_prog *xdp_prog; 543 unsigned int xdp_headroom = (unsigned long)ctx; 544 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom; 545 unsigned int headroom = vi->hdr_len + header_offset; 546 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 547 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 548 struct page *page = virt_to_head_page(buf); 549 unsigned int delta = 0, err; 550 struct page *xdp_page; 551 len -= vi->hdr_len; 552 553 rcu_read_lock(); 554 xdp_prog = rcu_dereference(rq->xdp_prog); 555 if (xdp_prog) { 556 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset; 557 struct xdp_buff xdp; 558 void *orig_data; 559 u32 act; 560 561 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags)) 562 goto err_xdp; 563 564 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) { 565 int offset = buf - page_address(page) + header_offset; 566 unsigned int tlen = len + vi->hdr_len; 567 u16 num_buf = 1; 568 569 xdp_headroom = virtnet_get_headroom(vi); 570 header_offset = VIRTNET_RX_PAD + xdp_headroom; 571 headroom = vi->hdr_len + header_offset; 572 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 573 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 574 xdp_page = xdp_linearize_page(rq, &num_buf, page, 575 offset, header_offset, 576 &tlen); 577 if (!xdp_page) 578 goto err_xdp; 579 580 buf = page_address(xdp_page); 581 put_page(page); 582 page = xdp_page; 583 } 584 585 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len; 586 xdp.data = xdp.data_hard_start + xdp_headroom; 587 xdp_set_data_meta_invalid(&xdp); 588 xdp.data_end = xdp.data + len; 589 xdp.rxq = &rq->xdp_rxq; 590 orig_data = xdp.data; 591 act = bpf_prog_run_xdp(xdp_prog, &xdp); 592 593 switch (act) { 594 case XDP_PASS: 595 /* Recalculate length in case bpf program changed it */ 596 delta = orig_data - xdp.data; 597 break; 598 case XDP_TX: 599 if (unlikely(!__virtnet_xdp_xmit(vi, &xdp))) 600 trace_xdp_exception(vi->dev, xdp_prog, act); 601 else 602 *xdp_xmit = true; 603 rcu_read_unlock(); 604 goto xdp_xmit; 605 case XDP_REDIRECT: 606 err = xdp_do_redirect(dev, &xdp, xdp_prog); 607 if (!err) 608 *xdp_xmit = true; 609 rcu_read_unlock(); 610 goto xdp_xmit; 611 default: 612 bpf_warn_invalid_xdp_action(act); 613 case XDP_ABORTED: 614 trace_xdp_exception(vi->dev, xdp_prog, act); 615 case XDP_DROP: 616 goto err_xdp; 617 } 618 } 619 rcu_read_unlock(); 620 621 skb = build_skb(buf, buflen); 622 if (!skb) { 623 put_page(page); 624 goto err; 625 } 626 skb_reserve(skb, headroom - delta); 627 skb_put(skb, len + delta); 628 if (!delta) { 629 buf += header_offset; 630 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len); 631 } /* keep zeroed vnet hdr since packet was changed by bpf */ 632 633 err: 634 return skb; 635 636 err_xdp: 637 rcu_read_unlock(); 638 dev->stats.rx_dropped++; 639 put_page(page); 640 xdp_xmit: 641 return NULL; 642 } 643 644 static struct sk_buff *receive_big(struct net_device *dev, 645 struct virtnet_info *vi, 646 struct receive_queue *rq, 647 void *buf, 648 unsigned int len) 649 { 650 struct page *page = buf; 651 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE); 652 653 if (unlikely(!skb)) 654 goto err; 655 656 return skb; 657 658 err: 659 dev->stats.rx_dropped++; 660 give_pages(rq, page); 661 return NULL; 662 } 663 664 static struct sk_buff *receive_mergeable(struct net_device *dev, 665 struct virtnet_info *vi, 666 struct receive_queue *rq, 667 void *buf, 668 void *ctx, 669 unsigned int len, 670 bool *xdp_xmit) 671 { 672 struct virtio_net_hdr_mrg_rxbuf *hdr = buf; 673 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers); 674 struct page *page = virt_to_head_page(buf); 675 int offset = buf - page_address(page); 676 struct sk_buff *head_skb, *curr_skb; 677 struct bpf_prog *xdp_prog; 678 unsigned int truesize; 679 unsigned int headroom = mergeable_ctx_to_headroom(ctx); 680 681 head_skb = NULL; 682 683 rcu_read_lock(); 684 xdp_prog = rcu_dereference(rq->xdp_prog); 685 if (xdp_prog) { 686 struct page *xdp_page; 687 struct xdp_buff xdp; 688 void *data; 689 u32 act; 690 691 /* This happens when rx buffer size is underestimated */ 692 if (unlikely(num_buf > 1 || 693 headroom < virtnet_get_headroom(vi))) { 694 /* linearize data for XDP */ 695 xdp_page = xdp_linearize_page(rq, &num_buf, 696 page, offset, 697 VIRTIO_XDP_HEADROOM, 698 &len); 699 if (!xdp_page) 700 goto err_xdp; 701 offset = VIRTIO_XDP_HEADROOM; 702 } else { 703 xdp_page = page; 704 } 705 706 /* Transient failure which in theory could occur if 707 * in-flight packets from before XDP was enabled reach 708 * the receive path after XDP is loaded. In practice I 709 * was not able to create this condition. 710 */ 711 if (unlikely(hdr->hdr.gso_type)) 712 goto err_xdp; 713 714 /* Allow consuming headroom but reserve enough space to push 715 * the descriptor on if we get an XDP_TX return code. 716 */ 717 data = page_address(xdp_page) + offset; 718 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len; 719 xdp.data = data + vi->hdr_len; 720 xdp_set_data_meta_invalid(&xdp); 721 xdp.data_end = xdp.data + (len - vi->hdr_len); 722 xdp.rxq = &rq->xdp_rxq; 723 724 act = bpf_prog_run_xdp(xdp_prog, &xdp); 725 726 if (act != XDP_PASS) 727 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len); 728 729 switch (act) { 730 case XDP_PASS: 731 /* recalculate offset to account for any header 732 * adjustments. Note other cases do not build an 733 * skb and avoid using offset 734 */ 735 offset = xdp.data - 736 page_address(xdp_page) - vi->hdr_len; 737 738 /* We can only create skb based on xdp_page. */ 739 if (unlikely(xdp_page != page)) { 740 rcu_read_unlock(); 741 put_page(page); 742 head_skb = page_to_skb(vi, rq, xdp_page, 743 offset, len, PAGE_SIZE); 744 return head_skb; 745 } 746 break; 747 case XDP_TX: 748 if (unlikely(!__virtnet_xdp_xmit(vi, &xdp))) 749 trace_xdp_exception(vi->dev, xdp_prog, act); 750 else 751 *xdp_xmit = true; 752 if (unlikely(xdp_page != page)) 753 goto err_xdp; 754 rcu_read_unlock(); 755 goto xdp_xmit; 756 default: 757 bpf_warn_invalid_xdp_action(act); 758 case XDP_ABORTED: 759 trace_xdp_exception(vi->dev, xdp_prog, act); 760 case XDP_DROP: 761 if (unlikely(xdp_page != page)) 762 __free_pages(xdp_page, 0); 763 goto err_xdp; 764 } 765 } 766 rcu_read_unlock(); 767 768 truesize = mergeable_ctx_to_truesize(ctx); 769 if (unlikely(len > truesize)) { 770 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 771 dev->name, len, (unsigned long)ctx); 772 dev->stats.rx_length_errors++; 773 goto err_skb; 774 } 775 776 head_skb = page_to_skb(vi, rq, page, offset, len, truesize); 777 curr_skb = head_skb; 778 779 if (unlikely(!curr_skb)) 780 goto err_skb; 781 while (--num_buf) { 782 int num_skb_frags; 783 784 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx); 785 if (unlikely(!buf)) { 786 pr_debug("%s: rx error: %d buffers out of %d missing\n", 787 dev->name, num_buf, 788 virtio16_to_cpu(vi->vdev, 789 hdr->num_buffers)); 790 dev->stats.rx_length_errors++; 791 goto err_buf; 792 } 793 794 page = virt_to_head_page(buf); 795 796 truesize = mergeable_ctx_to_truesize(ctx); 797 if (unlikely(len > truesize)) { 798 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 799 dev->name, len, (unsigned long)ctx); 800 dev->stats.rx_length_errors++; 801 goto err_skb; 802 } 803 804 num_skb_frags = skb_shinfo(curr_skb)->nr_frags; 805 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) { 806 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC); 807 808 if (unlikely(!nskb)) 809 goto err_skb; 810 if (curr_skb == head_skb) 811 skb_shinfo(curr_skb)->frag_list = nskb; 812 else 813 curr_skb->next = nskb; 814 curr_skb = nskb; 815 head_skb->truesize += nskb->truesize; 816 num_skb_frags = 0; 817 } 818 if (curr_skb != head_skb) { 819 head_skb->data_len += len; 820 head_skb->len += len; 821 head_skb->truesize += truesize; 822 } 823 offset = buf - page_address(page); 824 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) { 825 put_page(page); 826 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1, 827 len, truesize); 828 } else { 829 skb_add_rx_frag(curr_skb, num_skb_frags, page, 830 offset, len, truesize); 831 } 832 } 833 834 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len); 835 return head_skb; 836 837 err_xdp: 838 rcu_read_unlock(); 839 err_skb: 840 put_page(page); 841 while (--num_buf) { 842 buf = virtqueue_get_buf(rq->vq, &len); 843 if (unlikely(!buf)) { 844 pr_debug("%s: rx error: %d buffers missing\n", 845 dev->name, num_buf); 846 dev->stats.rx_length_errors++; 847 break; 848 } 849 page = virt_to_head_page(buf); 850 put_page(page); 851 } 852 err_buf: 853 dev->stats.rx_dropped++; 854 dev_kfree_skb(head_skb); 855 xdp_xmit: 856 return NULL; 857 } 858 859 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq, 860 void *buf, unsigned int len, void **ctx, bool *xdp_xmit) 861 { 862 struct net_device *dev = vi->dev; 863 struct sk_buff *skb; 864 struct virtio_net_hdr_mrg_rxbuf *hdr; 865 int ret; 866 867 if (unlikely(len < vi->hdr_len + ETH_HLEN)) { 868 pr_debug("%s: short packet %i\n", dev->name, len); 869 dev->stats.rx_length_errors++; 870 if (vi->mergeable_rx_bufs) { 871 put_page(virt_to_head_page(buf)); 872 } else if (vi->big_packets) { 873 give_pages(rq, buf); 874 } else { 875 put_page(virt_to_head_page(buf)); 876 } 877 return 0; 878 } 879 880 if (vi->mergeable_rx_bufs) 881 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit); 882 else if (vi->big_packets) 883 skb = receive_big(dev, vi, rq, buf, len); 884 else 885 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit); 886 887 if (unlikely(!skb)) 888 return 0; 889 890 hdr = skb_vnet_hdr(skb); 891 892 ret = skb->len; 893 894 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) 895 skb->ip_summed = CHECKSUM_UNNECESSARY; 896 897 if (virtio_net_hdr_to_skb(skb, &hdr->hdr, 898 virtio_is_little_endian(vi->vdev))) { 899 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n", 900 dev->name, hdr->hdr.gso_type, 901 hdr->hdr.gso_size); 902 goto frame_err; 903 } 904 905 skb->protocol = eth_type_trans(skb, dev); 906 pr_debug("Receiving skb proto 0x%04x len %i type %i\n", 907 ntohs(skb->protocol), skb->len, skb->pkt_type); 908 909 napi_gro_receive(&rq->napi, skb); 910 return ret; 911 912 frame_err: 913 dev->stats.rx_frame_errors++; 914 dev_kfree_skb(skb); 915 return 0; 916 } 917 918 /* Unlike mergeable buffers, all buffers are allocated to the 919 * same size, except for the headroom. For this reason we do 920 * not need to use mergeable_len_to_ctx here - it is enough 921 * to store the headroom as the context ignoring the truesize. 922 */ 923 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq, 924 gfp_t gfp) 925 { 926 struct page_frag *alloc_frag = &rq->alloc_frag; 927 char *buf; 928 unsigned int xdp_headroom = virtnet_get_headroom(vi); 929 void *ctx = (void *)(unsigned long)xdp_headroom; 930 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom; 931 int err; 932 933 len = SKB_DATA_ALIGN(len) + 934 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 935 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp))) 936 return -ENOMEM; 937 938 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 939 get_page(alloc_frag->page); 940 alloc_frag->offset += len; 941 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom, 942 vi->hdr_len + GOOD_PACKET_LEN); 943 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 944 if (err < 0) 945 put_page(virt_to_head_page(buf)); 946 return err; 947 } 948 949 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq, 950 gfp_t gfp) 951 { 952 struct page *first, *list = NULL; 953 char *p; 954 int i, err, offset; 955 956 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2); 957 958 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */ 959 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) { 960 first = get_a_page(rq, gfp); 961 if (!first) { 962 if (list) 963 give_pages(rq, list); 964 return -ENOMEM; 965 } 966 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE); 967 968 /* chain new page in list head to match sg */ 969 first->private = (unsigned long)list; 970 list = first; 971 } 972 973 first = get_a_page(rq, gfp); 974 if (!first) { 975 give_pages(rq, list); 976 return -ENOMEM; 977 } 978 p = page_address(first); 979 980 /* rq->sg[0], rq->sg[1] share the same page */ 981 /* a separated rq->sg[0] for header - required in case !any_header_sg */ 982 sg_set_buf(&rq->sg[0], p, vi->hdr_len); 983 984 /* rq->sg[1] for data packet, from offset */ 985 offset = sizeof(struct padded_vnet_hdr); 986 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset); 987 988 /* chain first in list head */ 989 first->private = (unsigned long)list; 990 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2, 991 first, gfp); 992 if (err < 0) 993 give_pages(rq, first); 994 995 return err; 996 } 997 998 static unsigned int get_mergeable_buf_len(struct receive_queue *rq, 999 struct ewma_pkt_len *avg_pkt_len) 1000 { 1001 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 1002 unsigned int len; 1003 1004 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len), 1005 rq->min_buf_len, PAGE_SIZE - hdr_len); 1006 return ALIGN(len, L1_CACHE_BYTES); 1007 } 1008 1009 static int add_recvbuf_mergeable(struct virtnet_info *vi, 1010 struct receive_queue *rq, gfp_t gfp) 1011 { 1012 struct page_frag *alloc_frag = &rq->alloc_frag; 1013 unsigned int headroom = virtnet_get_headroom(vi); 1014 char *buf; 1015 void *ctx; 1016 int err; 1017 unsigned int len, hole; 1018 1019 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len); 1020 if (unlikely(!skb_page_frag_refill(len + headroom, alloc_frag, gfp))) 1021 return -ENOMEM; 1022 1023 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1024 buf += headroom; /* advance address leaving hole at front of pkt */ 1025 get_page(alloc_frag->page); 1026 alloc_frag->offset += len + headroom; 1027 hole = alloc_frag->size - alloc_frag->offset; 1028 if (hole < len + headroom) { 1029 /* To avoid internal fragmentation, if there is very likely not 1030 * enough space for another buffer, add the remaining space to 1031 * the current buffer. 1032 */ 1033 len += hole; 1034 alloc_frag->offset += hole; 1035 } 1036 1037 sg_init_one(rq->sg, buf, len); 1038 ctx = mergeable_len_to_ctx(len, headroom); 1039 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1040 if (err < 0) 1041 put_page(virt_to_head_page(buf)); 1042 1043 return err; 1044 } 1045 1046 /* 1047 * Returns false if we couldn't fill entirely (OOM). 1048 * 1049 * Normally run in the receive path, but can also be run from ndo_open 1050 * before we're receiving packets, or from refill_work which is 1051 * careful to disable receiving (using napi_disable). 1052 */ 1053 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq, 1054 gfp_t gfp) 1055 { 1056 int err; 1057 bool oom; 1058 1059 do { 1060 if (vi->mergeable_rx_bufs) 1061 err = add_recvbuf_mergeable(vi, rq, gfp); 1062 else if (vi->big_packets) 1063 err = add_recvbuf_big(vi, rq, gfp); 1064 else 1065 err = add_recvbuf_small(vi, rq, gfp); 1066 1067 oom = err == -ENOMEM; 1068 if (err) 1069 break; 1070 } while (rq->vq->num_free); 1071 virtqueue_kick(rq->vq); 1072 return !oom; 1073 } 1074 1075 static void skb_recv_done(struct virtqueue *rvq) 1076 { 1077 struct virtnet_info *vi = rvq->vdev->priv; 1078 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)]; 1079 1080 virtqueue_napi_schedule(&rq->napi, rvq); 1081 } 1082 1083 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi) 1084 { 1085 napi_enable(napi); 1086 1087 /* If all buffers were filled by other side before we napi_enabled, we 1088 * won't get another interrupt, so process any outstanding packets now. 1089 * Call local_bh_enable after to trigger softIRQ processing. 1090 */ 1091 local_bh_disable(); 1092 virtqueue_napi_schedule(napi, vq); 1093 local_bh_enable(); 1094 } 1095 1096 static void virtnet_napi_tx_enable(struct virtnet_info *vi, 1097 struct virtqueue *vq, 1098 struct napi_struct *napi) 1099 { 1100 if (!napi->weight) 1101 return; 1102 1103 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only 1104 * enable the feature if this is likely affine with the transmit path. 1105 */ 1106 if (!vi->affinity_hint_set) { 1107 napi->weight = 0; 1108 return; 1109 } 1110 1111 return virtnet_napi_enable(vq, napi); 1112 } 1113 1114 static void virtnet_napi_tx_disable(struct napi_struct *napi) 1115 { 1116 if (napi->weight) 1117 napi_disable(napi); 1118 } 1119 1120 static void refill_work(struct work_struct *work) 1121 { 1122 struct virtnet_info *vi = 1123 container_of(work, struct virtnet_info, refill.work); 1124 bool still_empty; 1125 int i; 1126 1127 for (i = 0; i < vi->curr_queue_pairs; i++) { 1128 struct receive_queue *rq = &vi->rq[i]; 1129 1130 napi_disable(&rq->napi); 1131 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL); 1132 virtnet_napi_enable(rq->vq, &rq->napi); 1133 1134 /* In theory, this can happen: if we don't get any buffers in 1135 * we will *never* try to fill again. 1136 */ 1137 if (still_empty) 1138 schedule_delayed_work(&vi->refill, HZ/2); 1139 } 1140 } 1141 1142 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit) 1143 { 1144 struct virtnet_info *vi = rq->vq->vdev->priv; 1145 unsigned int len, received = 0, bytes = 0; 1146 void *buf; 1147 1148 if (!vi->big_packets || vi->mergeable_rx_bufs) { 1149 void *ctx; 1150 1151 while (received < budget && 1152 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) { 1153 bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit); 1154 received++; 1155 } 1156 } else { 1157 while (received < budget && 1158 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) { 1159 bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit); 1160 received++; 1161 } 1162 } 1163 1164 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) { 1165 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) 1166 schedule_delayed_work(&vi->refill, 0); 1167 } 1168 1169 u64_stats_update_begin(&rq->stats.syncp); 1170 rq->stats.bytes += bytes; 1171 rq->stats.packets += received; 1172 u64_stats_update_end(&rq->stats.syncp); 1173 1174 return received; 1175 } 1176 1177 static void free_old_xmit_skbs(struct send_queue *sq) 1178 { 1179 struct sk_buff *skb; 1180 unsigned int len; 1181 unsigned int packets = 0; 1182 unsigned int bytes = 0; 1183 1184 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) { 1185 pr_debug("Sent skb %p\n", skb); 1186 1187 bytes += skb->len; 1188 packets++; 1189 1190 dev_consume_skb_any(skb); 1191 } 1192 1193 /* Avoid overhead when no packets have been processed 1194 * happens when called speculatively from start_xmit. 1195 */ 1196 if (!packets) 1197 return; 1198 1199 u64_stats_update_begin(&sq->stats.syncp); 1200 sq->stats.bytes += bytes; 1201 sq->stats.packets += packets; 1202 u64_stats_update_end(&sq->stats.syncp); 1203 } 1204 1205 static void virtnet_poll_cleantx(struct receive_queue *rq) 1206 { 1207 struct virtnet_info *vi = rq->vq->vdev->priv; 1208 unsigned int index = vq2rxq(rq->vq); 1209 struct send_queue *sq = &vi->sq[index]; 1210 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index); 1211 1212 if (!sq->napi.weight) 1213 return; 1214 1215 if (__netif_tx_trylock(txq)) { 1216 free_old_xmit_skbs(sq); 1217 __netif_tx_unlock(txq); 1218 } 1219 1220 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1221 netif_tx_wake_queue(txq); 1222 } 1223 1224 static int virtnet_poll(struct napi_struct *napi, int budget) 1225 { 1226 struct receive_queue *rq = 1227 container_of(napi, struct receive_queue, napi); 1228 unsigned int received; 1229 bool xdp_xmit = false; 1230 1231 virtnet_poll_cleantx(rq); 1232 1233 received = virtnet_receive(rq, budget, &xdp_xmit); 1234 1235 /* Out of packets? */ 1236 if (received < budget) 1237 virtqueue_napi_complete(napi, rq->vq, received); 1238 1239 if (xdp_xmit) 1240 xdp_do_flush_map(); 1241 1242 return received; 1243 } 1244 1245 static int virtnet_open(struct net_device *dev) 1246 { 1247 struct virtnet_info *vi = netdev_priv(dev); 1248 int i, err; 1249 1250 for (i = 0; i < vi->max_queue_pairs; i++) { 1251 if (i < vi->curr_queue_pairs) 1252 /* Make sure we have some buffers: if oom use wq. */ 1253 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 1254 schedule_delayed_work(&vi->refill, 0); 1255 1256 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i); 1257 if (err < 0) 1258 return err; 1259 1260 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 1261 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi); 1262 } 1263 1264 return 0; 1265 } 1266 1267 static int virtnet_poll_tx(struct napi_struct *napi, int budget) 1268 { 1269 struct send_queue *sq = container_of(napi, struct send_queue, napi); 1270 struct virtnet_info *vi = sq->vq->vdev->priv; 1271 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq)); 1272 1273 __netif_tx_lock(txq, raw_smp_processor_id()); 1274 free_old_xmit_skbs(sq); 1275 __netif_tx_unlock(txq); 1276 1277 virtqueue_napi_complete(napi, sq->vq, 0); 1278 1279 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1280 netif_tx_wake_queue(txq); 1281 1282 return 0; 1283 } 1284 1285 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb) 1286 { 1287 struct virtio_net_hdr_mrg_rxbuf *hdr; 1288 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest; 1289 struct virtnet_info *vi = sq->vq->vdev->priv; 1290 int num_sg; 1291 unsigned hdr_len = vi->hdr_len; 1292 bool can_push; 1293 1294 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest); 1295 1296 can_push = vi->any_header_sg && 1297 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) && 1298 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len; 1299 /* Even if we can, don't push here yet as this would skew 1300 * csum_start offset below. */ 1301 if (can_push) 1302 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len); 1303 else 1304 hdr = skb_vnet_hdr(skb); 1305 1306 if (virtio_net_hdr_from_skb(skb, &hdr->hdr, 1307 virtio_is_little_endian(vi->vdev), false)) 1308 BUG(); 1309 1310 if (vi->mergeable_rx_bufs) 1311 hdr->num_buffers = 0; 1312 1313 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2)); 1314 if (can_push) { 1315 __skb_push(skb, hdr_len); 1316 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len); 1317 if (unlikely(num_sg < 0)) 1318 return num_sg; 1319 /* Pull header back to avoid skew in tx bytes calculations. */ 1320 __skb_pull(skb, hdr_len); 1321 } else { 1322 sg_set_buf(sq->sg, hdr, hdr_len); 1323 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len); 1324 if (unlikely(num_sg < 0)) 1325 return num_sg; 1326 num_sg++; 1327 } 1328 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC); 1329 } 1330 1331 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) 1332 { 1333 struct virtnet_info *vi = netdev_priv(dev); 1334 int qnum = skb_get_queue_mapping(skb); 1335 struct send_queue *sq = &vi->sq[qnum]; 1336 int err; 1337 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum); 1338 bool kick = !skb->xmit_more; 1339 bool use_napi = sq->napi.weight; 1340 1341 /* Free up any pending old buffers before queueing new ones. */ 1342 free_old_xmit_skbs(sq); 1343 1344 if (use_napi && kick) 1345 virtqueue_enable_cb_delayed(sq->vq); 1346 1347 /* timestamp packet in software */ 1348 skb_tx_timestamp(skb); 1349 1350 /* Try to transmit */ 1351 err = xmit_skb(sq, skb); 1352 1353 /* This should not happen! */ 1354 if (unlikely(err)) { 1355 dev->stats.tx_fifo_errors++; 1356 if (net_ratelimit()) 1357 dev_warn(&dev->dev, 1358 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err); 1359 dev->stats.tx_dropped++; 1360 dev_kfree_skb_any(skb); 1361 return NETDEV_TX_OK; 1362 } 1363 1364 /* Don't wait up for transmitted skbs to be freed. */ 1365 if (!use_napi) { 1366 skb_orphan(skb); 1367 nf_reset(skb); 1368 } 1369 1370 /* If running out of space, stop queue to avoid getting packets that we 1371 * are then unable to transmit. 1372 * An alternative would be to force queuing layer to requeue the skb by 1373 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be 1374 * returned in a normal path of operation: it means that driver is not 1375 * maintaining the TX queue stop/start state properly, and causes 1376 * the stack to do a non-trivial amount of useless work. 1377 * Since most packets only take 1 or 2 ring slots, stopping the queue 1378 * early means 16 slots are typically wasted. 1379 */ 1380 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) { 1381 netif_stop_subqueue(dev, qnum); 1382 if (!use_napi && 1383 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) { 1384 /* More just got used, free them then recheck. */ 1385 free_old_xmit_skbs(sq); 1386 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) { 1387 netif_start_subqueue(dev, qnum); 1388 virtqueue_disable_cb(sq->vq); 1389 } 1390 } 1391 } 1392 1393 if (kick || netif_xmit_stopped(txq)) 1394 virtqueue_kick(sq->vq); 1395 1396 return NETDEV_TX_OK; 1397 } 1398 1399 /* 1400 * Send command via the control virtqueue and check status. Commands 1401 * supported by the hypervisor, as indicated by feature bits, should 1402 * never fail unless improperly formatted. 1403 */ 1404 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd, 1405 struct scatterlist *out) 1406 { 1407 struct scatterlist *sgs[4], hdr, stat; 1408 unsigned out_num = 0, tmp; 1409 1410 /* Caller should know better */ 1411 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)); 1412 1413 vi->ctrl_status = ~0; 1414 vi->ctrl_hdr.class = class; 1415 vi->ctrl_hdr.cmd = cmd; 1416 /* Add header */ 1417 sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr)); 1418 sgs[out_num++] = &hdr; 1419 1420 if (out) 1421 sgs[out_num++] = out; 1422 1423 /* Add return status. */ 1424 sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status)); 1425 sgs[out_num] = &stat; 1426 1427 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs)); 1428 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC); 1429 1430 if (unlikely(!virtqueue_kick(vi->cvq))) 1431 return vi->ctrl_status == VIRTIO_NET_OK; 1432 1433 /* Spin for a response, the kick causes an ioport write, trapping 1434 * into the hypervisor, so the request should be handled immediately. 1435 */ 1436 while (!virtqueue_get_buf(vi->cvq, &tmp) && 1437 !virtqueue_is_broken(vi->cvq)) 1438 cpu_relax(); 1439 1440 return vi->ctrl_status == VIRTIO_NET_OK; 1441 } 1442 1443 static int virtnet_set_mac_address(struct net_device *dev, void *p) 1444 { 1445 struct virtnet_info *vi = netdev_priv(dev); 1446 struct virtio_device *vdev = vi->vdev; 1447 int ret; 1448 struct sockaddr *addr; 1449 struct scatterlist sg; 1450 1451 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL); 1452 if (!addr) 1453 return -ENOMEM; 1454 1455 ret = eth_prepare_mac_addr_change(dev, addr); 1456 if (ret) 1457 goto out; 1458 1459 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 1460 sg_init_one(&sg, addr->sa_data, dev->addr_len); 1461 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 1462 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) { 1463 dev_warn(&vdev->dev, 1464 "Failed to set mac address by vq command.\n"); 1465 ret = -EINVAL; 1466 goto out; 1467 } 1468 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) && 1469 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) { 1470 unsigned int i; 1471 1472 /* Naturally, this has an atomicity problem. */ 1473 for (i = 0; i < dev->addr_len; i++) 1474 virtio_cwrite8(vdev, 1475 offsetof(struct virtio_net_config, mac) + 1476 i, addr->sa_data[i]); 1477 } 1478 1479 eth_commit_mac_addr_change(dev, p); 1480 ret = 0; 1481 1482 out: 1483 kfree(addr); 1484 return ret; 1485 } 1486 1487 static void virtnet_stats(struct net_device *dev, 1488 struct rtnl_link_stats64 *tot) 1489 { 1490 struct virtnet_info *vi = netdev_priv(dev); 1491 unsigned int start; 1492 int i; 1493 1494 for (i = 0; i < vi->max_queue_pairs; i++) { 1495 u64 tpackets, tbytes, rpackets, rbytes; 1496 struct receive_queue *rq = &vi->rq[i]; 1497 struct send_queue *sq = &vi->sq[i]; 1498 1499 do { 1500 start = u64_stats_fetch_begin_irq(&sq->stats.syncp); 1501 tpackets = sq->stats.packets; 1502 tbytes = sq->stats.bytes; 1503 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start)); 1504 1505 do { 1506 start = u64_stats_fetch_begin_irq(&rq->stats.syncp); 1507 rpackets = rq->stats.packets; 1508 rbytes = rq->stats.bytes; 1509 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start)); 1510 1511 tot->rx_packets += rpackets; 1512 tot->tx_packets += tpackets; 1513 tot->rx_bytes += rbytes; 1514 tot->tx_bytes += tbytes; 1515 } 1516 1517 tot->tx_dropped = dev->stats.tx_dropped; 1518 tot->tx_fifo_errors = dev->stats.tx_fifo_errors; 1519 tot->rx_dropped = dev->stats.rx_dropped; 1520 tot->rx_length_errors = dev->stats.rx_length_errors; 1521 tot->rx_frame_errors = dev->stats.rx_frame_errors; 1522 } 1523 1524 #ifdef CONFIG_NET_POLL_CONTROLLER 1525 static void virtnet_netpoll(struct net_device *dev) 1526 { 1527 struct virtnet_info *vi = netdev_priv(dev); 1528 int i; 1529 1530 for (i = 0; i < vi->curr_queue_pairs; i++) 1531 napi_schedule(&vi->rq[i].napi); 1532 } 1533 #endif 1534 1535 static void virtnet_ack_link_announce(struct virtnet_info *vi) 1536 { 1537 rtnl_lock(); 1538 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE, 1539 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL)) 1540 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n"); 1541 rtnl_unlock(); 1542 } 1543 1544 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 1545 { 1546 struct scatterlist sg; 1547 struct net_device *dev = vi->dev; 1548 1549 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ)) 1550 return 0; 1551 1552 vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs); 1553 sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq)); 1554 1555 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ, 1556 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) { 1557 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n", 1558 queue_pairs); 1559 return -EINVAL; 1560 } else { 1561 vi->curr_queue_pairs = queue_pairs; 1562 /* virtnet_open() will refill when device is going to up. */ 1563 if (dev->flags & IFF_UP) 1564 schedule_delayed_work(&vi->refill, 0); 1565 } 1566 1567 return 0; 1568 } 1569 1570 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 1571 { 1572 int err; 1573 1574 rtnl_lock(); 1575 err = _virtnet_set_queues(vi, queue_pairs); 1576 rtnl_unlock(); 1577 return err; 1578 } 1579 1580 static int virtnet_close(struct net_device *dev) 1581 { 1582 struct virtnet_info *vi = netdev_priv(dev); 1583 int i; 1584 1585 /* Make sure refill_work doesn't re-enable napi! */ 1586 cancel_delayed_work_sync(&vi->refill); 1587 1588 for (i = 0; i < vi->max_queue_pairs; i++) { 1589 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 1590 napi_disable(&vi->rq[i].napi); 1591 virtnet_napi_tx_disable(&vi->sq[i].napi); 1592 } 1593 1594 return 0; 1595 } 1596 1597 static void virtnet_set_rx_mode(struct net_device *dev) 1598 { 1599 struct virtnet_info *vi = netdev_priv(dev); 1600 struct scatterlist sg[2]; 1601 struct virtio_net_ctrl_mac *mac_data; 1602 struct netdev_hw_addr *ha; 1603 int uc_count; 1604 int mc_count; 1605 void *buf; 1606 int i; 1607 1608 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */ 1609 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX)) 1610 return; 1611 1612 vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0); 1613 vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0); 1614 1615 sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc)); 1616 1617 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 1618 VIRTIO_NET_CTRL_RX_PROMISC, sg)) 1619 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n", 1620 vi->ctrl_promisc ? "en" : "dis"); 1621 1622 sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti)); 1623 1624 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 1625 VIRTIO_NET_CTRL_RX_ALLMULTI, sg)) 1626 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n", 1627 vi->ctrl_allmulti ? "en" : "dis"); 1628 1629 uc_count = netdev_uc_count(dev); 1630 mc_count = netdev_mc_count(dev); 1631 /* MAC filter - use one buffer for both lists */ 1632 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) + 1633 (2 * sizeof(mac_data->entries)), GFP_ATOMIC); 1634 mac_data = buf; 1635 if (!buf) 1636 return; 1637 1638 sg_init_table(sg, 2); 1639 1640 /* Store the unicast list and count in the front of the buffer */ 1641 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count); 1642 i = 0; 1643 netdev_for_each_uc_addr(ha, dev) 1644 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 1645 1646 sg_set_buf(&sg[0], mac_data, 1647 sizeof(mac_data->entries) + (uc_count * ETH_ALEN)); 1648 1649 /* multicast list and count fill the end */ 1650 mac_data = (void *)&mac_data->macs[uc_count][0]; 1651 1652 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count); 1653 i = 0; 1654 netdev_for_each_mc_addr(ha, dev) 1655 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 1656 1657 sg_set_buf(&sg[1], mac_data, 1658 sizeof(mac_data->entries) + (mc_count * ETH_ALEN)); 1659 1660 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 1661 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg)) 1662 dev_warn(&dev->dev, "Failed to set MAC filter table.\n"); 1663 1664 kfree(buf); 1665 } 1666 1667 static int virtnet_vlan_rx_add_vid(struct net_device *dev, 1668 __be16 proto, u16 vid) 1669 { 1670 struct virtnet_info *vi = netdev_priv(dev); 1671 struct scatterlist sg; 1672 1673 vi->ctrl_vid = vid; 1674 sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid)); 1675 1676 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 1677 VIRTIO_NET_CTRL_VLAN_ADD, &sg)) 1678 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid); 1679 return 0; 1680 } 1681 1682 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, 1683 __be16 proto, u16 vid) 1684 { 1685 struct virtnet_info *vi = netdev_priv(dev); 1686 struct scatterlist sg; 1687 1688 vi->ctrl_vid = vid; 1689 sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid)); 1690 1691 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 1692 VIRTIO_NET_CTRL_VLAN_DEL, &sg)) 1693 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid); 1694 return 0; 1695 } 1696 1697 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu) 1698 { 1699 int i; 1700 1701 if (vi->affinity_hint_set) { 1702 for (i = 0; i < vi->max_queue_pairs; i++) { 1703 virtqueue_set_affinity(vi->rq[i].vq, -1); 1704 virtqueue_set_affinity(vi->sq[i].vq, -1); 1705 } 1706 1707 vi->affinity_hint_set = false; 1708 } 1709 } 1710 1711 static void virtnet_set_affinity(struct virtnet_info *vi) 1712 { 1713 int i; 1714 int cpu; 1715 1716 /* In multiqueue mode, when the number of cpu is equal to the number of 1717 * queue pairs, we let the queue pairs to be private to one cpu by 1718 * setting the affinity hint to eliminate the contention. 1719 */ 1720 if (vi->curr_queue_pairs == 1 || 1721 vi->max_queue_pairs != num_online_cpus()) { 1722 virtnet_clean_affinity(vi, -1); 1723 return; 1724 } 1725 1726 i = 0; 1727 for_each_online_cpu(cpu) { 1728 virtqueue_set_affinity(vi->rq[i].vq, cpu); 1729 virtqueue_set_affinity(vi->sq[i].vq, cpu); 1730 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i); 1731 i++; 1732 } 1733 1734 vi->affinity_hint_set = true; 1735 } 1736 1737 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node) 1738 { 1739 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1740 node); 1741 virtnet_set_affinity(vi); 1742 return 0; 1743 } 1744 1745 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node) 1746 { 1747 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1748 node_dead); 1749 virtnet_set_affinity(vi); 1750 return 0; 1751 } 1752 1753 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node) 1754 { 1755 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 1756 node); 1757 1758 virtnet_clean_affinity(vi, cpu); 1759 return 0; 1760 } 1761 1762 static enum cpuhp_state virtionet_online; 1763 1764 static int virtnet_cpu_notif_add(struct virtnet_info *vi) 1765 { 1766 int ret; 1767 1768 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node); 1769 if (ret) 1770 return ret; 1771 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD, 1772 &vi->node_dead); 1773 if (!ret) 1774 return ret; 1775 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 1776 return ret; 1777 } 1778 1779 static void virtnet_cpu_notif_remove(struct virtnet_info *vi) 1780 { 1781 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 1782 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD, 1783 &vi->node_dead); 1784 } 1785 1786 static void virtnet_get_ringparam(struct net_device *dev, 1787 struct ethtool_ringparam *ring) 1788 { 1789 struct virtnet_info *vi = netdev_priv(dev); 1790 1791 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq); 1792 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq); 1793 ring->rx_pending = ring->rx_max_pending; 1794 ring->tx_pending = ring->tx_max_pending; 1795 } 1796 1797 1798 static void virtnet_get_drvinfo(struct net_device *dev, 1799 struct ethtool_drvinfo *info) 1800 { 1801 struct virtnet_info *vi = netdev_priv(dev); 1802 struct virtio_device *vdev = vi->vdev; 1803 1804 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver)); 1805 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version)); 1806 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info)); 1807 1808 } 1809 1810 /* TODO: Eliminate OOO packets during switching */ 1811 static int virtnet_set_channels(struct net_device *dev, 1812 struct ethtool_channels *channels) 1813 { 1814 struct virtnet_info *vi = netdev_priv(dev); 1815 u16 queue_pairs = channels->combined_count; 1816 int err; 1817 1818 /* We don't support separate rx/tx channels. 1819 * We don't allow setting 'other' channels. 1820 */ 1821 if (channels->rx_count || channels->tx_count || channels->other_count) 1822 return -EINVAL; 1823 1824 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0) 1825 return -EINVAL; 1826 1827 /* For now we don't support modifying channels while XDP is loaded 1828 * also when XDP is loaded all RX queues have XDP programs so we only 1829 * need to check a single RX queue. 1830 */ 1831 if (vi->rq[0].xdp_prog) 1832 return -EINVAL; 1833 1834 get_online_cpus(); 1835 err = _virtnet_set_queues(vi, queue_pairs); 1836 if (!err) { 1837 netif_set_real_num_tx_queues(dev, queue_pairs); 1838 netif_set_real_num_rx_queues(dev, queue_pairs); 1839 1840 virtnet_set_affinity(vi); 1841 } 1842 put_online_cpus(); 1843 1844 return err; 1845 } 1846 1847 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data) 1848 { 1849 struct virtnet_info *vi = netdev_priv(dev); 1850 char *p = (char *)data; 1851 unsigned int i, j; 1852 1853 switch (stringset) { 1854 case ETH_SS_STATS: 1855 for (i = 0; i < vi->curr_queue_pairs; i++) { 1856 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 1857 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s", 1858 i, virtnet_rq_stats_desc[j].desc); 1859 p += ETH_GSTRING_LEN; 1860 } 1861 } 1862 1863 for (i = 0; i < vi->curr_queue_pairs; i++) { 1864 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 1865 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s", 1866 i, virtnet_sq_stats_desc[j].desc); 1867 p += ETH_GSTRING_LEN; 1868 } 1869 } 1870 break; 1871 } 1872 } 1873 1874 static int virtnet_get_sset_count(struct net_device *dev, int sset) 1875 { 1876 struct virtnet_info *vi = netdev_priv(dev); 1877 1878 switch (sset) { 1879 case ETH_SS_STATS: 1880 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN + 1881 VIRTNET_SQ_STATS_LEN); 1882 default: 1883 return -EOPNOTSUPP; 1884 } 1885 } 1886 1887 static void virtnet_get_ethtool_stats(struct net_device *dev, 1888 struct ethtool_stats *stats, u64 *data) 1889 { 1890 struct virtnet_info *vi = netdev_priv(dev); 1891 unsigned int idx = 0, start, i, j; 1892 const u8 *stats_base; 1893 size_t offset; 1894 1895 for (i = 0; i < vi->curr_queue_pairs; i++) { 1896 struct receive_queue *rq = &vi->rq[i]; 1897 1898 stats_base = (u8 *)&rq->stats; 1899 do { 1900 start = u64_stats_fetch_begin_irq(&rq->stats.syncp); 1901 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 1902 offset = virtnet_rq_stats_desc[j].offset; 1903 data[idx + j] = *(u64 *)(stats_base + offset); 1904 } 1905 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start)); 1906 idx += VIRTNET_RQ_STATS_LEN; 1907 } 1908 1909 for (i = 0; i < vi->curr_queue_pairs; i++) { 1910 struct send_queue *sq = &vi->sq[i]; 1911 1912 stats_base = (u8 *)&sq->stats; 1913 do { 1914 start = u64_stats_fetch_begin_irq(&sq->stats.syncp); 1915 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 1916 offset = virtnet_sq_stats_desc[j].offset; 1917 data[idx + j] = *(u64 *)(stats_base + offset); 1918 } 1919 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start)); 1920 idx += VIRTNET_SQ_STATS_LEN; 1921 } 1922 } 1923 1924 static void virtnet_get_channels(struct net_device *dev, 1925 struct ethtool_channels *channels) 1926 { 1927 struct virtnet_info *vi = netdev_priv(dev); 1928 1929 channels->combined_count = vi->curr_queue_pairs; 1930 channels->max_combined = vi->max_queue_pairs; 1931 channels->max_other = 0; 1932 channels->rx_count = 0; 1933 channels->tx_count = 0; 1934 channels->other_count = 0; 1935 } 1936 1937 /* Check if the user is trying to change anything besides speed/duplex */ 1938 static bool 1939 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd) 1940 { 1941 struct ethtool_link_ksettings diff1 = *cmd; 1942 struct ethtool_link_ksettings diff2 = {}; 1943 1944 /* cmd is always set so we need to clear it, validate the port type 1945 * and also without autonegotiation we can ignore advertising 1946 */ 1947 diff1.base.speed = 0; 1948 diff2.base.port = PORT_OTHER; 1949 ethtool_link_ksettings_zero_link_mode(&diff1, advertising); 1950 diff1.base.duplex = 0; 1951 diff1.base.cmd = 0; 1952 diff1.base.link_mode_masks_nwords = 0; 1953 1954 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) && 1955 bitmap_empty(diff1.link_modes.supported, 1956 __ETHTOOL_LINK_MODE_MASK_NBITS) && 1957 bitmap_empty(diff1.link_modes.advertising, 1958 __ETHTOOL_LINK_MODE_MASK_NBITS) && 1959 bitmap_empty(diff1.link_modes.lp_advertising, 1960 __ETHTOOL_LINK_MODE_MASK_NBITS); 1961 } 1962 1963 static int virtnet_set_link_ksettings(struct net_device *dev, 1964 const struct ethtool_link_ksettings *cmd) 1965 { 1966 struct virtnet_info *vi = netdev_priv(dev); 1967 u32 speed; 1968 1969 speed = cmd->base.speed; 1970 /* don't allow custom speed and duplex */ 1971 if (!ethtool_validate_speed(speed) || 1972 !ethtool_validate_duplex(cmd->base.duplex) || 1973 !virtnet_validate_ethtool_cmd(cmd)) 1974 return -EINVAL; 1975 vi->speed = speed; 1976 vi->duplex = cmd->base.duplex; 1977 1978 return 0; 1979 } 1980 1981 static int virtnet_get_link_ksettings(struct net_device *dev, 1982 struct ethtool_link_ksettings *cmd) 1983 { 1984 struct virtnet_info *vi = netdev_priv(dev); 1985 1986 cmd->base.speed = vi->speed; 1987 cmd->base.duplex = vi->duplex; 1988 cmd->base.port = PORT_OTHER; 1989 1990 return 0; 1991 } 1992 1993 static void virtnet_init_settings(struct net_device *dev) 1994 { 1995 struct virtnet_info *vi = netdev_priv(dev); 1996 1997 vi->speed = SPEED_UNKNOWN; 1998 vi->duplex = DUPLEX_UNKNOWN; 1999 } 2000 2001 static void virtnet_update_settings(struct virtnet_info *vi) 2002 { 2003 u32 speed; 2004 u8 duplex; 2005 2006 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX)) 2007 return; 2008 2009 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config, 2010 speed)); 2011 if (ethtool_validate_speed(speed)) 2012 vi->speed = speed; 2013 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config, 2014 duplex)); 2015 if (ethtool_validate_duplex(duplex)) 2016 vi->duplex = duplex; 2017 } 2018 2019 static const struct ethtool_ops virtnet_ethtool_ops = { 2020 .get_drvinfo = virtnet_get_drvinfo, 2021 .get_link = ethtool_op_get_link, 2022 .get_ringparam = virtnet_get_ringparam, 2023 .get_strings = virtnet_get_strings, 2024 .get_sset_count = virtnet_get_sset_count, 2025 .get_ethtool_stats = virtnet_get_ethtool_stats, 2026 .set_channels = virtnet_set_channels, 2027 .get_channels = virtnet_get_channels, 2028 .get_ts_info = ethtool_op_get_ts_info, 2029 .get_link_ksettings = virtnet_get_link_ksettings, 2030 .set_link_ksettings = virtnet_set_link_ksettings, 2031 }; 2032 2033 static void virtnet_freeze_down(struct virtio_device *vdev) 2034 { 2035 struct virtnet_info *vi = vdev->priv; 2036 int i; 2037 2038 /* Make sure no work handler is accessing the device */ 2039 flush_work(&vi->config_work); 2040 2041 netif_device_detach(vi->dev); 2042 netif_tx_disable(vi->dev); 2043 cancel_delayed_work_sync(&vi->refill); 2044 2045 if (netif_running(vi->dev)) { 2046 for (i = 0; i < vi->max_queue_pairs; i++) { 2047 napi_disable(&vi->rq[i].napi); 2048 virtnet_napi_tx_disable(&vi->sq[i].napi); 2049 } 2050 } 2051 } 2052 2053 static int init_vqs(struct virtnet_info *vi); 2054 2055 static int virtnet_restore_up(struct virtio_device *vdev) 2056 { 2057 struct virtnet_info *vi = vdev->priv; 2058 int err, i; 2059 2060 err = init_vqs(vi); 2061 if (err) 2062 return err; 2063 2064 virtio_device_ready(vdev); 2065 2066 if (netif_running(vi->dev)) { 2067 for (i = 0; i < vi->curr_queue_pairs; i++) 2068 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 2069 schedule_delayed_work(&vi->refill, 0); 2070 2071 for (i = 0; i < vi->max_queue_pairs; i++) { 2072 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2073 virtnet_napi_tx_enable(vi, vi->sq[i].vq, 2074 &vi->sq[i].napi); 2075 } 2076 } 2077 2078 netif_device_attach(vi->dev); 2079 return err; 2080 } 2081 2082 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads) 2083 { 2084 struct scatterlist sg; 2085 vi->ctrl_offloads = cpu_to_virtio64(vi->vdev, offloads); 2086 2087 sg_init_one(&sg, &vi->ctrl_offloads, sizeof(vi->ctrl_offloads)); 2088 2089 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS, 2090 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) { 2091 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n"); 2092 return -EINVAL; 2093 } 2094 2095 return 0; 2096 } 2097 2098 static int virtnet_clear_guest_offloads(struct virtnet_info *vi) 2099 { 2100 u64 offloads = 0; 2101 2102 if (!vi->guest_offloads) 2103 return 0; 2104 2105 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM)) 2106 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM; 2107 2108 return virtnet_set_guest_offloads(vi, offloads); 2109 } 2110 2111 static int virtnet_restore_guest_offloads(struct virtnet_info *vi) 2112 { 2113 u64 offloads = vi->guest_offloads; 2114 2115 if (!vi->guest_offloads) 2116 return 0; 2117 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM)) 2118 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM; 2119 2120 return virtnet_set_guest_offloads(vi, offloads); 2121 } 2122 2123 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog, 2124 struct netlink_ext_ack *extack) 2125 { 2126 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr); 2127 struct virtnet_info *vi = netdev_priv(dev); 2128 struct bpf_prog *old_prog; 2129 u16 xdp_qp = 0, curr_qp; 2130 int i, err; 2131 2132 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) 2133 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) || 2134 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) || 2135 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) || 2136 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) { 2137 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first"); 2138 return -EOPNOTSUPP; 2139 } 2140 2141 if (vi->mergeable_rx_bufs && !vi->any_header_sg) { 2142 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required"); 2143 return -EINVAL; 2144 } 2145 2146 if (dev->mtu > max_sz) { 2147 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP"); 2148 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz); 2149 return -EINVAL; 2150 } 2151 2152 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs; 2153 if (prog) 2154 xdp_qp = nr_cpu_ids; 2155 2156 /* XDP requires extra queues for XDP_TX */ 2157 if (curr_qp + xdp_qp > vi->max_queue_pairs) { 2158 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available"); 2159 netdev_warn(dev, "request %i queues but max is %i\n", 2160 curr_qp + xdp_qp, vi->max_queue_pairs); 2161 return -ENOMEM; 2162 } 2163 2164 if (prog) { 2165 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1); 2166 if (IS_ERR(prog)) 2167 return PTR_ERR(prog); 2168 } 2169 2170 /* Make sure NAPI is not using any XDP TX queues for RX. */ 2171 for (i = 0; i < vi->max_queue_pairs; i++) 2172 napi_disable(&vi->rq[i].napi); 2173 2174 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp); 2175 err = _virtnet_set_queues(vi, curr_qp + xdp_qp); 2176 if (err) 2177 goto err; 2178 vi->xdp_queue_pairs = xdp_qp; 2179 2180 for (i = 0; i < vi->max_queue_pairs; i++) { 2181 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2182 rcu_assign_pointer(vi->rq[i].xdp_prog, prog); 2183 if (i == 0) { 2184 if (!old_prog) 2185 virtnet_clear_guest_offloads(vi); 2186 if (!prog) 2187 virtnet_restore_guest_offloads(vi); 2188 } 2189 if (old_prog) 2190 bpf_prog_put(old_prog); 2191 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2192 } 2193 2194 return 0; 2195 2196 err: 2197 for (i = 0; i < vi->max_queue_pairs; i++) 2198 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2199 if (prog) 2200 bpf_prog_sub(prog, vi->max_queue_pairs - 1); 2201 return err; 2202 } 2203 2204 static u32 virtnet_xdp_query(struct net_device *dev) 2205 { 2206 struct virtnet_info *vi = netdev_priv(dev); 2207 const struct bpf_prog *xdp_prog; 2208 int i; 2209 2210 for (i = 0; i < vi->max_queue_pairs; i++) { 2211 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2212 if (xdp_prog) 2213 return xdp_prog->aux->id; 2214 } 2215 return 0; 2216 } 2217 2218 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp) 2219 { 2220 switch (xdp->command) { 2221 case XDP_SETUP_PROG: 2222 return virtnet_xdp_set(dev, xdp->prog, xdp->extack); 2223 case XDP_QUERY_PROG: 2224 xdp->prog_id = virtnet_xdp_query(dev); 2225 xdp->prog_attached = !!xdp->prog_id; 2226 return 0; 2227 default: 2228 return -EINVAL; 2229 } 2230 } 2231 2232 static const struct net_device_ops virtnet_netdev = { 2233 .ndo_open = virtnet_open, 2234 .ndo_stop = virtnet_close, 2235 .ndo_start_xmit = start_xmit, 2236 .ndo_validate_addr = eth_validate_addr, 2237 .ndo_set_mac_address = virtnet_set_mac_address, 2238 .ndo_set_rx_mode = virtnet_set_rx_mode, 2239 .ndo_get_stats64 = virtnet_stats, 2240 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid, 2241 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid, 2242 #ifdef CONFIG_NET_POLL_CONTROLLER 2243 .ndo_poll_controller = virtnet_netpoll, 2244 #endif 2245 .ndo_bpf = virtnet_xdp, 2246 .ndo_xdp_xmit = virtnet_xdp_xmit, 2247 .ndo_xdp_flush = virtnet_xdp_flush, 2248 .ndo_features_check = passthru_features_check, 2249 }; 2250 2251 static void virtnet_config_changed_work(struct work_struct *work) 2252 { 2253 struct virtnet_info *vi = 2254 container_of(work, struct virtnet_info, config_work); 2255 u16 v; 2256 2257 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS, 2258 struct virtio_net_config, status, &v) < 0) 2259 return; 2260 2261 if (v & VIRTIO_NET_S_ANNOUNCE) { 2262 netdev_notify_peers(vi->dev); 2263 virtnet_ack_link_announce(vi); 2264 } 2265 2266 /* Ignore unknown (future) status bits */ 2267 v &= VIRTIO_NET_S_LINK_UP; 2268 2269 if (vi->status == v) 2270 return; 2271 2272 vi->status = v; 2273 2274 if (vi->status & VIRTIO_NET_S_LINK_UP) { 2275 virtnet_update_settings(vi); 2276 netif_carrier_on(vi->dev); 2277 netif_tx_wake_all_queues(vi->dev); 2278 } else { 2279 netif_carrier_off(vi->dev); 2280 netif_tx_stop_all_queues(vi->dev); 2281 } 2282 } 2283 2284 static void virtnet_config_changed(struct virtio_device *vdev) 2285 { 2286 struct virtnet_info *vi = vdev->priv; 2287 2288 schedule_work(&vi->config_work); 2289 } 2290 2291 static void virtnet_free_queues(struct virtnet_info *vi) 2292 { 2293 int i; 2294 2295 for (i = 0; i < vi->max_queue_pairs; i++) { 2296 napi_hash_del(&vi->rq[i].napi); 2297 netif_napi_del(&vi->rq[i].napi); 2298 netif_napi_del(&vi->sq[i].napi); 2299 } 2300 2301 /* We called napi_hash_del() before netif_napi_del(), 2302 * we need to respect an RCU grace period before freeing vi->rq 2303 */ 2304 synchronize_net(); 2305 2306 kfree(vi->rq); 2307 kfree(vi->sq); 2308 } 2309 2310 static void _free_receive_bufs(struct virtnet_info *vi) 2311 { 2312 struct bpf_prog *old_prog; 2313 int i; 2314 2315 for (i = 0; i < vi->max_queue_pairs; i++) { 2316 while (vi->rq[i].pages) 2317 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0); 2318 2319 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 2320 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL); 2321 if (old_prog) 2322 bpf_prog_put(old_prog); 2323 } 2324 } 2325 2326 static void free_receive_bufs(struct virtnet_info *vi) 2327 { 2328 rtnl_lock(); 2329 _free_receive_bufs(vi); 2330 rtnl_unlock(); 2331 } 2332 2333 static void free_receive_page_frags(struct virtnet_info *vi) 2334 { 2335 int i; 2336 for (i = 0; i < vi->max_queue_pairs; i++) 2337 if (vi->rq[i].alloc_frag.page) 2338 put_page(vi->rq[i].alloc_frag.page); 2339 } 2340 2341 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q) 2342 { 2343 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs)) 2344 return false; 2345 else if (q < vi->curr_queue_pairs) 2346 return true; 2347 else 2348 return false; 2349 } 2350 2351 static void free_unused_bufs(struct virtnet_info *vi) 2352 { 2353 void *buf; 2354 int i; 2355 2356 for (i = 0; i < vi->max_queue_pairs; i++) { 2357 struct virtqueue *vq = vi->sq[i].vq; 2358 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) { 2359 if (!is_xdp_raw_buffer_queue(vi, i)) 2360 dev_kfree_skb(buf); 2361 else 2362 put_page(virt_to_head_page(buf)); 2363 } 2364 } 2365 2366 for (i = 0; i < vi->max_queue_pairs; i++) { 2367 struct virtqueue *vq = vi->rq[i].vq; 2368 2369 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) { 2370 if (vi->mergeable_rx_bufs) { 2371 put_page(virt_to_head_page(buf)); 2372 } else if (vi->big_packets) { 2373 give_pages(&vi->rq[i], buf); 2374 } else { 2375 put_page(virt_to_head_page(buf)); 2376 } 2377 } 2378 } 2379 } 2380 2381 static void virtnet_del_vqs(struct virtnet_info *vi) 2382 { 2383 struct virtio_device *vdev = vi->vdev; 2384 2385 virtnet_clean_affinity(vi, -1); 2386 2387 vdev->config->del_vqs(vdev); 2388 2389 virtnet_free_queues(vi); 2390 } 2391 2392 /* How large should a single buffer be so a queue full of these can fit at 2393 * least one full packet? 2394 * Logic below assumes the mergeable buffer header is used. 2395 */ 2396 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq) 2397 { 2398 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 2399 unsigned int rq_size = virtqueue_get_vring_size(vq); 2400 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu; 2401 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len; 2402 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size); 2403 2404 return max(max(min_buf_len, hdr_len) - hdr_len, 2405 (unsigned int)GOOD_PACKET_LEN); 2406 } 2407 2408 static int virtnet_find_vqs(struct virtnet_info *vi) 2409 { 2410 vq_callback_t **callbacks; 2411 struct virtqueue **vqs; 2412 int ret = -ENOMEM; 2413 int i, total_vqs; 2414 const char **names; 2415 bool *ctx; 2416 2417 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by 2418 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by 2419 * possible control vq. 2420 */ 2421 total_vqs = vi->max_queue_pairs * 2 + 2422 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ); 2423 2424 /* Allocate space for find_vqs parameters */ 2425 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL); 2426 if (!vqs) 2427 goto err_vq; 2428 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL); 2429 if (!callbacks) 2430 goto err_callback; 2431 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL); 2432 if (!names) 2433 goto err_names; 2434 if (!vi->big_packets || vi->mergeable_rx_bufs) { 2435 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL); 2436 if (!ctx) 2437 goto err_ctx; 2438 } else { 2439 ctx = NULL; 2440 } 2441 2442 /* Parameters for control virtqueue, if any */ 2443 if (vi->has_cvq) { 2444 callbacks[total_vqs - 1] = NULL; 2445 names[total_vqs - 1] = "control"; 2446 } 2447 2448 /* Allocate/initialize parameters for send/receive virtqueues */ 2449 for (i = 0; i < vi->max_queue_pairs; i++) { 2450 callbacks[rxq2vq(i)] = skb_recv_done; 2451 callbacks[txq2vq(i)] = skb_xmit_done; 2452 sprintf(vi->rq[i].name, "input.%d", i); 2453 sprintf(vi->sq[i].name, "output.%d", i); 2454 names[rxq2vq(i)] = vi->rq[i].name; 2455 names[txq2vq(i)] = vi->sq[i].name; 2456 if (ctx) 2457 ctx[rxq2vq(i)] = true; 2458 } 2459 2460 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks, 2461 names, ctx, NULL); 2462 if (ret) 2463 goto err_find; 2464 2465 if (vi->has_cvq) { 2466 vi->cvq = vqs[total_vqs - 1]; 2467 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN)) 2468 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER; 2469 } 2470 2471 for (i = 0; i < vi->max_queue_pairs; i++) { 2472 vi->rq[i].vq = vqs[rxq2vq(i)]; 2473 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq); 2474 vi->sq[i].vq = vqs[txq2vq(i)]; 2475 } 2476 2477 kfree(names); 2478 kfree(callbacks); 2479 kfree(vqs); 2480 kfree(ctx); 2481 2482 return 0; 2483 2484 err_find: 2485 kfree(ctx); 2486 err_ctx: 2487 kfree(names); 2488 err_names: 2489 kfree(callbacks); 2490 err_callback: 2491 kfree(vqs); 2492 err_vq: 2493 return ret; 2494 } 2495 2496 static int virtnet_alloc_queues(struct virtnet_info *vi) 2497 { 2498 int i; 2499 2500 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL); 2501 if (!vi->sq) 2502 goto err_sq; 2503 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL); 2504 if (!vi->rq) 2505 goto err_rq; 2506 2507 INIT_DELAYED_WORK(&vi->refill, refill_work); 2508 for (i = 0; i < vi->max_queue_pairs; i++) { 2509 vi->rq[i].pages = NULL; 2510 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll, 2511 napi_weight); 2512 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx, 2513 napi_tx ? napi_weight : 0); 2514 2515 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg)); 2516 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len); 2517 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg)); 2518 2519 u64_stats_init(&vi->rq[i].stats.syncp); 2520 u64_stats_init(&vi->sq[i].stats.syncp); 2521 } 2522 2523 return 0; 2524 2525 err_rq: 2526 kfree(vi->sq); 2527 err_sq: 2528 return -ENOMEM; 2529 } 2530 2531 static int init_vqs(struct virtnet_info *vi) 2532 { 2533 int ret; 2534 2535 /* Allocate send & receive queues */ 2536 ret = virtnet_alloc_queues(vi); 2537 if (ret) 2538 goto err; 2539 2540 ret = virtnet_find_vqs(vi); 2541 if (ret) 2542 goto err_free; 2543 2544 get_online_cpus(); 2545 virtnet_set_affinity(vi); 2546 put_online_cpus(); 2547 2548 return 0; 2549 2550 err_free: 2551 virtnet_free_queues(vi); 2552 err: 2553 return ret; 2554 } 2555 2556 #ifdef CONFIG_SYSFS 2557 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue, 2558 char *buf) 2559 { 2560 struct virtnet_info *vi = netdev_priv(queue->dev); 2561 unsigned int queue_index = get_netdev_rx_queue_index(queue); 2562 struct ewma_pkt_len *avg; 2563 2564 BUG_ON(queue_index >= vi->max_queue_pairs); 2565 avg = &vi->rq[queue_index].mrg_avg_pkt_len; 2566 return sprintf(buf, "%u\n", 2567 get_mergeable_buf_len(&vi->rq[queue_index], avg)); 2568 } 2569 2570 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute = 2571 __ATTR_RO(mergeable_rx_buffer_size); 2572 2573 static struct attribute *virtio_net_mrg_rx_attrs[] = { 2574 &mergeable_rx_buffer_size_attribute.attr, 2575 NULL 2576 }; 2577 2578 static const struct attribute_group virtio_net_mrg_rx_group = { 2579 .name = "virtio_net", 2580 .attrs = virtio_net_mrg_rx_attrs 2581 }; 2582 #endif 2583 2584 static bool virtnet_fail_on_feature(struct virtio_device *vdev, 2585 unsigned int fbit, 2586 const char *fname, const char *dname) 2587 { 2588 if (!virtio_has_feature(vdev, fbit)) 2589 return false; 2590 2591 dev_err(&vdev->dev, "device advertises feature %s but not %s", 2592 fname, dname); 2593 2594 return true; 2595 } 2596 2597 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \ 2598 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit) 2599 2600 static bool virtnet_validate_features(struct virtio_device *vdev) 2601 { 2602 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) && 2603 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX, 2604 "VIRTIO_NET_F_CTRL_VQ") || 2605 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN, 2606 "VIRTIO_NET_F_CTRL_VQ") || 2607 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE, 2608 "VIRTIO_NET_F_CTRL_VQ") || 2609 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") || 2610 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR, 2611 "VIRTIO_NET_F_CTRL_VQ"))) { 2612 return false; 2613 } 2614 2615 return true; 2616 } 2617 2618 #define MIN_MTU ETH_MIN_MTU 2619 #define MAX_MTU ETH_MAX_MTU 2620 2621 static int virtnet_validate(struct virtio_device *vdev) 2622 { 2623 if (!vdev->config->get) { 2624 dev_err(&vdev->dev, "%s failure: config access disabled\n", 2625 __func__); 2626 return -EINVAL; 2627 } 2628 2629 if (!virtnet_validate_features(vdev)) 2630 return -EINVAL; 2631 2632 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 2633 int mtu = virtio_cread16(vdev, 2634 offsetof(struct virtio_net_config, 2635 mtu)); 2636 if (mtu < MIN_MTU) 2637 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU); 2638 } 2639 2640 return 0; 2641 } 2642 2643 static int virtnet_probe(struct virtio_device *vdev) 2644 { 2645 int i, err = -ENOMEM; 2646 struct net_device *dev; 2647 struct virtnet_info *vi; 2648 u16 max_queue_pairs; 2649 int mtu; 2650 2651 /* Find if host supports multiqueue virtio_net device */ 2652 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ, 2653 struct virtio_net_config, 2654 max_virtqueue_pairs, &max_queue_pairs); 2655 2656 /* We need at least 2 queue's */ 2657 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN || 2658 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX || 2659 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 2660 max_queue_pairs = 1; 2661 2662 /* Allocate ourselves a network device with room for our info */ 2663 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs); 2664 if (!dev) 2665 return -ENOMEM; 2666 2667 /* Set up network device as normal. */ 2668 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE; 2669 dev->netdev_ops = &virtnet_netdev; 2670 dev->features = NETIF_F_HIGHDMA; 2671 2672 dev->ethtool_ops = &virtnet_ethtool_ops; 2673 SET_NETDEV_DEV(dev, &vdev->dev); 2674 2675 /* Do we support "hardware" checksums? */ 2676 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) { 2677 /* This opens up the world of extra features. */ 2678 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG; 2679 if (csum) 2680 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG; 2681 2682 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) { 2683 dev->hw_features |= NETIF_F_TSO 2684 | NETIF_F_TSO_ECN | NETIF_F_TSO6; 2685 } 2686 /* Individual feature bits: what can host handle? */ 2687 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4)) 2688 dev->hw_features |= NETIF_F_TSO; 2689 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6)) 2690 dev->hw_features |= NETIF_F_TSO6; 2691 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN)) 2692 dev->hw_features |= NETIF_F_TSO_ECN; 2693 2694 dev->features |= NETIF_F_GSO_ROBUST; 2695 2696 if (gso) 2697 dev->features |= dev->hw_features & NETIF_F_ALL_TSO; 2698 /* (!csum && gso) case will be fixed by register_netdev() */ 2699 } 2700 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM)) 2701 dev->features |= NETIF_F_RXCSUM; 2702 2703 dev->vlan_features = dev->features; 2704 2705 /* MTU range: 68 - 65535 */ 2706 dev->min_mtu = MIN_MTU; 2707 dev->max_mtu = MAX_MTU; 2708 2709 /* Configuration may specify what MAC to use. Otherwise random. */ 2710 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) 2711 virtio_cread_bytes(vdev, 2712 offsetof(struct virtio_net_config, mac), 2713 dev->dev_addr, dev->addr_len); 2714 else 2715 eth_hw_addr_random(dev); 2716 2717 /* Set up our device-specific information */ 2718 vi = netdev_priv(dev); 2719 vi->dev = dev; 2720 vi->vdev = vdev; 2721 vdev->priv = vi; 2722 2723 INIT_WORK(&vi->config_work, virtnet_config_changed_work); 2724 2725 /* If we can receive ANY GSO packets, we must allocate large ones. */ 2726 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) || 2727 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) || 2728 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) || 2729 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO)) 2730 vi->big_packets = true; 2731 2732 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) 2733 vi->mergeable_rx_bufs = true; 2734 2735 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) || 2736 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 2737 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 2738 else 2739 vi->hdr_len = sizeof(struct virtio_net_hdr); 2740 2741 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) || 2742 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 2743 vi->any_header_sg = true; 2744 2745 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 2746 vi->has_cvq = true; 2747 2748 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 2749 mtu = virtio_cread16(vdev, 2750 offsetof(struct virtio_net_config, 2751 mtu)); 2752 if (mtu < dev->min_mtu) { 2753 /* Should never trigger: MTU was previously validated 2754 * in virtnet_validate. 2755 */ 2756 dev_err(&vdev->dev, "device MTU appears to have changed " 2757 "it is now %d < %d", mtu, dev->min_mtu); 2758 goto free; 2759 } 2760 2761 dev->mtu = mtu; 2762 dev->max_mtu = mtu; 2763 2764 /* TODO: size buffers correctly in this case. */ 2765 if (dev->mtu > ETH_DATA_LEN) 2766 vi->big_packets = true; 2767 } 2768 2769 if (vi->any_header_sg) 2770 dev->needed_headroom = vi->hdr_len; 2771 2772 /* Enable multiqueue by default */ 2773 if (num_online_cpus() >= max_queue_pairs) 2774 vi->curr_queue_pairs = max_queue_pairs; 2775 else 2776 vi->curr_queue_pairs = num_online_cpus(); 2777 vi->max_queue_pairs = max_queue_pairs; 2778 2779 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */ 2780 err = init_vqs(vi); 2781 if (err) 2782 goto free; 2783 2784 #ifdef CONFIG_SYSFS 2785 if (vi->mergeable_rx_bufs) 2786 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group; 2787 #endif 2788 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs); 2789 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs); 2790 2791 virtnet_init_settings(dev); 2792 2793 err = register_netdev(dev); 2794 if (err) { 2795 pr_debug("virtio_net: registering device failed\n"); 2796 goto free_vqs; 2797 } 2798 2799 virtio_device_ready(vdev); 2800 2801 err = virtnet_cpu_notif_add(vi); 2802 if (err) { 2803 pr_debug("virtio_net: registering cpu notifier failed\n"); 2804 goto free_unregister_netdev; 2805 } 2806 2807 virtnet_set_queues(vi, vi->curr_queue_pairs); 2808 2809 /* Assume link up if device can't report link status, 2810 otherwise get link status from config. */ 2811 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) { 2812 netif_carrier_off(dev); 2813 schedule_work(&vi->config_work); 2814 } else { 2815 vi->status = VIRTIO_NET_S_LINK_UP; 2816 virtnet_update_settings(vi); 2817 netif_carrier_on(dev); 2818 } 2819 2820 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++) 2821 if (virtio_has_feature(vi->vdev, guest_offloads[i])) 2822 set_bit(guest_offloads[i], &vi->guest_offloads); 2823 2824 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n", 2825 dev->name, max_queue_pairs); 2826 2827 return 0; 2828 2829 free_unregister_netdev: 2830 vi->vdev->config->reset(vdev); 2831 2832 unregister_netdev(dev); 2833 free_vqs: 2834 cancel_delayed_work_sync(&vi->refill); 2835 free_receive_page_frags(vi); 2836 virtnet_del_vqs(vi); 2837 free: 2838 free_netdev(dev); 2839 return err; 2840 } 2841 2842 static void remove_vq_common(struct virtnet_info *vi) 2843 { 2844 vi->vdev->config->reset(vi->vdev); 2845 2846 /* Free unused buffers in both send and recv, if any. */ 2847 free_unused_bufs(vi); 2848 2849 free_receive_bufs(vi); 2850 2851 free_receive_page_frags(vi); 2852 2853 virtnet_del_vqs(vi); 2854 } 2855 2856 static void virtnet_remove(struct virtio_device *vdev) 2857 { 2858 struct virtnet_info *vi = vdev->priv; 2859 2860 virtnet_cpu_notif_remove(vi); 2861 2862 /* Make sure no work handler is accessing the device. */ 2863 flush_work(&vi->config_work); 2864 2865 unregister_netdev(vi->dev); 2866 2867 remove_vq_common(vi); 2868 2869 free_netdev(vi->dev); 2870 } 2871 2872 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev) 2873 { 2874 struct virtnet_info *vi = vdev->priv; 2875 2876 virtnet_cpu_notif_remove(vi); 2877 virtnet_freeze_down(vdev); 2878 remove_vq_common(vi); 2879 2880 return 0; 2881 } 2882 2883 static __maybe_unused int virtnet_restore(struct virtio_device *vdev) 2884 { 2885 struct virtnet_info *vi = vdev->priv; 2886 int err; 2887 2888 err = virtnet_restore_up(vdev); 2889 if (err) 2890 return err; 2891 virtnet_set_queues(vi, vi->curr_queue_pairs); 2892 2893 err = virtnet_cpu_notif_add(vi); 2894 if (err) 2895 return err; 2896 2897 return 0; 2898 } 2899 2900 static struct virtio_device_id id_table[] = { 2901 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID }, 2902 { 0 }, 2903 }; 2904 2905 #define VIRTNET_FEATURES \ 2906 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \ 2907 VIRTIO_NET_F_MAC, \ 2908 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \ 2909 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \ 2910 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \ 2911 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \ 2912 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \ 2913 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \ 2914 VIRTIO_NET_F_CTRL_MAC_ADDR, \ 2915 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \ 2916 VIRTIO_NET_F_SPEED_DUPLEX 2917 2918 static unsigned int features[] = { 2919 VIRTNET_FEATURES, 2920 }; 2921 2922 static unsigned int features_legacy[] = { 2923 VIRTNET_FEATURES, 2924 VIRTIO_NET_F_GSO, 2925 VIRTIO_F_ANY_LAYOUT, 2926 }; 2927 2928 static struct virtio_driver virtio_net_driver = { 2929 .feature_table = features, 2930 .feature_table_size = ARRAY_SIZE(features), 2931 .feature_table_legacy = features_legacy, 2932 .feature_table_size_legacy = ARRAY_SIZE(features_legacy), 2933 .driver.name = KBUILD_MODNAME, 2934 .driver.owner = THIS_MODULE, 2935 .id_table = id_table, 2936 .validate = virtnet_validate, 2937 .probe = virtnet_probe, 2938 .remove = virtnet_remove, 2939 .config_changed = virtnet_config_changed, 2940 #ifdef CONFIG_PM_SLEEP 2941 .freeze = virtnet_freeze, 2942 .restore = virtnet_restore, 2943 #endif 2944 }; 2945 2946 static __init int virtio_net_driver_init(void) 2947 { 2948 int ret; 2949 2950 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online", 2951 virtnet_cpu_online, 2952 virtnet_cpu_down_prep); 2953 if (ret < 0) 2954 goto out; 2955 virtionet_online = ret; 2956 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead", 2957 NULL, virtnet_cpu_dead); 2958 if (ret) 2959 goto err_dead; 2960 2961 ret = register_virtio_driver(&virtio_net_driver); 2962 if (ret) 2963 goto err_virtio; 2964 return 0; 2965 err_virtio: 2966 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 2967 err_dead: 2968 cpuhp_remove_multi_state(virtionet_online); 2969 out: 2970 return ret; 2971 } 2972 module_init(virtio_net_driver_init); 2973 2974 static __exit void virtio_net_driver_exit(void) 2975 { 2976 unregister_virtio_driver(&virtio_net_driver); 2977 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 2978 cpuhp_remove_multi_state(virtionet_online); 2979 } 2980 module_exit(virtio_net_driver_exit); 2981 2982 MODULE_DEVICE_TABLE(virtio, id_table); 2983 MODULE_DESCRIPTION("Virtio network driver"); 2984 MODULE_LICENSE("GPL"); 2985